Nvidia has provided access to its new Vera data center CPU for independent testing, marking a significant step in the company's expansion into server infrastructure. The benchmarks compare the standalone Vera processor against Intel's Granite Rapids-AP and AMD's EPYC Turin chips. While the market is currently dominated by x86 architectures from Intel and AMD, as well as custom Arm chips, Vera is specifically designed to address the rising demand for CPUs capable of handling agentic AI workflows. Although standardized end-to-end benchmarks for these specific workloads do not yet exist, initial data suggests Vera offers strong indicators of utility in this sector.
The Vera CPU is built on Nvidia's first custom core design, named Olympus, which utilizes the Armv9.2-A architecture. The chip features 88 cores with two threads per core via spatial multithreading, resulting in a total of 176 threads. It includes 2 MB of L2 cache per core and a massive 164 MB of L3 cache distributed across the die through Nvidia's second-generation Scalable Coherency Fabric (SCF). This mesh architecture provides up to 3.4 TB/s of bandwidth across the chip, a critical factor for performance in data-rich operations. The processor supports up to 1.5 TB of SOCAMM2 LPPDDR5X memory running at speeds up to 9600 MT/s, delivering an aggregate memory bandwidth of 1.2 TB/s, or approximately 14 GB/s per core.
In terms of raw performance, Vera demonstrates commanding capabilities in Python, a language that forms the backbone of much of the AI software ecosystem including frameworks like LangGraph and CrewAI. In PyBench tests, Vera led the competition by 4% over AMD's EPYC 9755, which itself outperformed Intel's Xeon offerings and the larger EPYC 9965. The advantage in Python is particularly notable given that these tasks are largely single-threaded. Furthermore, in the SPEC CPU 2026 intrate suite, Vera showed a 1.5% lead over the EPYC 9755 despite having significantly fewer threads (352 versus 512). When normalized per core, Vera's performance advantage becomes even more pronounced.
The testing also highlighted Vera's strengths in floating-point operations and memory-intensive tasks. The chip topped charts in several float benchmarks, driven by its high memory bandwidth and the efficiency of the SCF mesh. However, the processor showed some weaknesses in specific areas; cryptography tests using SEAL indicated weaker performance compared to rivals, and Zstd compression results showed Vera trailing behind AMD and Intel, although decompression performance remained competitive. In general-purpose server workloads, Vera often matched or beat comparable x86 chips, trading the core density of Intel and AMD for superior performance in backend, data-heavy operations.
The benchmarks were conducted on a dual-socket Nvidia MGX reference platform equipped with 1.5 TB of Samsung LPDDR5-9600 ECC memory. Competing systems included a Supermicro platform with Intel Xeon 6980P processors and an AMD Volcano reference platform featuring EPYC 9965 and 9755 chips. The tests covered a range of scenarios applicable to agentic workflows, such as code compilation, database queries, and JSON parsing, alongside general server tasks like Java, cryptography, and rendering. Nvidia did not impose restrictions on the testing methodology, allowing for an independent evaluation of the chip's capabilities across both specialized AI workloads and traditional server applications.
Sources: Tom's Hardware.