Nvidia Vera CPU Review: 2TB/S Memory Bandwidth, a Real x86 Rival

NVIDIA Vera is a move from supplying CPUs mainly as companions to its GPUs toward competing in general-purpose server computing. Its 88-core, 176-thread Olympus Arm processor replaces Grace’s standard Neoverse cores with a custom design, and pairs them with up to 1.5TB of replaceable LPDDR5X memory per socket. The system’s standout feature is bandwidth: each socket is rated for 1.2TB/s, and two sockets sustained more than 2TB/s in testing - 2.3 times the best tested DDR5 server. Vera also supports a high-bandwidth, coherent NVLink-C2C connection to GPUs. NVIDIA is positioning it for GPU systems, standalone servers, and CPU-heavy agentic AI workloads, while broad OEM support could make it a third major server-CPU option.

Results show a capable all-rounder rather than a universal winner. A single Vera core beat AMD’s 5.0GHz EPYC 9175F by 12% in 7-Zip compression, and Vera excelled in memory-bound OpenFOAM, finishing the largest test 1.44 times faster than the 384-core EPYC 9965. It also led that processor in PostgreSQL and MariaDB benchmarks, completed ClickBench faster, and served more Valkey requests. The main weakness is vector and cryptographic work: Olympus’s 128-bit SVE2 units fell behind x86 processors with 512-bit AVX-512, with the EPYC 9965 leading Vera by about threefold in some full-system tests. Vera’s appeal therefore rests on strong bandwidth, competitive general-purpose performance, and close integration with NVIDIA’s platform.