Nvidia's Vera CPU Beats EPYC Turin and Xeon 6980P in Independent Tests

Nvidia's 88-core Vera CPU beats AMD's EPYC 9755 by 1.5% on estimated SPEC CPU 2026 integer rate scores while running only 352 threads against the AMD system's 512. That is the headline result of independent benchmarking conducted by Tom's Hardware on a dual-socket Vera reference platform, and it places Nvidia's first custom-core server CPU ahead of contemporary x86 flagships on aggregate throughput.

Vera pairs 88 Olympus cores, built on the Armv9.2-A architecture with two threads per core, with up to 1.5 TB of replaceable SOCAMM2 LPDDR5X-9600 memory. Peak aggregate memory bandwidth reaches 1.2 TB/s, or roughly 14 GB/s per core — more than double AMD's dense Turin designs, according to the test data. The chip carries a 10-wide decode, a second-generation Scalable Coherency Fabric mesh that distributes 164 MB of L3 cache and moves up to 3.4 TB/s across the die, and 88 PCIe 6.4 lanes with CXL 3.1 support. Configurable TDP spans 250W to 450W.

What did the benchmarks show?

Testers compared a 2P Nvidia MGX Vera platform with 1.5 TB of Samsung LPDDR5-9600 SOCAMM2 against a Supermicro SYS-222HA carrying Intel Xeon 6980P processors with Micron DDR5-8800 MRDIMMs, and an AMD Volano reference platform with EPYC 9965 and EPYC 9755 chips on SK hynix DDR5-6000 RDIMMs. All SPEC runs used GCC 15.2 with common flags and the jemalloc allocator.

Nvidia's own white paper claimed a 3% lead over the EPYC 9755; the independent estimate came in at 1.5%. The Xeon 6980P trailed badly: even a retest with Intel's oneAPI optimized binaries, which lifted performance 10%, still left Vera 8.6% ahead with generic GCC.

Vera also topped the SPEC floating-point rate charts, driven by strong showings in memory-bound tests such as ROMS, Marian and LBM, with weaker results in FemFlow and CACTUS. It led every comparison in the STREAM memory bandwidth benchmark, with 1.8 TB/s of bandwidth available between the two Vera sockets. Applying the OpenMP "spread" policy gave Vera a further large bandwidth gain, while the same setting slightly regressed on AMD and Intel platforms.

Where does Vera fall behind?

The results are uneven. Vera posted chart-topping Zstd compression at both level 3 and level 19, but lost decompression to the Turin chips. It placed last in LLVM compilation tests with both Makefiles and Ninja, though it beat a dual 9755 system by 12% in timed GCC compilation. Cryptography is a clear weak point: Vera trailed significantly in John The Ripper and fell behind the large core arrays of its rivals in OpenSSL. AVX-512-dependent workloads such as simdjson JSON parsing and the OSPRay "dim_512" scientific visualization test favored AMD, sometimes by three or four times.

Python performance stands out as Vera's strongest suit. In PyBench total average test times, Vera led the EPYC 9755 by 4%, and it topped nearly every test in the PyPerformance suite outside a few web server cases. Since agentic frameworks such as LangGraph, CrewAI and AutoGen are built on Python, the testers treat this as the closest available proxy for agentic AI performance.

Database results were also strong. Vera led the like-for-like 64-thread and 128-thread SQLite tests, and on a cold-cache first run of ClickHouse it held nearly a 40% advantage over a dual EPYC 9755 server.

Why does this matter for the server market?

No standardized end-to-end agentic benchmark exists yet, so all current results are proxies. Mark Lohmeyer, VP and GM of AI and computing infrastructure at Google Cloud, described the workload pattern: agentic compute arrives in bursts — "a burst of use at the start, maybe for an hour, and then it goes away" — creating long-running but often idle workloads with heavy hardware contention.

Vera enters a market where AMD and Intel hold entrenched customer bases and granular SKU lists. Nvidia's Ian Buck acknowledged the limit at Vera's March reveal: "The world is not going to be served by one SKU of CPU. And that's not our intention." The competition responds soon: AMD's Zen 6-based Venice rolls out through the end of the year with much higher memory bandwidth, and Intel's Diamond Rapids targets 2027, both promising to narrow the bandwidth gap Vera currently enjoys.