The Epic Gains Made In 5 Years For AMD EPYC 7601 Naples vs. Newest 4th Gen EPYC Genoa

If taking the geometric mean of the nearly 200 benchmarks carried out, this is how the performance is at a high level... The EPYC 7601 2P was running at half the speed of a single EPYC 9374F that is just 32 cores / 64 threads. If running a 3~5 year old server, upgrading to Genoa can easily provide a compelling answer if wanting to consolidate your physical footprint and power. It's with much sadness the power reporting wasn't working for Zen 1 CPUs on Linux, but keep in mind the EPYC 7601 is a 180 Watt TDP or obviously 360 Watt for 2P. The EPYC 9374F in a 1P configuration has a 320 Watt TDP, or around 15% less than the Naples 2P flagship while delivering twice the performance overall.

If going from the flagship EPYC 7601 2P to EPYC 9654 2P for top-of-the-stack, that was a 4.26x improvement over the past five years since AMD introduced their EPYC processors.

If limiting the geometric mean just to the code compilation benchmarks, such as if running an aging CI/build server, moving to Genoa can mean much faster build times while also the possibility of consolidating your servers and reducing power use.

When looking at just the high performance computing (HPC) benchmarks is where the results begin to really show for how far AMD has come in five years with EPYC...

Or if looking at the machine learning / AI benchmarks is where it hits the point of insanity. Over a 13x improvement from the original EPYC Naples 2P flagship to the EPYC Genoa 2P flagship... Due to the heavy use of AVX-512 in these benchmarks is what leads to such an incredible difference for five years of AMD EPYC performance.

Or as another fun metric, the strides made when looking at all of the Intel oneAPI workloads tested... Embree, oneDNN. Open Image Denoise, OSPRay / OSPRay Studio, and OpenVINO. Intel has been putting out great open-source software for years. With Intel's open-source software they have invested a lot in AVX-512 usage to benefit their Xeon processors while now Zen 4 with its efficient AVX-512 implementation shows off massive improvements.

Outside of AVX-512, moving to twelve channels of DDR5 system memory helps in many of the workloads, the higher TDP CPUs, and the countless other architectural improvements from Zen 1 to Zen 4.

Well, that five year comparison ended up being much more fun and impactful than even originally anticipated... The AMD EPYC Genoa performance continues to demonstrate incredible performance uplift generationally or now as shown here if running an aging ~5 year old server, the raw performance gains can be huge or if simply looking to upgrade in order to downgrade your physical rackspace footprint or power consumption, there is a lot of opportunity with Genoa. Heck looking at my own numbers now makes me want to expedite upgrading Phoronix Media web servers currently using Rome CPUs to go for Genoa (when available at HiVelocity data centers). It's too bad though that the Zen 1 CPU power monitoring wasn't working accurately for even more data, but if just looking at the rated TDPs it's a big improvement there too. Now if finding time over the holidays, maybe I'll dust off some old AMD Opteron CPUs and see how those compare to Genoa?

On a value basis, EPYC processors continued to be priced competitively. The EPYC 7601 flagship launched at around $4200 USD in 2017 (or obviously, ~$8400 for 2P) while now in 2022 with the EPYC 9374F that in 1P form is delivering so much better performance than two Naples is priced at ~$4598. At the top end the EPYC 9654 96-core processor commands a list price of around $11,805.

It's also exciting to think about that this performance is just the at-launch AMD EPYC 9004 series on Linux... The GCC 12 stable compiler used with Ubuntu 22.10 for building the open-source benchmarks doesn't even have Zen 4 (znver4) tuning and in upstream GCC that work is still pending for delivering more compiler tuning. As shown separately in my AMD AOCC 4.0 compiler benchmarks there can be measurable performance gains out of AMD's tuned compiler or what will hopefully see similar levels of performance in future Clang and GCC versions. It will also be interesting to see what other software optimizations may come to the Linux kernel and elsewhere with developers now being able to leverage 192 core / 384 thread 2P servers, continued possibilities around AVX-512 in more software now that Zen 4 supports it across the board down to desktop CPUs, etc.

See my launch day AMD EPYC Genoa review for the Milan comparison points, how these CPUs stack up against current Intel Xeon Scalable processors, and also the performance benefits of running the CPUs in the power determinism mode. It will be very interesting in the new year to see how well AMD EPYC Genoa in turn competes with Intel's long-awaited Sapphire Rapids processors.