<body><div><p>In this GPU age, CPUs are often considered second citizens because most of the performance comes from the GPU. In most systems, GPUs are separate PCIe devices used by the CPU to accelerate numeric workloads. The exception, of course, is integrated (CPU and GPU) processors like the Nvidia Grace Hopper GH200 or the AMD Instinct MI300A.</p><p>Fundamentally, CPU performance and features are still an important part of the HPC equation. Indeed, there are still many HPC applications that run exclusively on CPUs.</p><p>In recent weeks, HPCwire published some benchmarks for Intel’s new Granite Rapids CPU (Xeon 6 6980P). The numbers and power envelope were impressive based on previous Intel offerings.</p><p>The benchmarks included AMD EPYC Genoa processors that were released in November 2022. This old vs. new comparison issue is common with many benchmarks because the release cadence is out of sync between vendor product lines. Of course, the benchmarks are valid and useful for what is currently available in the market “right now,” but if a new CPU is imminent, then users may wait for the new measurements.</p><p>In the case of Intel and AMD, the x86 planets have aligned. About a month after Intel released its new Xeon 6 Granite Ridge processors AMD released its 5th Gen Epyc Turin (Zen-5 architecture) processors. The new Intel-AMD agreement notwithstanding, there is now a possibility for a latest and greatest x86 head-head comparison.</p><p>Of course, such a comparison is available thanks to the premier benchmark site Phoronix!</p><h2>Standard Benchmark Disclaimer</h2><p>Benchmarking systems is important in HPC. The community focuses on the Top500 list twice a year to see which systems run the HPLinpack benchmark the fastest. The list is important from both historical and current perspectives, yet many users can find no correlation between the Top500 results and how well their specific application set will run on a modern system.</p><p>Obviously, the more benchmarks results for a specific platform, the better. It is even possible that your application (or something similar) was part of a benchmarking suite. Outside of independent 3rd party tests, the best benchmarks are the ones performed on your applications with your data profiles.</p><p>In HPC, the top-line FLOPS/sec numbers are important for measuring processor performance. These numbers provide a great start. However, they are not the whole story. Other issues need to be considered when accessing new hardware. For instance;</p><ol><li>How much faster is the new hardware versus my current hardware?</li><li>What is the price-to-performance and performance-per-watt for a fully configured server?</li><li>How well does it work in a clustered environment?</li><li>Will my current storage system be a bottleneck with new hardware?</li><li>What is the overall power usage when idle and fully loaded?</li><li>What level of tools and libraries are available for the platform?</li></ol><p>It is important to keep these issues in mind when evaluating benchmarks.</p><h2>New x86 Contenders</h2><p>The following results are some snapshots of popular HPC benchmarks. Consult the full article for a complete account of all the results and test setups. The results include a large list of processors both in 1P and 2P configurations (the new processors are in <b>bold</b>):</p><ul><li>Xeon Platinum 8490H</li><li>Xeon Max 9480</li><li>Xeon Platinum 8592+</li><li>Xeon 6766E</li><li>Xeon 6780E</li><li><b>Xeon 6980P @ DDR5-6400</b></li><li><b>Xeon 6980P @ MRDIMM-8800</b></li></ul><p>The new Xeon 6980P processor is paired with either MRDIMM-8800 or DDR5-6400 memory in an Intel AvenueCity reference platform. The 6980P has 128 P-cores, (Performance-cores) a 2.0 GHz base clock, a 3.2 boost cock, 504 MB L3 cache, and a 500 W TDP.</p><ul><li>Epyc 9654</li><li>Epyc 9754</li><li>Epyc 9684X</li><li><b>Epyc 9575F</b></li><li><b>Epyc 9755</b></li><li><b>Epyc 9965</b></li></ul><p>The new AMD Turin processor contenders are:</p><ol><li>Epyc 9755 – 128-core, base clock 2.7 GHz, boost to 4.1GHz, with a 500 Watt TDP, </li><li>Epyc 9575F – 64-core, base 3.3 GHz, boost 5.0 GHz, with a 400 Watt TDP</li><li>Epyc 9965 – 192-core, base clock 2.25 GHz, boost to 3.7 GHz, with a 500 Watt TDP </li></ol><h2>Select Benchmark Results</h2><p>Phoronix provides benchmarks for a good cross-section of applications. The following are highlights from some of the HPC benchmarks; consult the full article for more details and other applications. The highlights also include most of the benchmarks presented in the previous Granite Rapids article.</p><h2>GROMACS</h2><p>GROMACS is a molecular dynamics package mainly designed to simulate proteins, lipids, and nucleic acids. It was originally developed in the Biophysical Chemistry department of the University of Groningen and is now maintained by contributors in universities and research centers worldwide.</p><figure><img src="https://img.particlenews.com/image.php?type=thumbnail_580x000&amp;url=1pi0TQ_0wKtw3fC00" alt="https://img.particlenews.com/image.php?url=1pi0TQ_0wKtw3fC00" width="835" height="680" ><figcaption>GROMACS 1P results, some results truncated (Source Phoronix)</figcaption></figure><p>The results for GROMACS show excellent performance for both the AMD 9965 (192-core) and the 9755 (128-core) processors.</p><p>Intel catches up in the 2P results, with the Xeon 6980 (DDR-6400) coming in second to the Epyc 9755. Interestingly, the Xeon 6980 with the faster MRDIMM sightly lags behind the DDR result. GROMACS performance is more sensitive to processor (core) performance than memory bandwidth.</p><figure><img src="https://img.particlenews.com/image.php?type=thumbnail_580x000&amp;url=1zJqLn_0wKtw3fC00" alt="https://img.particlenews.com/image.php?url=1zJqLn_0wKtw3fC00" width="826" height="543" ><figcaption>GROMACS 2P results, some results truncated (Source Phoronix)</figcaption></figure><h2>OpenFOAM</h2><p>OpenFOAM is a C++ toolbox for developing customized numerical solvers and pre-/post-processing utilities for solving continuum mechanics problems, most prominently including computational fluid dynamics.</p><p>In the small OpenFOAM benchmark, the 2P Epyc 9755 does the best, with the 2P 9965 (with more cores) coming in second. The 2P Xeon 6980 with MRDIMM pulls into the third spot. These results show the move away from compute to memory-sensitive performance (the extra core in the 9965 actually slowed the benchmark)</p><figure><img src="https://img.particlenews.com/image.php?type=thumbnail_580x000&amp;url=0l4bXc_0wKtw3fC00" alt="https://img.particlenews.com/image.php?url=0l4bXc_0wKtw3fC00" width="592" height="502" ><figcaption>OpenFOAM small model results, some results truncated (Source Phoronix)</figcaption></figure><p>In the medium benchmark, the effect of memory bandwidth appears to take over performance. The results are close for all processors, but the 2P Xeon 6980 with DDR5 memory provides the best performance. Again, the extra cores in the 9965 reduce performance)</p><figure><img src="https://img.particlenews.com/image.php?type=thumbnail_580x000&amp;url=1wnH5r_0wKtw3fC00" alt="https://img.particlenews.com/image.php?url=1wnH5r_0wKtw3fC00" width="835" height="578" ><figcaption>OpenFOAM medium model 2P results, some results truncated (Source Phoronix)</figcaption></figure><h2>HPCG</h2><p>The High-Performance Conjugate Gradients (HPCG) Benchmark is intended to model the data access patterns of real-world applications such as sparse matrix calculations, thus testing/stressing memory subsystems.</p><figure><img src="https://img.particlenews.com/image.php?type=thumbnail_580x000&amp;url=0xbTeZ_0wKtw3fC00" alt="https://img.particlenews.com/image.php?url=0xbTeZ_0wKtw3fC00" width="830" height="582" ><figcaption>HPCG Results, some results truncated (Source Phoronix)</figcaption></figure><p>As expected, the 2P Xeon 6980 with MRDIMM wins this benchmark, with the DDR5 systems coming in second. Interstingly, the Epyc 9575F(64-core) comes in a distant third in front of the higher-core Epyc processors (128 and 196 cores). This indicates that the cores could be in contention for memory bandwidth. </p><h2>MiniBUDE</h2><p>The miniBUDE benchmark is an implementation of the core computation of the Bristol University Docking Engine (BUDE) implemented using Advanced Vector Extensions 512 (AVX-512).</p><p>The new Zen-5 AMD Epyc “Turin” processors have a full 512-bit data path for faster AVX-512 processing. The previous generation, Zen-4, implemented AVX-512 in a 256-bit “double pumped” approach that worked well and proved to be very efficient.</p><p>Phoronix has recently tested the new Zen-5 against the Zen-4 AVX-512 implementation and found it the performance improved in Zen-5.</p><figure><img src="https://img.particlenews.com/image.php?type=thumbnail_580x000&amp;url=2zh5qK_0wKtw3fC00" alt="https://img.particlenews.com/image.php?url=2zh5qK_0wKtw3fC00" width="844" height="664" ><figcaption>MiniBUDE AVX-512 results, some results truncated (Source Phoronix)</figcaption></figure><p>The improved AVX-512 can be seen in the MiniBUDE result in the figure above. All of the top finishers, both 1P and 2P (with the exception of the 1P 64-core 9575F), are the new Zen-5 processors.</p><h2>Just the Beginning</h2><p>As the benchmarks show, each new processor has its strengths. The Xeon 6 seems to offer</p><p>more memory performance, while the new Epyc Zen-5 processors excel at crunching numbers. There will be more application performance data as these processors make there way into HPC systems.</p><figure><img src="https://img.particlenews.com/image.php?type=thumbnail_580x000&amp;url=2PYVG4_0wKtw3fC00" alt="https://img.particlenews.com/image.php?url=2PYVG4_0wKtw3fC00" width="828" height="427" ><figcaption>Geometric mean of all tests run by Phoronix (source Phoronix)</figcaption></figure><p>In the overall Geometric Mean of all tests, the new AMD Epyc processors do quite well. Using these kinds of summaries is dangerous when your application is sensitive to memory bandwidth or processor performance. Check assumptions, run benchmarks, and remember the devil is in the details.</p></div></body>

Mountain View, CA

NewsBreak

Xeon 6 vs. Zen-5 HPC Benchmark Showdown

In this GPU age, CPUs are often considered second citizens because most of the performance comes from the GPU. In most systems, GPUs are separate PCIe devices used by the CPU to accelerate numeric workloads. The exception, of course, is integrated (CPU and GPU) processors like the Nvidia Grace Hopper GH200 or the AMD Instinct MI300A.

Xeon 6 vs. Zen-5 HPC Benchmark Showdown

Related Search

Comments /

YOU MAY ALSO LIKE

Comments / 0