Early public benchmark results show NVIDIA's Vera CPU-built on 88 Olympus cores and a second‑generation LPDDR5X

Early public benchmark results show NVIDIA's Vera CPU-built on 88 Olympus cores and a second‑generation LPDDR5X subsystem — delivers sustained multi‑core throughput and class‑leading memory bandwidth per watt, with a reported 1.

NVIDIA's Vera CPU posted strong single‑socket throughput and memory efficiency in early public benchmark testing, meeting the sustained multi‑core and bandwidth demands for agentic AI workloads. The results matter because they highlight a platform optimized for continuous, branch‑heavy runtimes and multi‑tenant orchestration where predictable memory performance limits overall system throughput.

Vera combines 88 custom Olympus CPU cores on a monolithic die with wide cores, advanced branch prediction and a second‑generation Scalable Coherency Fabric to move data across cores. The platform pairs that CPU design with a second‑generation LPDDR5X memory subsystem rated up to 1.2 TB/s peak bandwidth. NVIDIA lists a single‑socket thermal design power of 450 watts, with under 30 watts attributed to memory.

Public tests concentrated on CPU‑heavy, agentic workloads rather than narrow synthetic metrics: tasks included code compilation, file compression, video transcoding, Python and Java runtimes, and database management. The testing reported both cross‑platform advantages and generational gains; the results include a 1.6x geometric‑mean improvement over NVIDIA's prior Grace CPU. Michael Larabel, founder and principal author of a prominent benchmarking site, called Vera “the most formidable competition to Intel and AMD x86_64 processors ever realized.

Memory efficiency stood out. In STREAM TRIAD runs Vera sustained about 90% of its rated peak memory bandwidth — the highest percentage of rated peak bandwidth recorded in these tests — and delivered more than four times the memory bandwidth per core versus the tested x86 processors. NVIDIA emphasizes that LPDDR5X achieves dramatically lower energy per bit than DDR5, enabling higher bandwidth with substantially lower memory power draw.

Beyond peak numbers, the platform showed predictable behavior under parallel load. Separate testing cited by NVIDIA, conducted by Prime Intellect, found Vera maintained high bandwidth and low, consistent memory latency as additional sandboxes and workloads ran concurrently — behavior that orchestration layers and multi‑tool agent stacks require. Olympus cores are fully compatible with the Armv9.2 ISA and were designed for branch‑heavy runtimes, sandboxed code, data processing and orchestration.

For builders, the practical takeaway is a combination of sustained multi‑core throughput, high memory bandwidth per watt and a compact power envelope aimed at the sequential CPU work and memory pressure that constrain agentic AI factories today. The reported 1.6x generation‑on‑generation improvement over Grace and leadership in the tested field suggest Vera offers a new option emphasizing memory efficiency and predictable multi‑tenant performance for AI‑oriented data centers.