<?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 Mitali Soni on Medium]]></title>
        <description><![CDATA[Stories by Mitali Soni on Medium]]></description>
        <link>https://medium.com/@mitali.soni04?source=rss-9d3d629f0197------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*0BGd_gZ167Ttt1IRDY5Pkw.jpeg</url>
            <title>Stories by Mitali Soni on Medium</title>
            <link>https://medium.com/@mitali.soni04?source=rss-9d3d629f0197------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 22:29:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mitali.soni04/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[Decoding DRAM Timings : Part II]]></title>
            <link>https://medium.com/@mitali.soni04/decoding-dram-timings-part-ii-9541e0f72aca?source=rss-9d3d629f0197------2</link>
            <guid isPermaLink="false">https://medium.com/p/9541e0f72aca</guid>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Mitali Soni]]></dc:creator>
            <pubDate>Sun, 05 Nov 2017 12:44:10 GMT</pubDate>
            <atom:updated>2017-11-05T12:44:10.348Z</atom:updated>
            <content:encoded><![CDATA[<p>In the previous <a href="https://medium.com/p/4f97a5b90e82/edit">article</a>, we discussed sources of latency constraints in DRAM and motivation for studying them along with a brief summary of key timing parameters. In the second part of this series, we’ll try to cover few commonly used techniques to minimize the impact of added latency on the performance of memory system.</p><p><strong>A. Optimal Row Buffer Locality</strong></p><p>Once a row is activated, sense amplifiers contain data of the entire row. Any subsequent accesses to different columns in the same row can be scheduled after maximum of data burst duration or tCCD delay, whereas column accesses to different rows of the same bank can be scheduled after tRC delay. Since tRC is significantly greater than tCCD or time taken by data burst, it is desirable to read as much data as possible from different columns of the activated row before activating another row. Thus, working out optimal number of column accesses in an activated row can help in scheduling memory accesses in a more efficient manner so that subsequent memory accesses from the same row do not incur latency or energy costs of another row activation.</p><p>Burst length (BL) determines the maximum number of columns that can be accessed for a given command. When a read or write command is issued, data worth of BL columns is effectively selected. We already know that tRRD is the minimum delay between row activation commands to different banks within a same DRAM device.</p><p>For a DDR (double data rate) memory, data transfer occurs at both rising and falling edges of the clock and it takes BL/2 clock cycles to transfer data for a single read or write command. Thus, dividing tRRD by BL/2 would give the optimal number of commands to be sent to the same row (also referred as optimal row buffer locality) before activating next row.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/349/1*8fMdB0dKBtFlnEZaWFKYhQ.png" /></figure><p>For LPDDR4 at 1600 MHz with BL=16, the optimal row buffer locality would be 2. Therefore, to eliminate inefficiencies due to traffic pattern, every activate should be followed by minimum 2 transactions to the same row.</p><p><strong>B. Exploiting Bank Level Parallelism</strong></p><p>In DRAM, multiple banks on the same device can be accessed independently. Bank level parallelism can be exploited by interleaving requests across multiple banks of memory.</p><p>Average number of transactions per bank can be computed as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/618/1*q9VrV2r5sHqgqlaOuGDMHg.png" /></figure><p>tRC is the minimum time between successive activate commands to the same bank and hence:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/557/1*8LsPe7Jppnktz7D34Lf0MA.png" /></figure><p>The minimum number of banks that should be accessed in parallel to be efficient can be derived as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/570/1*r0cPf0_g6WI8VoBjJbf1AA.png" /></figure><p>However, we need to keep in mind that interleaving requests across banks without maintaining optimal row buffer locality per activate can lead to poor performance due to increased number of activate and precharge commands.</p><p><strong>C. Per Bank Refresh</strong></p><p>Owing to the volatile nature of DRAM, to retain data, DRAM cells need to be periodically refreshed. By issuing a refresh command, data from DRAM cells is read out and restored to full voltage before the electrical charges start decaying. Refresh cycles interfere with memory performance as they consume available bandwidth and power. In all bank refresh, memory controller sends a single refresh command and one row across all banks within a rank are concurrently refreshed at one point, stalling the normal memory accesses.</p><p>Per bank refresh is a common technique employed to reduce impact of refreshes on memory bandwidth. Instead of all bank concurrent refresh, per bank refresh allows separate refresh commands to refresh one row of a single bank at a time. Researchers have further optimized the scheduling of per bank refreshes to further minimize the impact on bandwidth. There are several other mechanisms which try to minimize the power consumption in all bank refresh.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9541e0f72aca" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Decoding DRAM Timings : Part I]]></title>
            <link>https://medium.com/@mitali.soni04/decoding-dram-timings-part-i-4f97a5b90e82?source=rss-9d3d629f0197------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f97a5b90e82</guid>
            <category><![CDATA[dram]]></category>
            <category><![CDATA[memory-systems]]></category>
            <category><![CDATA[computer-architecture]]></category>
            <dc:creator><![CDATA[Mitali Soni]]></dc:creator>
            <pubDate>Sun, 05 Nov 2017 12:38:04 GMT</pubDate>
            <atom:updated>2017-11-05T12:38:04.552Z</atom:updated>
            <content:encoded><![CDATA[<p>Hardware experts have long been anticipating demise of DRAMs, yet with latest products like HBM2 and DDR5, this technology still remains a popular choice of leading semiconductor companies in the memory products space. Though there are quite a few emerging non-volatile memory technologies like PCM (phase-change memory), ReRAM (resistive RAM), FRAM (ferroelectric RAM)or MRAM (magnetoresistive RAM), these memory types are still in research or early stages of deployment, and are struggling with limitations on speed, endurance and manufacturing costs.</p><p>Though DRAM capacity and cost have continously been scaling, latency of DRAM has remained almost constant. Consequently, in order to maximize utilization from DRAM devices, one has to schedule memory accesses in such a way that minimum latency restrictions to guarantee reliable operation, are hidden well behind actual<strong> </strong>data transfers<strong> </strong>through multiple levels of parallelism<strong>. </strong>In-depth<strong> </strong>understanding of DRAM timing parameters is crucial to achieve this goal and this series of articles is an attempt in this direction. These articles assume that readers are familiar with DRAM system organization and basic memory access protocols.</p><p>DRAM chips use capacitor based cells to store data in form of electrical charge. Storing and retrieving data is essentially moving electrical charge to and from these cells, respectively. Such movement of charge is the source of DRAM latency for the following reasons:<br> i. For storing (writing) data, charge is injected into the cell through a wire called bitline. To reduce cost-per-bit of, many cells are connected to one bitline. Thus, owing to large resistance and the large capacitance of bitline, the cell experiences a large RC-delay, which increases the time it takes to become fully charged.</p><p>ii. To increase capacity, the DRAM cell size has been scaled down aggressively limiting the amount of charge that can be stored in its small capacitor which weakens its ability to drive charge into the bitline, while retrieving the data. As a result, cell cannot charge the bitline quickly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/460/1*yGI20v4y2zbcUTr7KHQhbQ.png" /></figure><p>Here’s a summary of few key timing parameters:</p><p><strong>tRCD</strong>: Delay in moving data from DRAM cells to sense amplifiers as a part of row activation command (referred to as activate command)</p><p><strong>tRAS</strong>: The minimum time from the initial issuance of the row access command, till the DRAM cells are ready for a precharge</p><p><em>For reads, tRAS ≥ tRCD+ tCAS +Max( tCCD, data burst duration)</em></p><p><em>In case of writes, tRAS ≥ tRCD + tCWD +tCCD + tWR</em></p><p><strong>tCAS</strong>: Time between column read command and valid data on the data bus. Modern DRAM move data in short and continuous bursts.</p><p><strong>tCCD</strong>: Number of cycles that must be allowed to elapse between successive column commands, determined by internal burst length of the DRAM device.</p><p><strong>tCWD</strong>: Delay between issuance of the column-write command and placement of data on the data bus by the memory controller</p><p><strong>tWR</strong>: The minimum time between the end of a write data burst and the start of a precharge command</p><p><strong>tWTR</strong>: The minimum time interval between the end of a write data burst and the start of a column-read command. Time taken by the write command to release I/O gating resources.</p><p><strong>tRP</strong>: Delay to reset the bitlines and the sense amplifiers</p><p><strong>tRC</strong>: The time interval between accesses to different rows within a same bank</p><p><em>tRC = tRAS + tRP</em></p><p><strong>tRRD</strong>: Minimum delay between row activation commands to different banks within a same DRAM device</p><p><strong>tRFC:</strong> Time between a refresh command and a successive refresh or activate</p><p><strong>tFAW : </strong>Rolling time window in which maximum of four banks can be activated</p><p><strong>tRTRS:</strong> Rank to rank switching penalty due to system level data strobe resynchronization</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/585/1*6xEcqA8ot_r9IIqV3XDMKQ.png" /></figure><p>In the next article, we would cover commonly used mechanisms to improve DRAM utilization and efficiency within the existing latency constraints.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f97a5b90e82" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>