GoKawiilFollowing the launch of AMD's EPYC 'Venice' CPUs in July, AMD extended the performance claims for its upcoming generation of server chips on Friday. The high-level claim hasn't changed. AMD still says a 96-core, high-frequency Venice chip is around 20% faster than Nvidia's 88-core Vera in SPEC CPU 2026's Integer Rate test. However, the company went into far greater detail about the benchmarks in a new white paper.
There are several configuration differences depending on the benchmark throughout AMD's white paper, and although we'll call out those differences here to the best of our ability, we don't have all of the details. For the Vera comparison, in particular, AMD is mixing data from different sources, and in some cases, using different major releases of the GNU Compiler Collection (GCC). That can have a substantial impact on performance, so keep your salt shaker handy.
(Image credit: AMD)
First up are results in SPEC CPU 2026 with the intrate test, looking at total throughput. These are older numbers, gathered in July with GCC 15.2. The intrate test runs multiple copies of an application on the same CPU, and the SOP is to run one copy per thread. Presumably, that's what AMD did here, but the white paper doesn't clarify, even in the footnotes.
Latest Videos From Tom's Hardware Watch full video here:
The 256-core 9996 is 2.37x faster than the Intel Xeon 6980P and 2.24x faster than Vera according to the slide. The white paper clarifies the mystery 9006 CPU is the 256-core flagship. Perhaps most impressive is AMD's gen-on-gen comparison. According to these results, the 9996 is around 78% faster than last-gen's 192-core EPYC 9965.
Although the high-level results bring in data from Intel and AWS, much of the white paper focused squarely on the comparison between Venice and Vera. AMD broke down the individual subtests of SPEC CPU 2026 intrate in the white paper, which you can see below.
(Image credit: AMD)
The comparison looks good for AMD, naturally, though there are a few wrinkles in the configuration. AMD is testing a down-cored EPYC 9996, dropping from 256 cores to 96 cores. It made no mention of power budget, but when AMD originally shared SPEC numbers, the 96-core model had access to the same 600W as the 256-core model — AMD's 96-core, high-frequency Venice SKU tops out at 500W. More consequential is the compiler, however. AMD is using GCC 16.1 and comparing the results to the ones Nvidia shared in its Vera white paper. Nvidia used GCC 15.2.
Michael Larabel over at Phoronix has a nice write-up about the difference between GCC 15 and 16, but the short story is that there are performance differences, not always for the better. GCC 16 takes longer to compile due to better optimizations, hence the lower scores on the GCC and LLVM compilations above. However, that leads to faster binaries. By how much depends on the flags, software, and a whole host of other factors. Regardless, it's not best practice to compare benchmarks using two different compiler versions.
... continue reading