Eighty-eight custom NVIDIA Olympus cores and 1.2 TB/s of memory bandwidth just delivered a 1.5x overall performance advantage over a latest-generation 128-core x86 processor in Phoronix testing. This shift toward agentic AI requires a new CPU profile for the AI factory: high core utilization, massive memory bandwidth, and the ability to sustain performance when all cores are active.
Custom Olympus Cores Drive Agentic Workloads
At the heart of the Vera platform are NVIDIA's custom Olympus CPU cores, which are fully compatible with the Armv9.2 instruction set architecture. These cores are specifically engineered for the sequential, branch-heavy runtimes that underpin agentic AI, including sandboxed code execution, data processing, and orchestration.
Phoronix's testing of a single-socket Vera CPU—rated at a 450-watt thermal design power—showed significant generational gains across a broad array of workloads. In practical developer tasks, the Vera platform compiled a default Linux kernel in just 20 seconds, the fastest result measured in the test. On a per-core basis, Vera delivered 2x faster Linux kernel compilation compared to a 128-core x86 processor. Michael Larabel, founder of Phoronix, noted that the performance represents the most formidable competition to Intel and AMD x86_64 processors realized to date.
LPDDR5X Memory Delivers 4x Bandwidth Per Core
Agentic workloads are rarely limited by core count alone; they demand sustained memory bandwidth and low, predictable latency. Vera addresses this by incorporating a second-generation LPDDR5X memory subsystem. This allows the platform to reach 1.2 TB/s of bandwidth while consuming less than 30 watts of memory power—a stark contrast to traditional DDR5 setups that often require more than 100 watts for similar throughput.
In Phoronix STREAM TRIAD testing, Vera sustained 90% of its peak memory bandwidth, achieving the highest percentage of rated peak bandwidth of any CPU tested by the publication. This resulted in over 4x the memory bandwidth per core compared to traditional x86 CPUs. Furthermore, testing by Prime Intellect confirmed that Vera maintains high bandwidth and low, consistent memory latency even as multiple workloads run in parallel, providing the predictable performance necessary for complex AI agent tool calls and data services.
NVIDIA expects Vera to be available from partners in both single- and dual-socket configurations, with air-cooled and liquid-cooled options, starting in the second half of the year.
Source: NVIDIA Vera CPU Is 'Packing a Heavy-Hitting Punch' Against Competition
Domain: blogs.nvidia.com
Comments load interactively on the live page.