GoKawiilThe very first set of Nvidia Vera CPU benchmarks have just been released by Phoronix, with results from a set of common Linux benchmarks. While the tests were curated by Nvidia at its Santa Clara headquarters, the early data from those tests indicates that Vera is highly competitive compared to AMD's EPYC and Intel Xeon offerings, at least in the workloads Nvidia is targeting with the chip.
Phoronix was invited to NVIDIA's Santa Clara headquarters to test the upcoming 88-core CPUs. Vera is notable for all kinds of reasons, but most especially because it doesn't license an Arm processor core. Instead, like Apple's chips, it uses the ARM instruction set on a fully custom CPU core known as "Olympus." This isn't the first time Nvidia has produced a custom CPU core; that would be "Denver" in the Tegra K1 from 12 years ago. However, where Denver was a desktop-class CPU constrained by a mobile power budget, Vera is a server-class monster fed with a server-class power budget, and the proof is in the benchmark results.
The 88-core NVIDIA Vera CPU, flanked by eight SOCAMM2 memory modules equipped with LPDDR5X memory on tiny gumstick-like PCBs. (Image credit: Nvidia/YouTube)
Phoronix's Michael Larabel was able to test Vera across a range of benchmarks, including code compilation tests, synthetic memory benchmarks, AV1 video encoding, Python, Java OpenJDK, file compression, Lua JIT, and some database benchmarks. In most of the tests, the Vera processors are highly competitive against the competition, which includes AMD EPYC "Turin" processors in single- and dual-CPU configurations as well as an Intel "Granite Rapids" Xeon chip. As Larabel says, Vera offers "competitiveness to Intel/AMD x86_64 CPUs that [he has] never seen out of any other ARM or non-x86_64 processors." It doesn't win outright in most tests, but it's running very close to the EPYC configurations, which is shockingly good for a first-generation custom server core from Nvidia.
Latest Videos From
What's even more impressive is if you switch things over to a per-thread view of the data, which Phoronix does for some tests. Historically, Arm server vendors have achieved similar performance versus AMD and Intel in multi-threaded workloads by simply cramming a huge pile of cores onto a chip, but the single-threaded performance of x86 was beyond reach. Well, in a timed Gem5 compilation, only the AMD EPYC 9575F was able to beat Nvidia's Vera on a "performance per core" metric. In a Linux kernel build, Vera tops the stack of server chips by a decent little margin. It's real eyebrow-raising stuff, and speaks to the strength of the Vera core.
Looking at the geometric mean of test results, Vera comes out on top by a fair margin, in part thanks to very strong performance in LuaJIT FFTs, ClickHouse database server, and the Renaissance JVM benchmark, where it absolutely dusted the competition. This is genuinely exceptional, but there are caveats to this data. Larabel notes that Nvidia limited the scope of these initial benchmarks to its "intended markets and target use-cases," but it's also true that these tests are benchmarks that Larabel is familiar with and already runs on server CPUs for performance analysis. He says that today's results are a "small subset" of his typical preferred approach, which is to say that this data is valid, but may not represent all workloads or use-cases.
Vera was created to support the Rubin GPU, as seen here, but Nvidia's also selling servers with just Vera. (Image credit: Tom's Hardware)
One key detail that the site wasn't able to test comes down to power efficiency. This is a big deal for consumer systems, but it's a much bigger deal for AI data center operators who are currently struggling with build-outs because the power infrastructure worldwide simply isn't robust enough to support megawatt AI training clusters popping up all over the place. Nvidia says that Vera has a 450W TDP, while the fast SOCAMM2 memory it uses consumes another 50 watts; the Xeon and EPYC chips used for comparisons are rated for 500W before accounting for platform memory power. Real power consumption can be very removed from TDP figures, though, so we'll have to wait and see how that plays out.
Another point in favor of Vera is software support. According to Larabel, Vera has "great upstream open-source support," which is encouraging for the future of these chips. Over the decades, we've seen many examples of promising hardware that falls by the wayside due to insufficient software support. All of the testing was performed on a mainline Linux kernel with no need for nasty Device Trees or hacky bespoke driver solutions.
... continue reading