AMD Bergamo And Genoa-X Performance Review: EPYC Chips For Big Iron Workloads
AMD Bergamo And Genoa-X: General Compute And Web Server Benchmarks
The Phoronix Test Suite offers a wide variety of workloads to choose from and eases reproducibility. We’d encourage you to visit openbenchmarking.org for additional information, reference numbers, and to compare your own existing infrastructure against these workloads, if you’d like. All of our results have been run fresh for this review except as noted with asterisks.
Core Mark 1.0.1 Benchmark
We begin our testing with Coremark. Coremark is a very quick no-nonsense multi-threaded CPU test intended for quick comparisons.
The AMD Bergamo EPYC 9754 starts off with a bang here, putting all 512 threads to great use. Per-core scaling isn’t perfect relative to the EPYC 9654, due to its lower clock speeds, but that is to be expected. The Genoa-X EPYC 9684X trails the EPYC 9654 by a small margin, but is certainly within striking distance.
7-Zip 1.10.0 Compression/Decompression
Next, we looked at 7-Zip compression and decompression. The compression workload is influenced by memory and cache performance as well as out of order processing. Decompression is much more integer-driven, but also stresses the branch prediction pipeline.
The EPYC 9754 rockets out in front once again, at least in the compute-heavy decompression workload. As compression stresses the cache more, we find its performance taper off to fall in line with the 96C parts. The EPYC 9654 retains its edge over the 9684X for decompression, but the extra cache does bring the EPYC 9684X to the victory podium for compression workloads.
Linux Kernel Compilation 1.15.0
Software compiling is a common task and building the Linux kernel itself has long been used as a performance benchmark. We tested with defconfig and allmodconfig with results reported in seconds.
The EPYC 9684X and 9654 virtually tie here with the larger allmodconfig favoring the cache-strapped model ever so slightly. As the EPYC 9554 shows, more cores aren’t necessarily the end-all-be-all here so we find the EPYC 9754 slip to the middle of the pack.
DaCapo Benchmark 1.0.1
The DaCapo benchmark suite evaluates the performance of the CPU and system as a whole in the context of Java. In particular, we are using the Jython test which has become popular among Python developers who want to develop in a Java environment.
We missed retesting DaCapo with the EPYC 9374F and Xeon 8462Y+, so those results are recycled from our intel Sapphire Rapids performance review. It doesn’t really impact our conclusions for these new chips though, neither of which is a strong performer in this situation which favors lower core counts and higher frequencies.
Blender BMD 3.6.0 Model 3D Rendering
Blender is a staple 3D rendering benchmark. We queued up the tried-and-true BMW scene and gauged the time to render in seconds.
Once again, the EPYC 9754 impresses with its immense thread count, taking the top spot by a comfortable margin. Blender is not particularly cache-intensive which works to its benefit and also results in the EPYC 9684X narrowly trailing the EPYC 9654.
X264 2.7.0 Video Encode
Another common use of these servers is to serve as a render farm for video. We used the multithreaded x264 encoder with both 1080p and 4K test footage.
The 4K x264 encode has some perplexing scaling characteristics. The 96-core EPYC 9654 performs best with its Genoa-X twin close behind, but these are followed by the 32-core EPYC 9374F while the 128-core EPYC 9754 brings up the rear for AMD while Intel's latest Xeons trail once again.
The 1080p encode is entirely inverted with regard to core counts, though results fall into two distinct bands with the EPYC 9754 joining the lower-performing Xeon chips.
PostgreSQL PGBench 1.13.0 Database Transaction Benchmark
PostgreSQL is very popular and provides us with a look at database processing performance. PGBench provides a rating of database transactions per second and corresponding average latency using both read-only and read-write workloads.
The Bergamo and Genoa-X chips post the lowest Read Only scores from AMD, but are still well ahead of Intel’s best showing.
Moving to Read Write, both are in the mix with the top-performing EPYC 9554 while the EPYC 9374F falls off to just narrowly remain ahead of most Intel Xeon Scalable Processors.
Apache HTTP Server 3.0.0 Benchmark
Apache is an open-source web server initially released in 1995 but holds onto a massive install base to this day. It supports a wide variety of modules to add various features.
The EPYC 9754 and EPYC 9684X both put up competitive results for Apache, serving in the neighborhood of 100,000 concurrent requests. This can be as much of a memory benchmark as it is compute, allowing the Genoa-X chip to float above the EPYC 9654.
NGINX 3.0.1 Web Server Benchmark
Open-source NGINX is commonly employed for web servers as well as for mail proxy, reverse proxy, and other uses like load balancing. It can be used in conjunction with other software like Apache or standalone and is heavily deployed in the cloud.
The Bergamo and Genoa-X chips rank at the bottom in our testing. Intel’s accelerators appear to buoy performance significantly here. Of course, NGINX is not usually run as monolithically as this testing is performed, but rather in Docker containers which may change the real-world landscape. Still, Intel's dominance is profound here.
SMHasher 1.1.0
SMHasher is a benchmark tool to assess the performance of various hashing algorithms. Such workloads pop up all over in the datacenter and are often critical for security applications. Wyhash is not considered completely secure, but it is very efficient, portable, and fast.
Intel’s accelerators offer those chips a meaningful advantage for hashing, leaving all of AMD’s processors trailing behind. Within AMD’s stack, performance is in line with expectations, with the higher clocked 96-core parts inching past the 128-core Bergamo chip.
With a look at general performance out of the way, let's move on to AI and HPC workloads...
