Nvidia Vera leads in Python and memory benchmarks

Nvidia’s Vera CPU delivers strong Python performance and memory bandwidth while competing closely with AMD and Intel in general-purpose server workloads. The 88-core processor uses Nvidia’s custom Olympus cores, marking a substantial advance over Grace’s off-the-shelf Arm design. Its strengths align with agentic AI tasks, including executing Python code and querying large databases, but standardized benchmarks for complete agentic workflows remain unavailable.

In PyBench, Vera finished 4% ahead of AMD’s EPYC 9755 and led most PyPerformance tests. Estimated SPEC CPU 2026 integer throughput results put it 1.5% ahead of the EPYC 9755 configuration despite having fewer threads. Vera also topped the floating-point throughput rankings. Its memory interface supports up to 1.2 TB/s, with STREAM testing confirming a substantial bandwidth advantage over the competing systems. Database results were strong, including a nearly 40% advantage over a dual EPYC 9755 server on a cold-cache ClickHouse run.

Performance was less consistent elsewhere. Vera lagged in cryptography, decompression and some heavily threaded workloads. Compilation results varied: it led in building GCC but finished last in LLVM compilation tests. Intel led CPU-only prompt processing, while AMD retained advantages in several vector-intensive tasks. Vera therefore emerges as a credible standalone server option rather than a universal replacement, with suitability depending on workload and software requirements.