<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Linzhi ASICs on Medium]]></title>
        <description><![CDATA[Stories by Linzhi ASICs on Medium]]></description>
        <link>https://medium.com/@Linzhi?source=rss-5f5a9e27955b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*KUjUu5ExCjhOuDiH.jpg</url>
            <title>Stories by Linzhi ASICs on Medium</title>
            <link>https://medium.com/@Linzhi?source=rss-5f5a9e27955b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 19:48:35 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@Linzhi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[History of Bitcoin Mining Hardware]]></title>
            <link>https://medium.com/@Linzhi/history-of-bitcoin-mining-hardware-60be773e5f5d?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/60be773e5f5d</guid>
            <category><![CDATA[history]]></category>
            <category><![CDATA[sha-256]]></category>
            <category><![CDATA[bitcoin]]></category>
            <category><![CDATA[semiconductors]]></category>
            <category><![CDATA[asic-mining]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Thu, 14 Nov 2019 11:08:27 GMT</pubDate>
            <atom:updated>2019-11-15T16:24:09.301Z</atom:updated>
            <content:encoded><![CDATA[<p>Optimizing SHA-256</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HOBEeHHQXEcvFfrQSQuLcA.jpeg" /></figure><p>A hardware maker optimizes one or more of: cost, performance, efficiency, time to market or reliability.</p><p>In this article we walk the interested reader through this process for SHA-256, with emphasis on ASICs, and finally leading to the supply chain. In parentheses you find the rough improvement. C = cost, E = efficiency, P = performance. Note that C+5% means cost was improved (=reduced) by 5%. The article is rather dry without drawings. With sufficient interest we might add more detail or drawings. From FPGA on it’s a firsthand account.</p><p>For now it’s good as it is. Enjoy.</p><p><strong>Phase I — CPU era</strong></p><p>Bitcoin mining started in 2009 on CPUs (~5000 J/GH).</p><p><strong>Phase II — GPU era</strong></p><p>In 2010 and 2011, GPUs increased performance 10x-100x (~500 J/GH)</p><p><strong>Phase III — FPGA era</strong></p><p>The first FPGAs didn’t provide better performance per chip, but better efficiency and slightly better cost (~50 J/GH). The biggest optimization during the FPGA era was a <strong>pipeline architecture</strong>.</p><p>The performance bottleneck of SHA-256 was its deep algebra chain. Adding one more pipeline stage in the middle of the algebra chain doubled performance (P+100%, from 100 MHz to 200 MHz), for example on the Icarus FPGA miner. Further distribution of the algebra to different pipeline stages increased performance to 300 MHz (P+50%), for example on the Lancelot FPGA miner or Bitfury’s FPGA miner.</p><p>Other optimizations during the FPGA era were:</p><ol><li><strong>Pre-calculation</strong>: Remove as much as logic as possible from the hashing loop (C+5%, E+5%).</li><li><strong>Toggle rate reduction</strong>: Reduce average gate toggle rate (E+1%).</li><li><strong>Dynamic frequency control</strong>: Let every hash core operate on it’s best frequency (P+20%).</li><li><strong>Hand place and route</strong>: Makes space for more hash cores in the FPGA (P+15%). FPGA toolchain limitations made this a very hard effort.</li><li><strong>Over Voltage</strong>: The power distribution system of FPGAs is not designed for high power operation, so it drops a lot of voltage. Even when you power 1.2V at the PCB, on silicon you will have less than 1.0V due to package drop, which is a big performance hit. We couldn’t change the chip design (yet), so the only way to counter this was to supply the chip with higher voltage (P+30%).</li><li><strong>Miner management bus and protocol:</strong> Design for managing large amounts of miners, using minimum wire count. That saved cost for large scale operation (C+5%), and continued into the ASIC era.</li></ol><p><strong>Phase IV — Function ASICs</strong></p><p>In 2012 many teams started designing mining ASICs, such as Butterfly Labs, Avalon Miner, Bitfury, Friedcat, etc. At that time the design quality was not so important, time to market was paramount. Dozens of teams rushed into the market, hashrate increase reached 40% a week.</p><p>By <strong>removing programmability</strong>, even an old process ASIC cost much less than an FPGA (C+1000%).</p><p>The first ASICs fixed FPGA weaknesses in power distribution and thermal resistance. With only 1 clock pin, 1 reset pin and 4 signal pins, <strong>all other pins were used for power and ground</strong>. <strong>QFN packages</strong> provided good thermal performance at low cost (C+50%, E+50%, P+100%).</p><p>Good <strong>bus design</strong> also contributed, allowing for maximum pin count used to enhance power and thermal performance. Using a wide parallel bus wouldn’t provide much package benefit. An improved <strong>control protocol</strong> enabled the connection of thousands of chips with low cost and low pin count.</p><p>Good thermal design was more important in the ASIC era because the heat density of an ASIC is much higher than CPU/GPU/FPGA and most other chips. It reached 100 W/cm2 or more. <strong>High thermal packaging material</strong> (P+30%) proved to be very useful.</p><p>In 2014 the Bitcoin price crashed, and the era of function ASICs came to an end. Hashrate was already high, low-tech ASICs didn’t make sense anymore, and energy efficiency became the top priority.</p><p><strong>Phase V — Efficiency ASICs</strong></p><p>Many of the early ASIC teams had failed by now, but several big players started to optimize efficiency, which turned out to be a difficult and long path.</p><p>First was the semiconductor process. In the function ASIC era, people used mature processes, such as 180/153nm, and later 90/80nm or 65/55nm. Driven by performance, cost, maybe other reasons. Deeper processes provided better transistors, lower power and lower cost per transistor.</p><p>The arms race finally reached deeper process nodes such as 45/40nm, 32/28nm, someone even tried 20/16nm. In 2014, 28nm was already a state-of-the-art process, 20/16nm was just coming out, but it was expensive and quite unstable still.</p><p>Reaching 28nm meant you were at the leading semiconductor process at that time, there was not much room to go deeper, so optimization of the design on 28nm became the top priority.</p><p>The <strong>28nm process itself</strong> provided ~5x better efficiency compared to 130nm (C+50%, E+500%, P+100%).</p><p>The efficiency formula of a CMOS circuit is</p><blockquote>Efficiency ∝ TCV²<br>(T=toggle rate, C=capacitance, V²=voltage squared)</blockquote><p>The <strong>toggle rate</strong> was already optimized since the FPGA era, now it was further optimized to improve efficiency (E+10%). <strong>Capacitance</strong> depends on transistor and process, optimizing it meant a redesign of the entire circuit library, which turned out to be a lot of work but provided good results (E+15%, P+15%). Normally smaller transistors have better efficiency.</p><p><strong>Voltage</strong> was the biggest deal in all optimizations, the only 2nd order (squared) effect in all of mining. Reducing the supply voltage from 0.9V (typical at 28nm) to 0.6V (most 28nm miners) provided an advantage (E+225%). However reaching such low voltage was hard. A single optimization may have enabled a reduction of only 0.01V, meaning that many optimizations had to be accumulated to reach low voltages. This played an even bigger role in FinFET processes such as 16/14/12/10/7/5nm.</p><p>16/14nm FinFET didn’t provide much benefit over 28nm at the same supply voltage. However FinFET can run down to 0.30V, which itself provided a big boost (E+400%), but was difficult to design.</p><p><strong>Dynamic circuits</strong> were the single biggest improvement. Normally ASICs will be designed using standard cells, static CMOS design. Due to Bitcoin not needing to store data longer than 1 clock cycle, we could DFF (D-flipflop) using a dynamic circuit. A typical non-scan DFF cell is 22 transistors (22T), with a well designed dynamic circuit we could reduce the transistor count to 10T, 8T and below, with latch design we could reduce it to 4T or less. It provided significant cost reduction and power savings (C+30%, E+30%, P+20%).</p><p><strong>Flipchip packages</strong> became common at this time. Flipchip and good substrate provided much lower electrical resistance, to enable lower voltage operation. It also provided good thermal performance on both top and bottom side. Large silicon die designs (&gt;100mm2) were abandoned after 28nm, because the heat density became so high that it was impossible to maintain a reasonable temperature. A small die can dissipate heat over 6 sides (top, bottom, 4 edges), but a large die can only dissipate heat over one side (top). Good packaging brought substantial benefit (E+50%) to the chip.</p><p><strong>Power cascading</strong> was another major cost optimization. When ASIC supply voltages became less and less, using DCDC supplies became uneconomic. Energy was lost on inductors, capacitors, wires, connectors, etc. A better choice was to cascade ASIC chips together, for example a 0.6V cascade of 20 stages to reach 12V, or 0.4V 30 stages, or 0.3V 40 stages. Power cascades turned out to be difficult. They were unstable, semiconductor processes have variations making every chip slightly different. Small differences caused big problems, especially during power-up sequences in cold weather. Stabilizing needed several iterations of trial and error. IO and clocking, reset, software, temperature distribution within the machine — all became important. Balance loss meant to lose a whole chain of chips. In the end power cascading led to reduced current, increased voltage, saved cost and increased efficiency (C+100%, E+50%).</p><p>The famous <strong>ASIC Boost</strong> (consider it a math optimization) brought some advantage (C+15%, E+15%), but in the context of all other improvements wasn’t outstanding or worthy of being considered an attack vector in our opinion. It didn’t provide more advantage than voltage or dynamic circuits. The efficiency improvement of ASIC Boost was just as good as an improved <strong>fan speed adjustment</strong> algorithm (E+15%), or keeping <strong>each chip at the optimum temperature</strong> for that chip (E+15%).</p><p><strong>Approximate adders</strong> (C+10%, E+10%, P+10%): Mining chips can tolerate an error rate of 1–2%. Making a circuit less accurate will reduce cost and provide better efficiency. Lack of toolchain support made this harder to implement than expected.</p><p><strong>Process adjustments</strong> (E+15%, P+15%): Semiconductor processes have many parameters to adjust, such as the transistor threshold voltage, annealing temperature, poly gate size. While such adjustments can provide a 30% benefit, they need a lot of sampling and testing resources, and then analysis of the results for efficiency and performance, a process which will take several months.</p><p><strong>Phase VI — Supply Chain</strong></p><p>When 16nm chips became stable, the process arms race stopped. There was no better process for a long time. Volume became the top priority. Big ASIC companies rushed to book materials. From foundry to packaging house, substrate vendor, thermal material vendor and other key components. <strong>Pre-ordering 80% of key supplier capacity</strong> or signing <strong>exclusive supply agreements</strong> became important tools to compete.</p><p>By now mining occupied a significant percentage of the semiconductor industry, every supplier was aware and had to <strong>prioritize their large customers</strong>. Large ASIC mining companies now dominated the supply chain.</p><p>Small companies could design a chip, but couldn’t order enough capacity from a foundry. Small companies had to prepay 100% for wafer orders, where large companies could prepay as little as 0% or 10%, giving them a big cashflow advantage.<br>The best packaging materials were sold exclusively to big companies, small companies wouldn’t even get a sample. Same for most other key components.</p><p><strong>Epilogue</strong></p><p>What does this mean? What’s next?<br>Who knows, you tell us! Come to our Telegram and let’s hash out ideas.</p><p>Chen Min<br>Linzhi, Shenzhen<br>Telegram: <a href="https://t.me/LinzhiCorp">https://t.me/LinzhiCorp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=60be773e5f5d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Linzhi E1400 — Architecture Overview]]></title>
            <link>https://medium.com/@Linzhi/linzhi-e1400-architecture-overview-6fed5a19ef70?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/6fed5a19ef70</guid>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[bitcoin]]></category>
            <category><![CDATA[asic-mining]]></category>
            <category><![CDATA[chip-design]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Mon, 07 Oct 2019 22:07:19 GMT</pubDate>
            <atom:updated>2019-10-07T22:23:37.858Z</atom:updated>
            <content:encoded><![CDATA[<h3>Linzhi E1400 — Architecture Overview</h3><p>[This article is an edited version of the talk given by Werner Almesberger at ETC Summit 2019 in Vancouver, on Oct 4th, 2019.]<br>Video: <a href="https://www.youtube.com/watch?v=_enGUG_Jbi4">https://www.youtube.com/watch?v=_enGUG_Jbi4</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F_enGUG_Jbi4%3Ffeature%3Doembed&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D_enGUG_Jbi4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F_enGUG_Jbi4%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/52b07c6208dd08074f474741c0fbba59/href">https://medium.com/media/52b07c6208dd08074f474741c0fbba59/href</a></iframe><p>Good afternoon, my name is Werner Almesberger.</p><p>I’ve worked on the Linux kernel many years ago, and today I talk about what I am working on at the moment. For a change that has nothing to do with software, it’s a mining ASIC for ETC, for Ethash. Linzhi is a Shenzhen-based fabless semiconductor company, we <a href="https://www.coindesk.com/a-new-line-of-powerful-asic-miners-is-coming-to-ethereum">announced our Ethash miner</a> at the last ETC Summit, expecting a performance of about 1.4 GH/sec at a power consumption of 1 kW.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BICY1ah0Ycn1Y-1Cx6eW_g.png" /></figure><p>Reviewing the beginnings of Bitcoin, there was CPU mining with SHA-256, running on your CPU with one or a small number of compute cores. People quickly realized that you can do better, because the CPU is a big chip with a lot of transistors, you have floating point and other things, even if none of it is used for mining. So why not reduce the functionality, and do only what you actually need to do, namely the hash calculation, and put this on a chip — then you can have a lot of little compute units in the same space, so the chip costs roughly the same. They all work in parallel, and you can do lots of hash operations with the same effort, and gain a huge benefit over the CPU. This is how ASICs were born. The best CPU miners in 2011, when people basically gave up even trying to CPU-mine, did something like 30 MH/sec, while a contemporary ASIC miner today does 60 TH/sec. If you adjust those values for technological progress over time, using Moore’s law, then we would get a performance difference between a CPU extrapolated to the present day, and the present day ASIC miner, of about 50,000 — in terms of hashing speed.</p><p>You may not be interested in hashing speed per se, but more in power efficiency. There the factor is a little bit smaller, maybe something like 5,000 or a bit more. Still, ASICs have a huge benefit over CPUs. If you had a time machine, sitting around somewhere in your basement, and you would like to put this to some creative use, you could go and shop and buy a state-of-the-art ASIC miner for Bitcoin, hop into your time machine, go back 8 years, get out of your DeLorean, plug in your miner, and suddenly you would have increased the global hashrate from 10 TH to 70 TH. We are not talking about a 51% attack, we are talking about an 85% attack. You basically own the network.</p><p>Clearly, when people saw what was happening with the ASICs, they were a bit shocked. And that caused a lot of bad blood with respect to ASICs. Some people started to think about ways to solve the ASIC problem, and this is how they came up with Ethash.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*522QUJ-tPqlQi5ExdkLBdQ.png" /></figure><p>In Ethash what you get is a header hash value, and you have your nonce, which you change on your own for the attempts you make. You hash those two together, and with this you have generated your task. Now you do the mixing, where you first calculate an address from your current value, and lookup this address in the DAG, a big block of memory, and you retrieve the corresponding memory line. You mix it with the result you have so far, and you repeat the whole operation 64 times, then you do some more hashing, and you compare with the difficulty target. The DAG is the new ingredient that should make the thing more difficult for ASICs. It makes it memory hard, and the DAG is very large, so you can’t just put it on a little corner on your chip. It’s currently about 3 GB and it’s growing over time.</p><p>Let’s look at the comparison:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PBxiSdvFh7TBZS-FGWECvg.png" /></figure><p>You have your CPU miners, a small number of cores, a reasonable amount of memory bandwidth. You need both, you need to compute, and you need the memory bandwidth.<br>With GPU miners, you can do better, they are more specialized, so your compute cores get smaller, you can have more of them at the same chip surface, and they are also optimized for memory bandwidth. They have vast memory bandwidth, they have wide buses. So they are better than CPUs for mining.<br>ASICs could of course still win on the side of the compute units, because the GPUs are still sort-of general purpose. You could make an ASIC that is only Ethash, but you would still have the memory bandwidth as your limiting factor. So unfortunately, you are not really winning. This means that you might be able to more or less pull even with the GPUs, but then GPUs are commodity hardware, they are easy to get, a big market, you have economies of scale on your side with a GPU.<br>So, ASICs just shouldn’t make sense, and this would mean that those evil ASICs would stay away from the precious Ethash. So far the theory.</p><p>[Please briefly review the Ethash diagram above: Green is the generation of the task, red the mixing, yellow the finalizing, and blue is the memory.]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VlZWxSFBj7bkkyBfU6F8kg.png" /></figure><p>In our architecture, this is what our chip more or less looks like. We have those little green boxes which do the generation. We have multiple of them so they don’t have to work too hard. And then, when the task is generated they send it to a so-called mixer node. The mixer node contains a bit of logic to do the computation, and a bit of memory. And there are lots of mixer nodes, 72 of them. They are connected with each other through a network-on-a-chip. And, at the bottom you can see the yellow boxes which are the finalizers which do the final hash and the comparison. Each of the 72 mixing nodes has 1 MB of memory. This means 72 MB of memory per chip. Which is too small for the DAG. We can solve this by using multiple chips, distributing the DAG over those chips, but then we do also need to communicate between chips, and so we add — which is shown on the side — communication interfaces where the chips can talk to each other.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KGRkYu1IAPeS2ebEXtACbw.png" /></figure><p>Now let’s look at how the hashing operation works. First of all you do the generation, which we call round 0, then you send it to the adjacent mixer node, where only the DAG address is calculated, because we don’t know yet where this thing has to go. Then we send it to the node that has the corresponding DAG line in it’s memory. The DAG line gets retrieved, the mixing is done, it calculates the next DAG address for the next round. Let’s say this DAG address is in the memory on a different chip, so it gets sent to the node where the communication interface connects, then it gets handed over to the the communication interface, arrives at the destination chip, and there it enters on the node where the communication interface is attached. This will then route it to the node which has the memory. There we again retrieve the memory line, do the mixing, calculate the next DAG address, and so on, 64 times in all, until the last round is done. And then it gets sent down to the finalizer on the chip. So generator and finalizer don’t have to be on the same chip, it doesn’t really matter where things start and where they end.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-XyLWQNgv3p98E7jmCTfaQ.png" /></figure><p>Let’s look at performance — memory bandwidth. There are a few examples for GPUs, contemporary GPUs. I’m not quite sure if they are the best for mining, but they have pretty good, pretty reasonable performance. One would be an AMD Radeon RX 5700, which has a memory bus of 256 bits, and it gets a throughput of about 450 GB/sec. Another one would be the Nvidia Titan RTX, which has a wider memory bus, and does about 670 GB/sec. In our system, memory has a speed of 600 MHz, and it needs 2 cycles for every access. Our memory has a width of 1024 bits, which means that we get throughput of almost 40 GB/sec, per mixer node in the chip. There are 72 mixers in a chip, which means we get to almost 3 TB/sec.<br>This is already starting to look good in comparison, right?, because the GPUs don’t even reach a TB. But then we have 64 of those chips, so our total memory bandwidth in the whole system is almost 180 TB/sec. I think we can confidently say that memory bandwidth is not really a limiting factor here. In fact, we have so much overkill there that we use it extremely inefficiently. We also store a queue in there, so for every access we do to the DAG, we do two accesses for just moving data around. But we have lots of spare bandwidth. However, this isn’t our real bottleneck.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rb9ZwLp_Y6wNdUMC23xAEQ.png" /></figure><p>Our real bottleneck is the communication now, because the chips have to talk to each other. We have up to 42 Gbits/sec, per communication link. All the chips are fully connected with each other, so each chip connects to all 63 others on the board. The links are bidirectional, this means that we have a peak throughput of about 130 Tbits/sec on a board, and for every task, for every complete hash — all 64 rounds, we need to send roughly 75 Kbits, if everything added together. This means that we have a theoretical limit of about 1.7 GH/sec. There are always some factors where we don’t use it perfectly, and where things get slowed down, there are some lossages along the way, so we expect a real performance of 1.4 GH/sec.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JaTUS5wr9tyDGPtbQYB_Tg.png" /></figure><p>There is a little bit more to making a miner. The miner can have one or two ASIC boards. There is a controller board, we have a Xilinx Zynq FPGA, which acts as a microcontroller. We have a few things that we like to do in the FPGA, but it doesn’t really do much work. All this goes into an enclosure, which has two fans on the rear. There is also power input, 12V standard mining power. On the front panel we have the Ethernet plug, a reset button, another button for some control functions, plus some blinken lights.</p><p>We have done our tapeout last month, finally. Tapeout is when you are finished with the design and your simulations and such, and you send the files to the semiconductor factory, and then they make the chip for you. Now we are waiting for those chips to come back, this will take about 2 months, so we expect them to be back around mid to end November, and then we will start with the bringup, which means we will test whether the chip converts power into, perhaps, some calculations, or if it’s more interested in converting power into smoke. Then we need to see if we can talk to it, and if it’s willing to communicate, and we can maybe try to mine a little bit on a single chip, which is really completely useless, because we don’t have the full DAG. But it would show us that the logic is working, we can test this, then combine multiple chips, and so on. Finally integrate everything, check that all the software is working, and we hope that things will be available for sale around first quarter next year.</p><p>By now you may be asking yourself the question: Given that there are now all these ASICs for Ethash — has Ethash failed? <br>You could say in one sense, if you use a very strict interpretation of what ASIC Resistance means, namely that there could not be any ASICs, they would be economically infeasible, then you could say yes. But, if you think about what happened to Bitcoin, where basically the CPUs got eradicated by ASICs, because they enjoyed such a huge performance difference, and apparently nobody really saw that coming, which is surprising. Then you see that actually Ethash did extremely well. Because now there are ASICs, and they are a bit more efficient than the GPUs, which is to be expected. You started your ASIC resistant design on using ASICs already. Your CPU, GPU, are just ASICs. So obviously, it can be done on ASICs, and if you are a specialized ASIC it’s extremely likely to be a little bit faster, at least. That is pretty much a given. But, you can also see that those ASICs are not a thousand times better than CPUs or GPUs.<br>That’s the good news — Ethash actually worked, it did its job, it prevented the threat of a hostile takeover.</p><p>If you want to know more, you can go to our website <a href="http://linzhi.io/">linzhi.io</a>, and we are also on Telegram as <a href="http://t.me/LinzhiCorp">t.me/LinzhiCorp</a>.<br>Thank you all.</p><p>Werner Almesberger, Buenos Aires/Vancouver</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q0nTRvZIjO7hN_fj7nawzA.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6fed5a19ef70" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Memory-Hard?]]></title>
            <link>https://medium.com/@Linzhi/what-is-memory-hard-45a363b59dfe?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/45a363b59dfe</guid>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[memory-hard]]></category>
            <category><![CDATA[ethash]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[ethereum-classic]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Fri, 04 Oct 2019 13:30:56 GMT</pubDate>
            <atom:updated>2019-10-04T16:50:56.994Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The Original Idea</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KhNxm65puAauTQoe-bjwsg.png" /></figure><p><a href="https://en.wikipedia.org/wiki/Scrypt">Scrypt</a> predates cryptocurrencies, and first appeared in the cryptocurrency scene as part of Tenebrix in September 2011, implemented by ArtForz. A few weeks later it was picked up by Charlie Lee for <a href="https://bitcointalk.org/index.php?topic=47417.0">Litecoin</a>, and is in use by Litecoin until today.</p><p>Tenebrix <a href="https://bitcointalk.org/index.php?topic=45667.0">promised Scrypt to be ‘GPU resistant’</a>, and to ‘resist the creation of efficient GPU, FPGA and even ASIC implementations’. It promised to ‘remain CPU friendly and GPU hostile’, and to become a ‘CPUclusive cryptocurrency’.</p><p>Early Tenebrix users reported hashrates of about 1.6 khash/sec on their CPUs.<br>Today Tenebrix is history, however Litecoin is still around and the Bitmain L3+ Scrypt miner can mine 500 MH/s at 900W, exceeding the 2011 CPUs by a factor of 300,000.</p><p>In the September-2011 <a href="https://github.com/Lolcust/Tenebrix/blob/cpumine/src/scrypt/scrypt.c">scrypt.c</a>, ArtForz achieved GPU resistance by using a 1MB memory structure. At the time, 1 MB fit into a CPU L2/L3 cache, but not into a GPU L2 cache.</p><p>“the GPU’s L2 cache is much smaller than a typical CPU’s L2 or L3 cache, but has much higher bandwidth available” <a href="http://supercomputingblog.com/cuda/cuda-memory-and-cache-architecture/">http://supercomputingblog.com/cuda/cuda-memory-and-cache-architecture/</a></p><p>The access pattern of memory accesses in Scrypt was read-many/write-few.<br>If it would have been read-many/write-many or read-many/write-once — things would have developed differently. It is unclear whether read-many/write-few was a deliberate design choice.<br>Other memory-hard algorithms such as Cryptonight or Grin are using read-many/write-many access patterns.</p><p>To summarize the beginnings of PoW memory-hardness in 2011 by example of Scrypt, we see that it meant:</p><ul><li>read-many, write-few</li><li>Memory size chosen to fit inside a CPU L2/L3 cache.</li><li>Lack of information about chip design and economics.</li></ul><p><strong>Down-Sampling Litecoin</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WADjtcnd_MnL0lLC5Q9u9g.png" /></figure><p>There are different reasons why a 2019 Bitmain L3+ outperforms a 2011 CPU in Scrypt by a factor of 300,000. The one we want to point out here, related to memory-hardness, is how Scrypt accesses memory.</p><p>Scrypt determines the random location in memory through a seed and a hash.</p><p>However, ASIC engineers found out that the function f() where memᵢ₊₁ = f(memᵢ) was relatively simple. They realized they could reduce the amount of memory to ⅛ by storing only every 8th item. If access to a %8 != 0 item was requested, the chip would read the %8 = 0 item, and then calculate up to the requested item.</p><p>That way, memory size was traded for logic effort, which was a net positive for an ASIC but most likely not anticipated by the software engineers that designed the 1MB structure. In fact any down-sampling ratio could be chosen to match the sweet spot of a given ASIC process. Memory could be down-sampled to ½ for a 1.5x logic effort, to ¼ for a 2.5x logic effort, and so on.</p><p>This meant Scrypt memory-hardness was broken, the algorithm was back to being bound by logic performance only, and there was an ASIC efficiency shock with efficiency gains of 300,000 times over a CPU.</p><p><strong>Why Ethash Is Stable</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A9XGG5xJpgH7CsX84eReBA.png" /></figure><p>Ethash designers learned from Scrypt, and in 2014 created a two-stage memory structure that is growing over time. The first stage is called the pseudorandom cache, and the second stage called the DAG.</p><p>The ratio between pseudorandom cache and DAG was set to 1:64.</p><p>The first-stage memory (pseudorandom cache) functions as a deterrent to memory down-sampling. We believe size and growth of it are secondary as long as it’s not smaller than 1 MB. The second-stage memory (DAG) functions as a deterrent to single-die solutions.<br>Ethash’s two-stage memory system works well to mitigate the use of logic to recalculate data (to replace actual reads from memory), but it now costs more to verify a block.</p><p><strong>Outlook</strong></p><p>Two desirables are currently being discussed around Ethash:</p><ol><li>Enable FlyClient<br>FlyClient seems to be held back by slow verification times, caused by the size of the pseudorandom cache, which is now over 50 MB. We will try to understand where the bottleneck is exactly and how to remove it if it exists.</li><li>Avoid bricking machines (ECIP 1043)<br>In April 2019 the DAG size exceeded 3GB, and at the end of 2020 it is expected to exceed 4 GB. It is not well established why this bricking of otherwise functioning machines is a net-positive for Ethereum, and if it is determined not to be, what the least impact change would be to avoid the bricking without damaging anything else.</li></ol><p>For these two desirables, we are proposing to consider the following points, which I will get to in an upcoming article:</p><ul><li>Don’t break things that work.</li><li>Make sure the unit of account (1 Ethash) remains comparable before and after PoW change.</li><li>Establish what constitutes a PoW failure and a PoW attack, and maintain or improve early warning systems like <a href="http://etcstatus.live/">etcstatus.live</a></li></ul><p>The introduction of memory into the Proof-of-Work algorithm successfully increased the cost of domination, but also brought challenges such as slower verification times and an increased capex:opex ratio because of memory transistors with lower activity ratio than logic.<br>For an existing coin with significant market cap, it seems prudent to make the least-damaging fix to alleviate past mistakes, because new mistakes will surely be made and if we don’t limit mistake-fixing radicalism, we will not be able to create long-term value.</p><p>Linzhi Team, Shenzhen</p><p>For questions around Proof-of-Work, memory hardness, ASICs, or anything else on your mind — please come to our Telegram <a href="https://t.me/LinzhiCorp">https://t.me/LinzhiCorp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45a363b59dfe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[On Whistleblowing]]></title>
            <link>https://medium.com/@Linzhi/on-whistleblowing-56132c069aa3?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/56132c069aa3</guid>
            <category><![CDATA[anonymity]]></category>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[predator]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[whistleblowing]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Sat, 21 Sep 2019 10:57:51 GMT</pubDate>
            <atom:updated>2019-09-22T06:44:10.018Z</atom:updated>
            <content:encoded><![CDATA[<p>We know more than what we have talked about so far.<br>We know but we cannot prove.</p><p>There are two types of knowledge: A priori and a posteriori knowledge.</p><p><a href="https://en.wikipedia.org/wiki/A_priori_and_a_posteriori">A priori and a posteriori</a></p><p>Our knowledge about the attack on Ethereum is a posteriori knowledge, so we don’t need to ask anyone, and we don’t need proof ourselves.<br>We know.<br>A posteriori are empirical facts unknowable by reason alone.</p><p>Since we don’t want to ally with the attackers, and we have no proof that would satisfy an audience that is tired in a world of fake news and FUD, what we decided to do is to try to open up eyes about what might be going on — plausibilities, motivations, personal and corporate backgrounds.<br>We do this for 9 months now, it binds a significant amount of our resources.<br>That is all we can do.</p><p><strong>Bob Rao’s Audit</strong></p><p>We are disappointed that so few independent Ethereum developers seem to read and think about Bob Rao’s excellent audit.</p><p><a href="https://github.com/ethereum-cat-herders/progpow-audit/raw/master/Bob%20Rao%20-%20ProgPOW%20Hardware%20Audit%20Report%20Final.pdf">https://github.com/ethereum-cat-herders/progpow-audit/raw/master/Bob%20Rao%20-%20ProgPOW%20Hardware%20Audit%20Report%20Final.pdf</a></p><p>From our view, Bob Rao’s audit confirms that ProgPoW will not achieve its goal of closing the efficiency gap between ASICs and GPUs, which adds to a list of people that have come to the same conclusion.</p><ol><li>Howard Chu</li></ol><h3>Howard Chu on Twitter</h3><p>@UPMEM ProgPoW&#39;s weak computational requirements don&#39;t stand a chance. Integer only, no floating point math, no branching - trivial to implement in ASICs, and with massive space &amp;amp; energy advantages over GPUs. Only a master spin-doctor could paint these audits as positive for ProgPoW.</p><p>2. ether4life</p><p><a href="https://ether4life.tumblr.com/">ether4life</a></p><blockquote>“This results in a 4x advantage in terms of energy efficiency over a GPU for ProgPoW” — ether4life</blockquote><p>(An anonymous Ethereum freedom fighter. We tried to track down who this person is, unsuccessfully. Please come forward privately if you read this.)</p><p>3. Tim Olson</p><p><a href="https://ethereum-magicians.org/t/progpow-audit-delay-issue/3309/47">https://ethereum-magicians.org/t/progpow-audit-delay-issue/3309/47</a></p><blockquote>“they are using more chip, which means more up-front cost (capex), but there will be a lot more power too (opex)” — Tim Olson</blockquote><p>4. Linzhi</p><p><a href="https://medium.com/@Linzhi/eip-1057-progpow-open-chip-design-for-only-1-cost-power-increase-eip-1057-progpow-d106d9baa6eb">EIP 1057 (ProgPoW): Open Chip Design for only 1% cost/power increase
EIP 1057 (ProgPoW)…</a></p><p>Against all this, EIP 1057 states right in its abstract<br><a href="https://github.com/ethereum/EIPs/blob/master/EIPS/eip-1057.md">https://github.com/ethereum/EIPs/blob/master/EIPS/eip-1057.md</a></p><blockquote>ProgPoW is … designed to close the efficiency gap available to specialized ASICs.</blockquote><p>It has now been sufficiently established that EIP 1057 fails to reach its main stated design goal.</p><p>That <em>alone</em> is enough to reject EIP 1057 on technical grounds. Why should Ethereum adopt a technology that doesn’t do what it says it will do?<br>It is enough to reject EIP 1057 without further input or agreement from its authors.</p><p>One must assume the EIP 1057 authors to respond to arguments with new arguments, but as engineers we should be able to think independently through the facts and come to our own judgment, and reject what we understand to be false.</p><p>We don’t need to speculate about business motives, mining fairness and other PoW algo change related risks.</p><p><strong>Linzhi and Hudson Jameson</strong></p><p>We surely made many mistakes in our whistleblowing journey. We may have construed connections with our bias. We followed dead-end leads. We used over-emotional language.</p><p>There was also a positive side: New friends, finally understood the power of FUD and that FUD must be rejected the same way we reject hate speech.<br>Learned a lot about bias<br><a href="https://en.wikipedia.org/wiki/List_of_cognitive_biases">https://en.wikipedia.org/wiki/List_of_cognitive_biases</a></p><p>We can say nothing but positive things about Hudson. Hudson is the greatest guy that could chair the all core devs. Hudson leads with integrity in an environment of horrendous intensity, like an ongoing F1 race.<br>If Hudson thinks Linzhi is not to be trusted, we can accept that. We will try to earn his respect the same way we try to earn respect from anyone.</p><p>We communicate with Hudson for 9 months now. None of us has ever met him. Our communication was 90% public, 99% one-directional (from us to him), plus some attempts to start an email conversation, basically repeating the same things we said publicly and maybe a few more speculative (to him) thoughts. I believe Hudson never replied.</p><p>He was and is faced with a very difficult situation: A team of malicious actors is attacking Ethereum, and the format of publicly recorded all-core-dev calls with a 100% engineer ratio is the perfect entry point for the attackers.<br>He wants to be neutral because Ethereum wants to be a neutral platform, but a team of collaborators cannot be neutral when faced with a predator, see Rick Falkvinge’s thoughts on that in the next section.</p><p>To maintain our own integrity, I will reply to Hudson’s tweets about Linzhi here, rather than taking it to Twitter.</p><h3>Hudson Jameson on Twitter</h3><p>@auryn_macmillan I have seen a particular ASIC company Linzhi rally against this for the last year+. Regardless of their intentions they are protecting their investment in an ETHash miner they have been public about.</p><blockquote>A: We are protecting our investment into an Ethash miner via the fight for the viability of the entire platform. The two motivations are inseparable for us.</blockquote><h3>Hudson Jameson on Twitter</h3><p>@BobSummerwill @OhGodAGirl I&#39;ve seen ASIC manufacturers publically lie about ProgPoW and other things, including Linzhi. They have an agenda too. I won&#39;t get trapped in completely being one sided due to political associations people have. The IP issues matter though.</p><blockquote>A: At some point neutrality and decision-making are mutually exclusive. If you are neutral forever you are setting up the perfect environment where a hidden agenda will succeed. Doacracy cannot work, in a doacracy the bank robber will just claim to be the CEO of the bank and get away with the robbery. Please point out what you believe our lies were so that we can reconsider the messaging or correct mistakes. We maintain a list of all substantive contributions to the debate at <a href="https://linzhi.io/">https://linzhi.io</a></blockquote><h3>Hudson Jameson on Twitter</h3><p>@fubuloubu @merkle_tree Linzhi has said a lot of stuff that was flatly untrue and pressuredme publically and privately so I stopped listening to them at the beginning of 2019 or earlier.</p><blockquote>A: There was no other known person that offered themselves as a recipient of whistleblowing-like information. We tried to contact people at the Ethereum Foundation, I believe noone ever replied. Big credits go to you for coming to our Telegram after seeing the Coindesk article. If you feel we pressured you unduly, I apologize. Same as before, please find the time to point out what you believe was flatly untrue so we can reconsider and correct. Thanks!</blockquote><h3>Hudson Jameson on Twitter</h3><p>@zooko Linzhi acted super unprofessional on their Telegram channel to myself and many others. That doesn&#39;t mean I shouldn&#39;t consider what they say, but they have a reason to be wanting to not have ProgPoW.</p><blockquote>A: Our Telegram chat from this time is archived, with permission from Hudson, in <a href="https://linzhi.io/docs/20190108-Telegram-Transcript.pdf">https://linzhi.io/docs/20190108-Telegram-Transcript.pdf</a> .<br>We were emotional, and the attackers were in our Telegram in full force. Today we have banned any and all accounts we believe to be associated with the attackers, and it’s much nicer again.<br>You are still in the group Hudson, if you have ASIC related questions — just ask.</blockquote><h3>Hudson Jameson on Twitter</h3><p>No one from Linzhi has reached out to the Ethereum Foundation or core developers to my knowledge. I&#39;d be happy to speak to them. In fact I&#39;m actively looking to talk to them. https://t.co/ecEFRtKOwM</p><blockquote>A: That is true. Before the core developer meeting January 4th we had not been in contact with core developers. Exactly as the Coindesk article says, we were shocked by the move of only working with one hardware company. That’s when we started to reach out.</blockquote><p><a href="https://www.coindesk.com/ethereum-miner-linzhi-calls-out-project-coders-for-proposed-asic-ban">Ethereum Miner Linzhi Calls Out Project Coders for Proposed ASIC Ban - CoinDesk</a></p><p>Thanks Hudson for everything you do for Ethereum! You <em>are</em> trying hard to maintain neutrality and we would be the last ones to criticize you for that.</p><p><strong>Predators and Collaborators</strong></p><p>The ProgPoW team among Ethereum developers, is a predator among collaborators.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FZkg_-_HwRcI%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DZkg_-_HwRcI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FZkg_-_HwRcI%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9200dd9fd05f90f45c7f612b6713c6f0/href">https://medium.com/media/9200dd9fd05f90f45c7f612b6713c6f0/href</a></iframe><blockquote>These predators hide among collaborators. They try to exhibit signs that make the collaborators see the predators as a collaborator, and then, when the time is right, they prey. It is not possible to imagine what these people are capable of. You cannot warn people who have not experienced a predator about what is likely to happen. Their behavior is simply inconceivable among people who are conditioned to collaborate. They do not feel embarrassment at lies. This is one defining factor. They are incapable of feeling embarrassment in the face of making outrageous lies.<br>When they are giving out outrageous lies, what they are feeling is something akin to a fisherman trying a new bait on their hook. No more.<br>People who see the signs can’t warn others, because they will be met by disbelief.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/451/1*pav3F64f6IhVMwRxiuGEKw.jpeg" /></figure><p>A collage floating around the Internet by people trying to warn others about the official ProgPoW author.</p><p><strong>Chipmakers and Collusion</strong></p><p>In the context of PoW, it is important that a chipmaker treats all customers equally. To us crypto engineers this is obvious, but to chipmakers that is a new thing. A chipmaker may initially not understand the intricacies of “global hashrate and difficulty”, and what that means in terms of cutting deals with large clients that come forward to propose a deal.</p><p>If a chipmaker shares optimizations, specialized or unsellable chips, offers delayed payments or revenue sharing to individual mining clients, they are colluding against their other customers.</p><p><a href="https://www.techspot.com/news/78548-analyst-nvidia-lied-about-cryptocurrency-earnings-avoid-stock.html">Analyst says Nvidia lied about its cryptocurrency earnings to avoid stock crash</a></p><p>The best answer to this problem in our opinion is to demand transparency and to foster competition. Currently there are 4 viable chipmakers competing in the Ethash mining market: Nvidia, AMD, Bitmain, Innosilicon.<br>With the collusion risk clearly understood, one must avoid any action that rewards monopolization efforts by a chipmaker, and instead try to motivate new entrants such as Linzhi to enter the market.<br>We think it’s good news that in addition to Linzhi there are multiple other teams trying to design competitive Ethash chips or FPGA solutions.</p><p>We wrote about the risks of the collaboration between Nvidia and Core Scientific in January, and encourage interested readers to revisit that.</p><p><a href="https://medium.com/@Linzhi/asics-and-51-achieving-mining-dominance-how-cost-advantage-drives-centralisation-cf28166d3d2a">ASICs and 51% — Achieving Mining Dominance
How cost advantage drives centralisation.</a></p><p><strong>Dark Empire and Fix</strong></p><p>Anonymous and decentralized value networks attract many people.<br>If you would be running a successful business line around stock fraud and money laundering, for example, you would see the chance to substantially reduce your operational risks.<br>Crypto mining (PoW and PoS) means cash-in, crypto-out. Cash losses become trading profits. You would want a media organization to orchestrate price swings you know about in advance.</p><p>Bob Summerwill started an important debate around anonymity of users versus anonymity of developers, and certificates of origin.</p><p><a href="https://developercertificate.org/">Developer Certificate of Origin</a></p><p>The anonymity of users in our value networks is what we want to uphold, but we must be realistic and cold-blooded, maybe even cynical, about who some of those anonymous users will be and how we can avoid that they infiltrate the culture and community of the builders of the network.</p><p>Because if they do, it’s all over.</p><p>Linzhi Team, Shenzhen, 2019-Sep-21<br>Telegram: <a href="http://t.me/LinzhiCorp">t.me/LinzhiCorp</a></p><p>About Linzhi: If you have questions about ASICs, please feel free to come to our Telegram group LinzhiCorp and ask whatever is on your mind.<br>If you are a known community member or developer with good reputation, casually feel invited to come to our office and factory in Shenzhen.</p><p>Linzhi believes that bridging the hardware-software gap is critical for crypto to succeed.<br>We don’t want crypto to be confined to a world of papers and github commits. PoW was a key innovation to connect the world of software with the world of energy and time. Hardware wallets, gas turbos, accelerators for signing, proving, verifying, sorting and other things are on the horizon.</p><p>We will continue to contribute to the Ethereum platform as a chipmaker.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/783/1*ZcMKdcKy-MJAx6iqXmqLIA.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=56132c069aa3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Linzhi E1400 Tapeout]]></title>
            <link>https://medium.com/@Linzhi/linzhi-e1400-tapeout-1688df3bb5d?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1688df3bb5d</guid>
            <category><![CDATA[asic-mining]]></category>
            <category><![CDATA[proof-of-work]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[asic-miner]]></category>
            <category><![CDATA[ethereum-classic]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Wed, 11 Sep 2019 04:01:21 GMT</pubDate>
            <atom:updated>2019-09-11T04:13:41.228Z</atom:updated>
            <content:encoded><![CDATA[<h3>Linzhi E1400 Ethash ASIC Tapeout</h3><figure><img alt="Linzhi E1400 masks released." src="https://cdn-images-1.medium.com/max/1024/1*PXXdLbzqL-DIYN2X4_t5gQ.jpeg" /></figure><p>Today we ordered the first set of wafers for our Ethash ASIC, <a href="https://www.coindesk.com/a-multi-million-dollar-bet-ethereums-proof-of-stake-isnt-coming-soon">announced almost a year ago</a> on September 13, 2018 at the ETC Summit 2018 in Seoul. This is called the chip tapeout.</p><p>The announced specs were for 1400 MHash at 1000 Watts. Pricing was announced as 4 months ROI, and the timeline was given as 12/18 for tapeout, 04/19 for samples and 06/19 for mass production.</p><p>The key points of today’s first wafer order for our new Linzhi E1400 Ethash ASIC are:</p><ul><li>Name of chip: <strong>Linzhi E1400</strong></li><li>Number of wafers ordered: <strong>37</strong></li><li>Theoretical maximum Ethash performance from ordered wafers: <strong>300 GH</strong></li><li>Theoretical maximum share of global ETC hashrate [2]: <strong>2.5%</strong></li><li>Theoretical maximum share of global ETH hashrate: <strong>0.17%</strong></li><li>Total costs up to and including first wafer order: <strong>3.85 mio USD</strong></li><li>Total people months so far: <strong>165</strong></li></ul><p><strong>FAQ</strong></p><p><strong>Q:</strong> Tapeout is delayed for 9 months. What are the reasons and what does this mean for the rest of the schedule?</p><p><strong>A:</strong> We underestimated the complexity of the chip and how long it would take to grow the team and make the company functional. We are cautiously optimistic that we can just move forward the rest of the schedule, which would mean 12/2019 for sample machines and 02/2020 for mass production.</p><p><strong>Q:</strong> Isn’t Ethereum switching to ProgPoW?</p><p><strong>A:</strong> We have not taken any pre-orders. When sales of our machines start, we will advise customers on all likely and ongoing efforts to change PoW algos in Ethash coins. Customer satisfaction is our #1 objective.</p><p><strong>Q:</strong> When are chips coming back from tapeout and what is the process of verifying their performance?</p><p><strong>A:</strong> We are hoping to have first chips back by mid-November. Then a multi-week effort known as chip bringup begins, in which we verify the different parts of the chip. The best way to follow chip bringup is our Telegram group <a href="https://t.me/LinzhiCorp">LinzhiCorp</a>, updates will also be shared via <a href="https://twitter.com/LinzhiCorp">Twitter</a>, <a href="https://instagram.com/LinzhiCorp">Instagram</a> and elsewhere.</p><p><strong>Q:</strong> What is the expected performance?</p><p><strong>A:</strong> We are still targeting the originally announced performance of 1400 MH at 1000 Watts. Acceptable outcomes would be for the power-to-hash ratio to be within 1.3x of the target, and hash performance to be within +/-15% of the target.</p><p><strong>Q:</strong> When can we see an actual machine running?</p><p><strong>A:</strong> Ideally we will have first prototype machines hashing in late-November. This will happen in our Shenzhen factory under our full control, with regular updates and AMA through social media. Our factory test environment is currently setup for a maximum load of 80 kW, which translates to 112 GH maximum Ethash hashrate.</p><p><strong>Q:</strong> Will you do chip testing on a testnet?</p><p><strong>A:</strong> We plan to do chip testing on the Ethereum Classic mainnet.</p><p><strong>Q:</strong> You are producing up to 300 GH Ethash, but sales are not open yet. Isn’t that self-mining?</p><p><strong>A:</strong> Self-mining is not part of our business model, we will try to sell 100% of the machines we are making as early as possible. However, the path to the first machine hashing at 1.4 GH is still long, and at this point the risks for customers are too unpredictable so that the machine price would be too low. We need to get the entire production to the point that performance, cost and yield are such that we can guarantee a satisfying customer experience. We will post regular transparency reports about production or testing related hashrate through social media.</p><p><strong>Q:</strong> What do you do against centralization?</p><p><strong>A:</strong> Decentralization is the entire purpose of PoW. Our sales will go to developers and community first, with a focus on geographical distribution, and potentially with a malus (reverse discount) for large orders. This means that small orders by individuals would be priced to hit the 4 month ROI and larger orders would pay more. This will offset some of the cost advantages that larger customers have and encourage decentralization.</p><p><strong>Q:</strong> What is the pricing?</p><p><strong>A:</strong> We have announced pricing at 4 month ROI and we remain committed to this target. Details may vary depending on time of order vs. shipment and efforts at decentralization, but haven’t been finalized yet. Typically Bitcoin machines are sold at 150–200 day paybacks to larger customers, and we remain committed to ensure that smaller customers should enjoy some advantage.</p><p><strong>Q:</strong> Why have you not taken pre-orders?</p><p><strong>A:</strong> The risks of an unproven design with engineering delays, unproven team with hiring delays and fast-changing crypto markets seemed too unpredictable to develop healthy customer relationships on. On the other hand we need to show that we believe in the approach ourselves, so the pressure to accept customer money remains high and is healthy.</p><p><strong>Q:</strong> What parts of the machine will you open, and what about future applications of the chip architecture?</p><p><strong>A:</strong> We will present an introduction to our chip architecture at the <a href="https://etcsummit.com/">Ethereum Classic Summit 2019 in Vancouver on October 3rd &amp; 4th</a>.We are committed to opening the entire software stack under a permissive MIT/BSD license at <a href="https://github.com/linzhichips">https://github.com/linzhichips</a> .</p><p><strong>Q:</strong> Does the USA-China trade war affect you?</p><p><strong>A:</strong> It does. Our USA customers will be affected by a 25% tariff, until we have found a way to manufacture machines outside of China. We will try to address this as early as possible.</p><p>Shenzhen, September 11, 2019<br>Linzhi Team<br>Telegram: <a href="https://t.me/LinzhiCorp">t.me/LinzhiCorp</a><br>email: <a href="/service/mailto:sonia@linzhi.io">sonia@linzhi.io</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1688df3bb5d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[EIP 1057 (ProgPoW): Open Chip Design for only 1% cost/power increase
EIP 1057 (ProgPoW)…]]></title>
            <link>https://medium.com/@Linzhi/eip-1057-progpow-open-chip-design-for-only-1-cost-power-increase-eip-1057-progpow-d106d9baa6eb?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/d106d9baa6eb</guid>
            <category><![CDATA[progpow]]></category>
            <category><![CDATA[proof-of-work]]></category>
            <category><![CDATA[chip-design]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[asics]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Fri, 29 Mar 2019 12:31:18 GMT</pubDate>
            <atom:updated>2019-03-29T12:36:23.908Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>EIP 1057 (ProgPoW): Open Chip Design for 1% cost/power increase <br>​仅带来1%能耗比提升的开源芯片设计</strong></p><p>Adding a pseudo-random program to the PoW algorithm is a key technical idea described in EIP 1057 (ProgPoW). The theory goes that the compute area of a GPU is underutilized compared to memory bandwidth. ProgPoW is then designed such that it “saturates both compute and memory bandwidth at once”.</p><p><a href="https://github.com/ethereum/EIPs/blob/master/EIPS/eip-1057.md">https://github.com/ethereum/EIPs/blob/master/EIPS/eip-1057.md</a></p><blockquote>Since the GPU is almost fully utilized, there’s little opportunity for specialized ASICs to gain efficiency.<br>There should be little opportunity for efficiency gains compared to a commodity GPU.<br>While a custom ASIC is still possible, the efficiency gains available are minimal.<br>These would result in minimal, roughly 1.1x-1.2x, efficiency gains.</blockquote><p>Bitmain and Innosilicon both currently have Ethash systems, a good starting point for ProgPoW. What would they need to add for the new EIP 1057 compute logic?</p><p><strong>Math Block</strong></p><p>First let’s look at Math() which includes 11 different instructions</p><pre>// Random math between two input values<br>uint32_t Math(uint32_t a, uint32_t b, uint32_t r)<br>{<br>    switch (r % 11)<br>    {<br>        case 0: return a + b;<br>        case 1: return a * b;<br>        case 2: return mul_hi(a, b);<br>        case 3: return min(a, b);<br>        case 4: return ROTL32(a, b);<br>        case 5: return ROTR32(a, b);<br>        case 6: return a &amp; b;<br>        case 7: return a | b;<br>        case 8: return a ^ b;<br>        case 9: return clz(a) + clz(b);<br>        case 10: return popcount(a) + popcount(b);<br>    }<br>}</pre><ul><li>modulo 11 operation, this is quite small logic, ~400gate, 1ck latency, can be a parallel process during mix read so we can hide this latency.</li><li>32-bit Add, simple logic, ~300gate for a fast one.</li><li>32-bit Multiplier, mature IP, ~20Kgate for a fast one, since multiplier only have ~4/11 activity rate, we can use a two cycle multiplier to half the area, small possibility to increase delay.</li><li>Rotation operation can easily map to a multiplier, for example I want to calculate ROTL(0x12345678, 8), I can do 0x12345678 * 0x00000100 = 0x0000001234567800, then we just need to OR higher word and lower word together to get 0x34567812. so just cost ~160gate extra logic</li><li>logic operation, A&amp;B only cost 32 gate, A|B 32 gate, A^B 96 gate, it looks like three different instructions but actually extremely small on silicon (&lt;30um²)</li><li>clz and popcount are also very small</li><li>We only need a multiplexer to select output.</li><li>Total size of Math() is about 0.0015mm² on a TSMC16ULP process.</li><li>Merge() is similar but even smaller, only shifter, adder, and tiny logic (no multipliers because constant multiply can be mapped into adder).</li><li>Size of Merge() is roughly ~0.0005mm².</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MigehHQ57q4G2Q4TtHxdfg.jpeg" /></figure><p><strong>Pipeline</strong></p><p>Now we can build our pipeline, 4-stage:</p><p>CK1: Mix regfile read, two operators in parallel (calculate %11 at same time)<br>CK2: Math()<br>CK3: Merge()<br>CK4: Mix write back to regfile</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5BfXAOd1iwHJqWKa2gqXQA.jpeg" /></figure><p>Task latency is 4 cycles, but we can have 4 independent threads so we can make this pipeline fully loaded.</p><p>Mix Regfile we use a 1W2R (1 write port, 2 read port) Regfile, mature IP, 8KB (for 4 threads), single cycle read/write, 1GHz operation.</p><p>If you want load/unload mix data on the fly without disturbing the pipeline, we can upgrade to a 12KB 2W3R Regfile. We then have extra read/write ports for load/unload, and 12KB is enough for 6 tasks (4 running, 1 loading, 1 unloading).</p><p>An ASIC can implement 10K sets of that block:</p><ul><li>running at 1GHz frequency</li><li>0.55V voltage (typical voltage for TSMC16ULP)</li><li>generating ~10T Math() + Merge() throughput per second</li><li>Power estimate roughly 3mW each pipeline, 30W in total.</li><li>No customized circuit/layout</li><li>All standard cells, auto placement route</li><li>Use mature IP only</li><li>No aggressive overclocking</li><li>No aggressive under voltage</li><li>~8.4K Merge() and ~4.8K Math() per hash</li><li>pipeline includes 1 Merge() and 1 Math()</li><li>pipeline is only the logic part, not the memory part</li><li>the xGB memory are not included</li><li>we have 10T throughput on single chip asic, divided by 8.4K Merge() per hash, means 1.2GHash</li></ul><p><strong>Performance</strong></p><p>The design (logic part only) can provide 1.2 GHash ProgPoW performance at 30W power.<br>Nvidia GTX1070Ti can provide 15.7 MHash at 115W, but including memory.</p><p><strong>Summary</strong></p><p>The random instruction of EIP 1057 increases die cost/power by about 1%, and causes a die increase of &lt;1mm². The proposed open design is demonstrating a logic-only performance of 1.2 GHash at 30W and could be deployed by Bitmain or Innosilicon, resulting in a machine with about half the hashrate of their predecessors, similar to the best GPUs.</p><p>If you are an independent chip designer and want to take the ProgPoW design idea further, please get in touch. We hereby release the design into the public domain. Hopefully the open design process can also inspire some developers to learn more about the world of chip design.</p><p>Linzhi Team, Shenzhen<br>Telegram: <a href="https://t.me/LinzhiCorp">https://t.me/LinzhiCorp</a><br>email: <a href="/service/mailto:sonia@linzhi.io">sonia@linzhi.io</a></p><p><strong>Thanks</strong></p><p>Peter Salanki provided the initial idea for this writeup and encouraged us to do it. Alexey Akhunov always encourages everyone to think for themselves.</p><p><strong>References</strong></p><p><a href="https://github.com/ethereum/EIPs/blob/master/EIPS/eip-1057.md">https://github.com/ethereum/EIPs/blob/master/EIPS/eip-1057.md</a><br><a href="https://medium.com/@ifdefelse/understanding-progpow-performance-and-tuning-d72713898db3">https://medium.com/@ifdefelse/understanding-progpow-performance-and-tuning-d72713898db3</a><br><a href="https://medium.com/@infantry1337/comprehensive-progpow-benchmark-715126798476">https://medium.com/@infantry1337/comprehensive-progpow-benchmark-715126798476</a><br><a href="https://etcsummit.com/2018-etc-summit/">https://etcsummit.com/2018-etc-summit/</a><br>Launch of Linzhi Ethash chip at ETC Summit: <a href="https://www.youtube.com/watch?v=LMofyroBfio">https://www.youtube.com/watch?v=LMofyroBfio</a><br>Linzhi Website: <a href="https://linzhi.io/">https://linzhi.io/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d106d9baa6eb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Studying the Feasibility of an ASIC]]></title>
            <link>https://medium.com/@Linzhi/studying-the-feasibility-of-an-asic-82634ac77d61?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/82634ac77d61</guid>
            <category><![CDATA[semiconductors]]></category>
            <category><![CDATA[asic-mining]]></category>
            <category><![CDATA[asic-design]]></category>
            <category><![CDATA[crypto]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Mon, 11 Mar 2019 10:52:21 GMT</pubDate>
            <atom:updated>2019-03-11T20:26:04.641Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Can you make an ASIC for algorithm X?<br>How fast can it be?<br>Can it be better than Y?</blockquote><p>Today we want to document the 5 steps that we typically think through, roughly, to answer such questions.<br>We intentionally leave out considerations for platform, programmability and software stack, industry trends or competition. This article is about the technical feasibility of a chip design for a given algorithm.</p><h3><strong>1. Math First</strong></h3><p>If we have an algorithm’s math description, we start from that. If we only have source code, we need to convert it to a math description first.<br>The reason for going to math first is that hardware may use a fundamentally different way to compute an algorithm.<br>Example task: Search how many data items X exist within a database P and locate them. Software may need a data structure, maybe a hash table, and many memory accesses. Hardware may be able to produce all results in one clock cycle using a CAM (content-addressable memory).<br>Many things like search and sort can have a different implementation in hardware.</p><h3><strong>2. Optimization Target</strong></h3><p>We need to clarify the goal of the optimization. Optimize for high throughput? For low latency? Low power? Low energy? Low cost? Fast time-to-market? Security? High reliability?<br>You cannot get all of them at once, so you have to pick priorities. For example in a Bitcoin miner, 1st is low energy, 2nd high throughput, 3rd low cost.<br>After this phase you have constraints, such as budget, target performance, maximum power, target delivery time, etc.</p><h3><strong>3. Hardware-Software Boundary</strong></h3><p>An algorithm may not benefit from a full hardware implementation. We need to decide which part to run in software, and which to run in hardware, after careful algorithm study.<br>Normally hardware will handle the major data flow for high-performance. Software can handle complex protocols. However if analysis shows the protocol processing to be the performance bottleneck, that part will also move to hardware.</p><h3><strong>4. Building Blocks</strong></h3><p>There are three major kinds of building blocks (called IP): logic, memory, IO. Each has nearly infinite types of them available.</p><p>For example let’s look at one of the smallest IP’s: an integer adder, to sum up two numbers. For smallest area we choose a Ripple-Carry adder, for lowest latency we choose Kogge-Stone. There are many others: carry-skip, carry-lookahead, Brent-Kung, Han-Carlson, conditional sum, Sklansky, Ling, Jackson, etc. You need to carefully choose because each one has a different PPA (power, performance, area).<br>The algebraic logic on non-critical paths can normally be optimized automatically by EDA tools.</p><p>Same for memory: First there is on-chip memory called SRAM. SRAM is fastest but also most expensive, and there are many kinds of SRAM.<br>eDRAM is much lower cost but very limited. eFlash is cheap but very slow, there is MRAM, RRAM, Regfile, etc.<br>Off-chip memory: PSRAM, DDR, LPDDR, QDR, RLDRAM, GDDR, HBM, HMC, NAND Flash, NOR Flash, FRAM, PCM, …</p><p>Based on project contraints, you can easily rule out most of them. For the rest you need to contact the vendor to sign an NDA, get detailed information, and price quote. Then you decide what IP and peripheral to use, based on trade-off of cost, performance, power, etc.<br>Like the Nvidia CEO said “It’s much more expensive for consumer goods. But storage prices will fall. Everything is fine with HBM. I love HBM. But I love GDDR6 (graphics memory) so much more.”<br><a href="https://optocrypto.com/ceo-of-nvidia-says-gddr6-is-cheaper-than-hbm-for-choosing-next-gen-gpu-memory/">https://optocrypto.com/ceo-of-nvidia-says-gddr6-is-cheaper-than-hbm-for-choosing-next-gen-gpu-memory/</a><br>They made a trade-off for cost.</p><p>If no good IP can be found for the target application, we may ask a vendor to customize an IP for our application. Or we can design it ourselves from the transistor level. In Bitcoin miners you find a large ratio of fully custom designed circuit blocks such as dynamic flip-flops.</p><h3><strong>5. Physical Implementation</strong></h3><ul><li>Foundry</li><li>Process node</li><li>Floor plan</li><li>Simulation</li><li>Clocking</li><li>Power distribution network</li><li>Timing corner</li><li>Signal integrity</li><li>Temperature range</li><li>Process variation</li><li>Aging</li><li>Pinout</li><li>Packaging</li><li>Thermal management</li><li>Mechanical reliability</li><li>ESD protection</li><li>Verification</li><li>Tapeout</li><li>Process Tuning</li><li>Tester</li><li>Binning/Grade</li><li>PCB</li><li>Assembly</li><li>Software integration</li><li>Burn-in</li><li>Shipping</li></ul><p>Done :)</p><p>Chen Min<br>Linzhi Shenzhen<br>Telegram: <a href="https://t.me/LinzhiCorp">https://t.me/LinzhiCorp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82634ac77d61" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ECIP 1043 Revisited]]></title>
            <link>https://medium.com/@Linzhi/ecip-1043-revisited-cd2e500b8550?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/cd2e500b8550</guid>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[ethash]]></category>
            <category><![CDATA[asic-mining]]></category>
            <category><![CDATA[crypto]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Sat, 09 Mar 2019 20:42:11 GMT</pubDate>
            <atom:updated>2019-03-09T20:44:50.545Z</atom:updated>
            <content:encoded><![CDATA[<p>ECIP 1043 proposed to set a limit on the maximum size of the DAG, and no longer increase it on epoch schedule.<br><a href="https://github.com/ethereumproject/ECIPs/pull/91">https://github.com/ethereumproject/ECIPs/pull/91</a></p><p>ECIP 1043 was closed by its author Cody Burns in December 2018, but today we take another look at motivation and specification, from our perspective of being a Shenzhen asicmaker (with upcoming Ethash chip).</p><p>Motivation (from the ECIP):</p><ul><li>DAG has failed at ASIC Resistance and now serves as deterrent to broader investment in ASICs</li><li>increasing DAG has negative impact on security and acts as anti-competitive bug preventing investment in mining</li><li>increasing DAG will eventually reach limit of GPU bus speed</li><li>smaller DAG allows more GPUs to mine while lowering entry cost</li></ul><p>Specification (from the ECIP):</p><ul><li><em>if(blkNumber &gt; forkblock) DAG epoch = 0</em></li></ul><p><strong>ASIC Resistance</strong></p><p>We see ASIC Resistance as an economic concept, it increases the cost of ASICs. As such the ASIC Resistance of Ethash has worked.<br>Maybe that was an unintended consequence, but unintended consequences of major decisions are inevitable and now it seems best to focus on learning instead of reacting with sudden changes that may lead to new unintended consequences.<br><a href="https://hackernoon.com/asic-resistance-is-nothing-but-a-blockchain-buzzword-b91d3d770366">https://hackernoon.com/asic-resistance-is-nothing-but-a-blockchain-buzzword-b91d3d770366</a><br>Keep calm and hash on.</p><p>We don’t think the DAG as it is growing in the current protocol (ca. 50 MB/month, 600 MB/year) creates sudden shocks when reaching any GB threshold, even the currently widely used 4 GB (obsolete 2020–11–30).<br>There is predictable obsolescence for 10 GB and up into the 2030&#39;s.<br><a href="https://investoon.com/tools/dag_size">https://investoon.com/tools/dag_size</a></p><p>Ethash has served well since 2015 in creating a stable and predictable mining environment, including the emergence of the Bitmain E3 and Innosilicon A10, as well as upcoming chips like ours or others.</p><p><strong>Current Mining Landscape</strong></p><p>Mining profitability (currently 1.3 cents/MH/day) should be looked at separately from PoW algorithm study. There are many mining business models, welcome ones (using peak wind energy at negative power rates, heaters), unwelcome ones (money laundering), and more.</p><p>Our estimates for the current Ethash mining memory distribution are<br>(obsolescence dates for ETC from investoon.com, ETH is comparable):</p><ul><li>2019–03–17 <strong>4%</strong> GPU <strong>3GB</strong></li><li>2020–11–30 <strong>10%</strong> Bitmain E3 <strong>4GB</strong></li><li>2020–11–30 <strong>36%</strong> GPU <strong>4GB</strong></li><li>2024–04–30 <strong>36%</strong> GPU <strong>6GB</strong></li><li>2027–09–29 <strong>9%</strong> GPU <strong>8GB</strong></li><li>2030+ <strong>5%</strong> GPU <strong>10GB</strong> and up</li></ul><p><strong>Objectives</strong></p><ul><li><strong>Predictability<br></strong>Predictability is connected to fairness, competition, stability and even recyling, and looks like the most important consideration to us. Predictability creates fairness. Shocks and nonlinear benefits should be minimized.</li><li><strong>Cost Advantages<br></strong>Decentralization is the result of three goals mainly — to have a system that is trustless, permissionless and resilient.<br>From a hardware perspective, avoiding centralization means avoiding persistant cost advantages. Thorough cost analysis must be standard practice.<br>What works against this is that you are never sure there is not some unknown way to be faster or cheaper. Calculations are based on one’s own assumptions and may not apply to others. We can calculate the limitations of things we know, but we cannot prove that there cannot be something better we don’t know.</li><li><strong>Competition<br></strong>There are currently 4 vendors for Ethash (Nvidia, AMD, Bitmain, Innosilicon), a 5th one announced (Linzhi). There may be other startups preparing for solutions in the future, but the current economics make investment difficult for everyone.</li><li><strong>Recycling<br></strong>We think existing miners should not be bricked frivolously. PoW changes should consider input from chip designers in how the resulting chips can be more useful to scale crypto applications beyond just PoW. Usefulness and security are not mutually exclusive.</li><li><strong>Immutability<br></strong>Every hardfork, and every feature in a hardfork, threatens the promise of immutability and should go through a rigorous process — ECIP.</li></ul><p><strong>Technical Considerations</strong></p><p><a href="https://github.com/ethereum/wiki/wiki/Ethash#definitions">https://github.com/ethereum/wiki/wiki/Ethash#definitions</a></p><ul><li>seed -&gt; cache -&gt; DAG</li><li>cache is DAG/64</li><li>current DAG is ‘relatively’ dynamic, the entire 3GB changes every epoch (4–5 days), and grows by approx. 8 MB/epoch</li><li>DAG growth approx. 50 MB/month, 600MB/year</li><li>ROM is 3x cheaper than RAM</li><li>constant data encourages over-specialized development that is technologically a dead-end</li><li>SRAM and eDRAM need to be considered</li><li>maximum chip size is, generally, 26mm x 33mm</li><li>we assume it would be easy for Bitmain to come out with a 6GB E3</li><li>we assume it would be easy for Innosilicon to come out with an 8GB A10</li></ul><p>The ECIP 1043 pseudo-code does not clearly specify whether the entire DAG becomes a constant, or whether the cache keeps being modified and the 1GB DAG changes every epoch. We assume that the author did not unintentionally want to create a constant DAG, and an actual implementation would have continued to modify the DAG every epoch, just not grow it anymore. A constant 1GB DAG could allow for single-die ROM miners.</p><p>Even a dynamic DAG at 1GB could open the possibility of an ASIC beating a GPU by 10x. 16nm or deeper processes allow integrated 1GB RAM on a single silicon die. On-chip SRAM is &gt;100x faster than GDDR or HBM.<br>Since Ethash loops are mostly memory accesses and few logic ops, it means the memory bandwidth bottleneck would be completely removed in this case.</p><p><strong>Summary</strong></p><ul><li>We think allowing the DAG to grow at the current approx. 8 MB / epoch (50 MB / month) rates high on predictability, and thus stability, fairness and security. This includes obsolescence of 4 GB GPUs (and Bitmain E3) in December 2020.</li><li>If people believe DAG growth should be limited, we propose 3 GB because that is proven to be stable right now. We also think it’s important that the data itself remains dynamic (changes with every epoch), and does not become one large constant. ROM is 3x cheaper than RAM.</li><li>To increase predictability and fairness, we propose a significant interval of one year between hardfork and activation of DAG fixing:<br><em>if(blkNumber &gt; forkblock + 2200000) DAG epoch = 250</em></li></ul><p>If there are no flaws in our analysis or reasoning, something is seriously wrong! If you spot mistakes or just plain disagree, please let us know.</p><p>Linzhi Shenzhen<br>Telegram: <a href="https://t.me/LinzhiCorp">https://t.me/LinzhiCorp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cd2e500b8550" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Bridging the Hardware-Software Divide: 4 More Claims.]]></title>
            <link>https://medium.com/@Linzhi/bridging-the-hardware-software-divide-4-more-claims-5b520ae00c0c?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b520ae00c0c</guid>
            <category><![CDATA[asic-mining]]></category>
            <category><![CDATA[open-hardware]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[proof-of-work]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Sun, 03 Feb 2019 21:48:53 GMT</pubDate>
            <atom:updated>2019-02-03T22:03:19.710Z</atom:updated>
            <content:encoded><![CDATA[<p>Linzhi Ltd. is a private fabless semiconductor company in Shenzhen.<br>On January 8th, we called upon Ethereum developers to publish rules and guidelines for what constitutes a good ProgPoW ASIC maker.<br><a href="https://www.coindesk.com/ethereum-miner-linzhi-calls-out-project-coders-for-proposed-asic-ban">https://www.coindesk.com/ethereum-miner-linzhi-calls-out-project-coders-for-proposed-asic-ban</a></p><p>Today Alexey Akhunov published a wonderful opinion piece with 4 claims, which we see as a massive effort to bridge the hardware-software divide.<br><a href="https://medium.com/@akhounov/my-take-on-progpow-looking-for-a-reasonable-answer-to-a-reasonable-question-9f13ffd4e9d7">https://medium.com/@akhounov/my-take-on-progpow-looking-for-a-reasonable-answer-to-a-reasonable-question-9f13ffd4e9d7</a><br>The Qi Hardware community provided a Chinese translation of Alexey’s claims:<br><a href="https://qihardware.org/2019/02/03/my-take-on-progpow-chinese.html">https://qihardware.org/2019/02/03/my-take-on-progpow-chinese.html</a></p><p>To respond, we first quote Chris Ziomkowski, founder of XTend.online (unaffiliated with us).</p><p>The reasonable question was: “<strong>What should chipmakers do to not be labeled bad actors after designing a crypto (here: PoW) chip?</strong>”</p><p>We quote Chris Ziomkowski of XTend.online (unaffiliated with us but we love his answer) for our reasonable answer: “<strong>Commit to the cause of making equipment that is ubiquitous and available to anyone on an egalitarian and non preferential basis. Accept competition from others in finding the best way to secure the network. Be patient.</strong>”</p><p>However we also liked Alexey’s open-minded claim format so much, we take the liberty to add another set of 4 claims on top of his. If you believe Alexey’s claims to be untrue, the next 4 might be even harder to prove. If you believe Alexey’s claims are true, they may be easier. Let us know.</p><h3>Claim 1: A change in PoW algorithm, or other reward algorithms in the future, must be preset to a fixed date at least 1, better 2 years in advance.</h3><p>The highest principle in reward algorithms must be enabling competition.<br>Free and fair competition is an extension of open-source principles into hardware. Hardware requires capital and time. This means software developers have a responsibility to commit to a fixed date for the rollout of any reward algorithm in advance, long enough to enable anyone who wants to to prepare.<br>Equality of opportunity must be achieved. For simple hardware, 6 months or 12 months may be enough. Even for the most complicated hardware, 2 years should suffice.</p><h3>Claim 2: Open-source principles, although universally loved in crypto, do not apply to hardware.</h3><p>Open does not mean open and honest. Open cannot extend well to schedules and wallets. Business negotiations and contracts are not open in many cases. One of the founding principles of open-source was that there is a mandate for code to be open, because it can be open. A bowl of spaghetti however, can only be eaten once.<br>The hardware world cannot make this any easier for open-source software developers. We cannot wait until the whole world becomes an open-source program. Software developers have to accept that open-source is a concept that is inherently intangible, and need to make the extra effort to define what exactly they want from the tangible hardware world.</p><h3>Claim 3: For crypto to succeed, bridging the hardware-software gap is critical.</h3><p>We don’t want crypto to be confined to a world of papers and github commits. PoW was a key innovation to connect the world of software with the world of energy and time. Hardware wallets, gas turbos, accelerators for signing, proving, verifying, sorting and other things are on the horizon. Environmental impact is real, safe and fair working conditions are real. The hardware-software divide is a big challenge, but also one of the biggest opportunities in crypto in the near future. We don’t believe this problem will go away with the promised switch to PoS or even an entire retreat from all things physical.<br>Open hardware projects like Qi Hardware may help.</p><h3>Claim 4: Contributing to the network must be more economic than attacking the network.</h3><p>Many of the problems we have with today’s PoW come from bad alignment of incentives. It shouldn’t matter that much whether all participants share the exact same goals and values, and whether their investment horizons are short-term or long-term. Protocol developers have a lot of power in structuring the rewards — issuance, monetary policy, fees. There must be no way to capitalize on inefficiencies. Devs need to understand the very real physical consequences of their economic models. Linzhi believes in a future of Proof-of-Contribution: Crypto projects will reward contribution to the actual scaling of the network: verifying transactions, STARK provers, signing, sorting, block speed, etc.<br>Contributing to the network has to be more economic than attacking the network. That’s the principle software developers have to bring to life with algorithms and protocols, so hardware and software can grow together and in harmony.</p><p>Never stop learning.<br>Linzhi Shenzhen<br>Telegram: <a href="https://t.me/LinzhiCorp">https://t.me/LinzhiCorp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b520ae00c0c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Our Telegram, January 7 to 29, 2019 — Lessons Learned]]></title>
            <link>https://medium.com/@Linzhi/our-telegram-january-7-to-29-2019-lessons-learned-cb5a1fc0f039?source=rss-5f5a9e27955b------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb5a1fc0f039</guid>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[progpow]]></category>
            <category><![CDATA[von-neumann]]></category>
            <category><![CDATA[open-hardware]]></category>
            <category><![CDATA[social-contract]]></category>
            <dc:creator><![CDATA[Linzhi ASICs]]></dc:creator>
            <pubDate>Fri, 01 Feb 2019 13:58:21 GMT</pubDate>
            <atom:updated>2019-02-03T21:58:11.512Z</atom:updated>
            <content:encoded><![CDATA[<h3>Our Telegram, January 7 to 29, 2019 — Lessons Learned</h3><p>On January 8, 2019, Linzhi Ltd, a private fabless semiconductor company in Shenzhen, called upon Ethereum developers to publish rules and guidelines for what constitutes a good ProgPoW ASIC maker.</p><p><a href="https://www.coindesk.com/ethereum-miner-linzhi-calls-out-project-coders-for-proposed-asic-ban">https://www.coindesk.com/ethereum-miner-linzhi-calls-out-project-coders-for-proposed-asic-ban</a></p><p>Little did we know that an intensive debate between many stakeholders would ensue, and since we now feel that has come to a conclusion, we want to summarize lessons learned, our takeaway. Biased.</p><p><strong>Ethereum Governance</strong></p><p>Ethereum is a wonderful project, decentralized, romantic. Anyone can be a developer, pseudonymous and anonymous members are welcome. Motives or second agendas are not questioned in crypto utopia.<br>Polkadot, cat herders, magicians, Vitalik, core devs, Ethereans, ProgPOW, classicists. We are all Ethereum.</p><p><strong>ProgPOW</strong></p><p>When we first read “technical papers” from ProgPOW supporters that didn’t add up, we knew something was different, but just what?<br>Today we know that ProgPOW follows the energists philosophy (more later), in line with their affiliates and business practices we have seen from Coingeek, Craig Wright, nChain, etc. There may be truth in this approach, but it’s not our truth.<br>We think ProgPOW is a bad idea not only because of motives and who will win or loose, but because the following values will be impaired:</p><ul><li>Technical Excellence: ASIC resistance not much different from Ethash.</li><li>Equality: Chinese miners will be bricked, creating waste.</li><li>Environmental Responsibility: Power consumption of ETH network seems to be going up, there is limited data on the subject.</li><li>Open Source: Open source audit cannot see power consumption, cannot see which existing chips favored, cannot see how hard or easy to design new chips.</li><li>Community Consensus: Process of PoW change in general (for any coin not just Ethereum) is not thought through, many problems are not identified or understood.</li><li>Decentralization: Verification slowed by unknown factor, asymmetry between proof and verification helps decentralization.</li></ul><p><strong>Lipstick on a Pig</strong></p><p>Aside from copious proofs of dirty backroom deals, we got some amazing anti-ProgPOW artworks from the community. Our favorite is this poem. Noone can make this up, thank you Anonymous!</p><blockquote>The ProgPOW initiative is teaching us to return to incandescent light bulbs. There’s no reason to improve energy efficiency beyond what GPU technology provides.</blockquote><blockquote>The ProgPOW initiative is teaching us there’s no reason to spend time advancing crypto mining technology. GPUs are as good as we’re allowed to have.</blockquote><blockquote>The ProgPOW initiative is teaching us that GPUs are god, and the GPU god sets the rules, and the GPU god favors his acolytes. Officers and investors in a mine who are heavily invested in GPU equipment and have insider connections to the GPU god plus influence in the Ethereum core development team write a protocol that favors their mine revenues and GPU makers.</blockquote><blockquote>The ProgPOW initiative is teaching us about arrogance. Those revolutionary advances in memory technology and semiconductor fabrication now heading for commercialization belong to the GPU god. Shh! Don’t tell them that those advances are going to render PropPOW worthless in a few years.</blockquote><blockquote>The ProgPOW initiative is teaching us that open source doesn‘t mean full transparency. Who originally developed and wrote the protocol? Who underwrote the original effort? Who’s been promised some future benefits? Did those involved fully disclose their self-interests? And what’s the real motive for this initiative?</blockquote><blockquote>The ProgPOW initiative is teaching us how socialism really works. Anyone who attempts to excel at mining is penalized until everyone mines with equally poor performance. Individual rights are subjugated to the crypto-state controlled by the privileged few.</blockquote><blockquote>The ProgPOW initiative is teaching us about hypocrisy. The developers of ProgPOW proclaim the socialist mantra of “fairness” while using the “fairness” algorithm to neuter potential competition that threatens their capitalist enterprises.</blockquote><blockquote>The ProgPOW initiative is teaching us again how the “fairness” mantra appeals to the masses whose educations didn’t include a history of the “fairness” movement. Only later do they discover that the price was their freedom. In this case, that freedom was to mine and otherwise prosper but those rights are now subjugated to the crypto-state controlled by the privileged few. Perhaps, advocates should consider the millions of people protesting and fleeing Venezuela to find freedom.</blockquote><blockquote>The ProgPOW initiative is teaching us a socialist subjugation tactic — spread “fairness” to neighboring crypto-currencies just in case you think about fleeing.</blockquote><blockquote>The ProgPOW initiative is teaching us an ugly anti-competitiveness tactic — secure adoption while the protocol specification remains a bit fluid and hint that further adjustments may be necessary. Is the intent to permanently discourage FPGA and ASIC development efforts? And who determines when adjustments are needed? The GPU god? An anointed acolyte?</blockquote><blockquote>The ProgPOW initiative is teaching us how to foster a legal quagmire for monopolistic restraint of trade practices that invite government intervention.</blockquote><blockquote>The ProgPOW initiative is teaching us to wonder why members of the core development team have not raised the issues above. Is this how they want to be remembered when the history of Ethereum is written? And that history may be much briefer because of the issues above.</blockquote><blockquote>The ProgPOW initiative is teaching me to look for another blockchain platform for a major program.</blockquote><blockquote>Anonymous</blockquote><p><strong>Chipmaker Social Contract</strong></p><p>It looks like manufacturer-neutral open hardware forums might pick this up (see Qi below), but the essence of our ideas for chipmaker guidelines to enable competition in the crypto era are (thanks to many who contributed):</p><ul><li>cost transparency</li><li>production transparency</li><li>sell to developers and community first</li><li>volume discounts topped out at fixed percentage</li><li>recycling program, protection against loss of value, bricking</li><li>no self-mining, sell to independent customers only</li><li>transparency in developer relations and funding</li><li>qualifying customers, no sales to ponzi, mlm, scammers, etc.</li></ul><p><strong>ASIC Resistance</strong></p><p>.. is a state of mind. MoneroCrusher’s article and our response remain our memory.</p><p><a href="https://medium.com/@MoneroCrusher/opinion-asic-resistance-is-a-state-of-mind-not-something-you-can-achieve-with-a-final-software-abdc878a50aa">https://medium.com/@MoneroCrusher/opinion-asic-resistance-is-a-state-of-mind-not-something-you-can-achieve-with-a-final-software-abdc878a50aa</a></p><p><a href="https://medium.com/@Linzhi/its-the-unknown-we-fear-response-to-monerocrusher-f7885bea5983">https://medium.com/@Linzhi/its-the-unknown-we-fear-response-to-monerocrusher-f7885bea5983</a></p><p><strong>Open Source</strong></p><p>.. is the over-arching value guideline. Unfortunately proven open source principles reach limits when dealing with energy and time. No amount of code study will tell you much about power consumption, business motives, hardware costs, etc.<br>We hope the discourse of how to extend open source principles into hardware will be continued. Open source principles are great!</p><p><strong>Hobby Mining</strong></p><p>.. is the desire of many, and we have too few answers for that. Mining needs to be FUN again. Mining should be possible with a 200 USD investment, with hardware that can easily be bought, and if mining doesn’t work out, can be sold again. Even with an unsellable GPU card you can still play Fortnite!</p><p><strong>GPUs for PoW</strong></p><p>.. are a bad idea in the long run. PoW is a cryptographically and thermodynamically provable security mechanism. To understand why GPUs are a PoW dinosaur to mainly benefit Nvidia/AMD insiders and large farms, we propose reading Chris Ziomkowski’s wonderful article.<br><a href="https://blog.usejournal.com/inside-the-new-crypto-mining-technology-that-will-redefine-the-industry-196529547c88">https://blog.usejournal.com/inside-the-new-crypto-mining-technology-that-will-redefine-the-industry-196529547c88</a></p><p>In our view it’s not ASICs that lead to centralization, it’s cost advantage.<br><a href="https://medium.com/@Linzhi/asics-and-51-achieving-mining-dominance-how-cost-advantage-drives-centralisation-cf28166d3d2a">https://medium.com/@Linzhi/asics-and-51-achieving-mining-dominance-how-cost-advantage-drives-centralisation-cf28166d3d2a</a></p><p>A major reason the computing industry evolved around the Von Neumann architecture is that it is easy to develop software for. The instruction set is intuitive for humans to understand.</p><p>Imagine how difficult it would be to compile an algorithm to Conway’s Game of Life hardware.<br><a href="https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life">https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life</a></p><p>There are thousands of computer architectures in history. The Von Neumann bottleneck does not apply to other computer architectures. Most of them never succeeded commercially although it is not difficult to actually build that hardware. Without a language, compiler, debugger or development method, these kinds of hardware can never be accepted by the industry.<br>Mining is a different market. Mining hardware mostly sells its service directly to the network, instead of selling to people who need to do development for the hardware. The network will reward the correct result automatically, but it cannot verify which implementation is providing the result. You could use paper and pencil to mine Bitcoin.</p><p>This opens the possibility of selling special hardware without programming language and development methods. Implementing one algorithm on special hardware is easy, but creating a toolchain or design method is much more difficult.</p><p>ProgPOW is still one algorithm, even if it has a pseudo-random sequence inside, it’s still a single algorithm.<br>ProgPOW’s parameters are optimized for a GPU, assuming the performance upper bound is limited by memory bandwidth, and the cost lower bound limited by (pseudo-random) logic requirements.<br>It will prove GPU limits, but will never prove that ASICs need to use the same implementation.</p><p><strong>Upcoming Ethash Miner</strong></p><p>.. is on track, maybe with a small delay, but we enjoy working on this chip and bringing more security to the ETC network and anyone else who wants Ethash security.</p><p><strong>Proof of Contribution</strong></p><p>.. is our vision for the way forward. Making contributing to the network more economic than attacking it. Making chips not just for PoW, but for signing, verifying, proving, sorting, block speed, other things. The biggest obstacle for any forecastable chip demand is lack of user adoption. Self-referential mining games with ever more sophisticated insider plays will end.<br>Our attention will go to hardware that enables financial liberty, user adoption.</p><p><strong>Energists</strong></p><p>ENERGISTS will always get energy as fast and efficiently as possible. #1 rule of ENERGISTS is to get energy for self. Golden Rule and other rules of humankind are thrown out or used in service of the needs of ENERGY gathering.<br>That is why the words do not add up if you use the old ways of pre-crypto thinking.<br>If you want to understand the ENERGIST, you must think like the ENERGIST.<br>Better to think like a plant or a tree, then the words will make sense…</p><p>Thanks to Jon Phillips.</p><p><strong>Qi Hardware</strong></p><p>Jon proposed Qi Hardware as a forum for open hardware collaboration, we will try to contribute as much as we can.<br><a href="https://qihardware.org/">https://qihardware.org/</a><br>Open source (MIT/GPL) code and documentation for Linzhi chips will be published at <a href="https://github.com/LinzhiChips">https://github.com/LinzhiChips</a> (there is nothing yet).</p><p>That was a fun month of January for us, thanks a lot to everyone who contributed and we hope others learned many things as well.</p><p>Never stop learning.<br>Linzhi Shenzhen<br>Telegram: <a href="https://t.me/LinzhiCorp">https://t.me/LinzhiCorp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cb5a1fc0f039" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>