NVIDIA Vera CPU Benchmarks Beat AMD and Intel

According to a report from Phoronix, the first benchmark results for NVIDIA’s upcoming Vera CPU have officially surfaced, revealing a major performance leap over the previous-generation Grace processor and positioning NVIDIA as a stronger competitor against AMD and Intel in the high-end data center CPU market.

Based on the published benchmark data, the Vera CPU, featuring 88 custom-designed Olympus cores, delivered approximately 63% higher performance in overall geometric mean (Geomean) testing compared with the 72-core Grace CPU. The results also showed Vera outperforming AMD’s Zen 5-based EPYC 9575F processor by around 10%, while leading Intel’s Granite Rapids Xeon 6980P by nearly 55%.

Vera is a key component of NVIDIA’s next-generation Rubin AI platform. Unlike Grace, which utilized Arm Neoverse-V2 cores, Vera adopts NVIDIA’s in-house Olympus Armv9+ architecture. The processor is specifically optimized for Agentic AI, AI inference, and GPU-accelerated computing workloads.

According to NVIDIA, Vera can deliver up to 50% higher performance than traditional x86 CPUs, double the performance-per-watt efficiency, and four times the rack density. The processor also supports FP8 precision computing, PCIe Gen 6, and CXL 3.1 connectivity, while offering memory bandwidth of up to 1.2TB/s.

From a design perspective, however, AMD’s upcoming Venice Zen 6 platform and NVIDIA Vera target different deployment strategies in the data center market.

AMD Venice Zen 6 is expected to utilize TSMC’s advanced 2nm process technology, while NVIDIA Vera is reportedly based on a 3nm-class process node. The newer manufacturing process may provide AMD with advantages in core density and overall computational efficiency.

The two platforms also differ significantly in memory architecture. NVIDIA Vera adopts LPDDR5X memory, focusing on lower power consumption, higher bandwidth efficiency, and improved rack density for AI inference and GPU collaboration workloads. In contrast, industry reports indicate that AMD Venice Zen 6 may support up to 16-channel DDR5 memory configurations, delivering larger memory capacity and higher bandwidth for hyperscale data centers and high-performance computing (HPC) applications.

Phoronix has not yet disclosed Vera’s power consumption or performance-per-watt benchmark data. Since the current testing platform is reportedly based on early pre-production hardware, NVIDIA has not enabled power or frequency monitoring features at this stage.