Overview
- Phoronix’s published tests, which appeared Tuesday, showed Vera delivered roughly a 1.6x geometric-mean gain over NVIDIA’s prior Grace CPU and beat the specific AMD EPYC and Intel Xeon chips included in the suite of workloads.
- Vera is an ARM-based, monolithic design with 88 custom Olympus cores, a second-generation coherency fabric and an LPDDR5X memory subsystem that Phoronix measured sustaining up to about 90% of peak STREAM TRIAD bandwidth and up to 1.2 TB/s peak.
- The benchmarks were run on a pre-production Vera module under Nvidia-approved test selections and with restrictions that prevented testers from recording some power and frequency measurements, so energy and performance-per-watt claims remain unverified.
- NVIDIA says it has begun delivering pre-production Vera racks to major AI firms and cloud providers and expects partner-available systems in the second half of 2026, making those deployments the next real test of production performance and reliability.
- Rivals including AMD, Intel and other ARM-based server vendors are readying next-generation chips, so independent, unrestricted testing and wide-scale cloud adoption will determine whether Vera reshapes CPU choices for agentic AI workloads.