<?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 Perry Stephenson on Medium]]></title>
        <description><![CDATA[Stories by Perry Stephenson on Medium]]></description>
        <link>https://medium.com/@perrystephenson?source=rss-9c1d39a3c2a1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*CRL0J7KdjgUv-CzZ.</url>
            <title>Stories by Perry Stephenson on Medium</title>
            <link>https://medium.com/@perrystephenson?source=rss-9c1d39a3c2a1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 02:31:27 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@perrystephenson/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[Building a cost-effective Deep Learning dream machine]]></title>
            <link>https://medium.com/@perrystephenson/building-a-cost-effective-deep-learning-dream-machine-147cf670ca0d?source=rss-9c1d39a3c2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/147cf670ca0d</guid>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[pc]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Perry Stephenson]]></dc:creator>
            <pubDate>Wed, 20 Dec 2017 02:38:32 GMT</pubDate>
            <atom:updated>2017-12-20T09:46:39.049Z</atom:updated>
            <content:encoded><![CDATA[<p>This post follows my previous post: <a href="https://medium.com/@perrystephenson/building-a-deep-learning-rig-that-pays-for-itself-6cc8f99fcf1b">Building a Deep Learning rig that pays for itself</a> — you should definitely go read that one if you’re interested in finding out how I justified spending the money 💰💰💰.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*yFaO6BQ1McwRpgLo_eDZqQ.jpeg" /><figcaption><a href="https://www.flickr.com/photos/pagedooley/19687972695">Source</a></figcaption></figure><p>With a solid justification under my belt, I had to figure out what I was going to build. Given the approximate payback rate of $1,500 per year I really wanted to keep it below that figure. Practical constraints drove it closer to $2,000, but given that the asset will have a residual value of more than $500 at the end of the year I think I’m in the right ballpark.</p><h4>GPU</h4><p>This is arguably the whole point of the exercise, and as I mentioned in the previous post I chose an <strong><em>nVidia GTX 1080</em></strong>, as it’s significantly cheaper than the market leading GTX 1080 Ti yet it performs almost as well. I spent a fair bit of time trying to understand why so many different companies are making products for nVidia, and I’m still a bit confused, but it seems like they’re either building cards around an nVidia chip or they’re building chips under licence. Either way, everything I read online suggests that it doesn’t really matter which manufacturer you use, and that it comes down to taste. I disagree with the taste bit though — for me it comes down to price and noise. This thing is going to run 24/7 so it needs to be quiet. Luckily the cheapest card on the market ($699, made by GALAX) happens to be one of the more quiet designs. Sold.</p><h4>CPU</h4><p>This is less important for deep learning and mining profitability, but most ML work is done on the CPU so I still need a solid performer here. The AMD Ryzen chips are getting really good reviews, and seem to be much better value than Intel’s latest offerings. The<strong> AMD <em>Ryzen 5 1600</em></strong> in particular seems excellent value, giving you 6 cores (which means 12 parallel processing threads with hyperthreading, i.e. I can train models in 12x parallel) and performing almost as well as the top-of-the-range 1600X. It also comes with a cooler, which means it’s WAY cheaper and gives you about 90% of the performance of the 1600X. Same as the GPU, I’ve tried to pick the elbow of the price-value curve, getting the most powerful thing available before the prices start going up sharply.</p><h4>Motherboard, RAM and SSD</h4><p>This was way easier than I expected. There is a magical new site called <a href="https://au.pcpartpicker.com/">PC Part Picker</a> which lets you select components based on a search (which means you can sort based on price), and checks to make sure all of your parts are compatible. The site suggested the <strong>ABRock AB350M Pro4</strong> motherboard (because it was the cheapest compatible board), 2x8GB RAM chips which I can’t remember what brand they are (you’ll definitely want 16GB or more of RAM), and a <strong>Samsung SSD</strong> (which I chose to be 500GB). It’s worth pointing out that an SSD is almost essential for the ML side of things, because the speed at which you can read data into RAM is important for training when you’re using large datasets. This means that you don’t actually need an SSD for Windows, and it might be worth picking up another cheap, smaller drive for Windows if you don’t already have one lying around.</p><h4>Case and Power Supply</h4><p>Again, PC Part Picker to the rescue — it does all the hard work and checks part compatibility for you. I chose a <strong>Corsair Carbide 400C</strong> because it’s relatively cheap, looks pretty slick, and doesn’t have all of that gross “gamer” stuff all over it. The power supply just needs to be large enough to cover PC Part Picker’s power estimate (329W in my case), and you’ll probably want to pick one which has some sort of precious metal in the name as that corresponds to energy efficiency. I chose a Corsair 550W “Gold” power supply.</p><h4>Other things I forgot to buy the first time</h4><p>I already had a monitor, but I had forgotten that non-laptop-people have to buy keyboards and mice. I also forgot that desktops typically don’t come with WiFi by default, so I had to buy a card for that too.</p><p>If you would like some alternative opinions about how to select parts, I also liked the approach taken in <a href="https://blog.slavv.com/the-1700-great-deep-learning-box-assembly-setup-and-benchmarks-148c5ebe6415">this guy’s post</a>.</p><p>The full list of equipment in the build is <a href="https://au.pcpartpicker.com/list/KVFmyf">here</a>, and the total cost of the build (without a monitor) is $1,798. Not bad!</p><h4>The build process</h4><p>I won’t document the whole build process here as I’m anything but an expert on building PCs (I followed <a href="https://www.youtube.com/watch?v=IhX0fOUYd8Q">this video</a>), but I will include a few photos to prove that I actually did it. I was pretty happy with my component selection, but I found the following things hard or surprising:</p><ul><li>The GPU card is <strong>huge</strong>. I saw the measurements and checked that it would fit, but I still wasn’t really expecting it to be so massive.</li><li>The case is great and has some really excellent features, but it was pretty tight getting the back cover on after hiding all the cables back there. Closing the back cover was the hardest part of the whole build.</li><li>I forgot that I had to put an operating system on it, so I had to use Windows Boot Camp on my Macbook to create a “Bootable Windows Installer USB” so that I could install Windows on the new machine. Not difficult, it just delayed launch by about an hour.</li><li>Everything comes with a CD. I’m not sure why — all the software downloads automatically.</li></ul><p>Photo time! I think it looks pretty good.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KGFg0YaEcjIcNpAnehxtxQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bB-uueuZbk1qppOEagtnfQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ugf_db079_IDWeCukuv-Ww.jpeg" /><figcaption>All the expensive bits, the RAM installed in front of the CPU fan, and the motherboard installed in the case.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L61V0s33ixlV3d6uhgquRg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ppMnsvNlHsOe4ca7fYi1wA.jpeg" /><figcaption>The GPU card next to a (very large) screndriver for scale, and the completed box sitting under my desk.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=147cf670ca0d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a Deep Learning rig that pays for itself]]></title>
            <link>https://medium.com/@perrystephenson/building-a-deep-learning-rig-that-pays-for-itself-6cc8f99fcf1b?source=rss-9c1d39a3c2a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/6cc8f99fcf1b</guid>
            <category><![CDATA[pc]]></category>
            <category><![CDATA[bitcoin]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Perry Stephenson]]></dc:creator>
            <pubDate>Tue, 19 Dec 2017 22:14:24 GMT</pubDate>
            <atom:updated>2017-12-20T09:50:06.264Z</atom:updated>
            <content:encoded><![CDATA[<p>I’ve been looking at building a deep learning rig for a little while, but kept putting it off by telling myself that AWS was cheaper. Unfortunately, AWS is also an outrageous pain-in-the-arse for ad-hoc projects, so I found myself simply avoiding deep learning because of the long training times on my Macbook. More recently I’ve been playing around with cryptocurrencies, and got drawn in by the idea that you can earn back the cost of hardware by mining currencies on GPUs. Can I build a deep learning rig which pays for itself by mining cryptocurrencies when I’m not using it? I smell a project!</p><h4>Can a PC pay itself back by mining cryptocurrencies?</h4><p>This is the first question I set out to investigate. Firstly, I had to check that the workflow is possible, so I used Boot Camp to install Windows on my Macbook Pro, and then started using the AMD GPU in my Macbook to mine for coins using <a href="http://www.awesomeminer.com/">Awesome Miner</a>. <em>Unfortunately, almost all of the mining software and tools are for Windows, and the most heavily optimised (profitable) are almost exclusively for Windows users. Sigh.</em></p><p>This actually worked, and Awesome Miner estimated I was making about $1 (USD) per day. This roughly lines up with what I calculated for the Macbook using <a href="http://whattomine.com/">whattomine.com</a>, which means that I can probably trust this site to be reasonably accurate when calculating potential profits for the new machine. I ran this setup for a couple of days to make sure I could test out the whole process, and after a few days and nights of mining on my fairly weak Macbook GPU, I had enough currency to cash out and transfer real profits to my own wallet. Given the small volume I took a 40% haircut due to the transaction fee, but given this is about testing the pipeline I’m not overly concerned about transaction fees at this point.</p><p>By the way, I’m using Litecoin instead of Bitcoin as the main currency for my earnings. Awesome Miner automatically switches algorithm (and coin) based on what is most profitable for my hardware at any given point in time, then I can use <a href="https://miningpoolhub.com/">Mining Pool Hub</a>’s auto-exchange feature to automatically convert all of these currencies into a single currency for me to accumulate. Litecoin has faster transaction times and lower fees, so given the size and frequency of these transactions I lose a lot less currency and have much lower payout thresholds than if I was using Bitcoin. I can always swap for other coins later using <a href="https://shapeshift.io/">Shapeshift</a>.</p><p>With the end-to-end process verified (the money turned up in my wallet 👌) I set out to calculate the potential revenue from a new system. I made some assumptions here:</p><ul><li>GPU mining only (CPU mining is barely profitable)</li><li>Electricity cost is AUD$0.23/kWh (Sydney)</li><li>System power usage when mining is approx 333W (based on <a href="https://au.pcpartpicker.com/">PC Part Picker</a>, and being 1/3 of a kW makes the maths clean)</li><li>The GPU is an nVidia GTX 1080, as this is pretty much the second most powerful card on the market, and almost definitely the most cost effective for mining</li><li>Mining revenue is USD$5/day or AUD$6.5/day (based on the top three currencies returning higher than this on <a href="http://whattomine.com/">whattomine.com</a> when using the GTX 1080</li><li>Currencies and mining revenues remain constant (not a valid assumption, but recently they’ve been going <strong>up</strong> so this improves profitability)</li></ul><p>Based on all these assumptions, it looks like total profits per day in AUD are roughly $6.5 (revenue) minus $1.84 (8kWh of electricity), leaving roughly $4.50/day profit. I’m being a bit rubbery with the numbers, because mining revenues move all over the place, and staying on the conservative side. Using $4.50/day profit, you can pay off about $1,500 worth of deep learning rig every 12 months, allowing for a bit of down time so you can actually use it to train models!</p><p>With that all said and done, I’ve at least convinced myself that I’m allowed to build myself a powerful deep learning rig 😊. Stay tuned for the next part in what is likely to become a bit of a saga — building a cost-effective deep learning machine!</p><p>Part 2: <a href="https://medium.com/@perrystephenson/building-a-cost-effective-deep-learning-dream-machine-147cf670ca0d">Building a cost-effective Deep Learning dream machine</a></p><p>Edit: Now that I’ve built the machine, I can verify my calculations! Right now it’s making US$6.80/day on the GPU and $0.70/day on the CPU, so I’m definitely making profit! Happy days.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8RON1AeWMB7USJ46hy_l4A.jpeg" /><figcaption>Proof that I’m earning good money! Also some really strange artefacts from the camera and the monitor — not sure what’s going on there.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6cc8f99fcf1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>