Nvidia Vera CPU Benchmarks: Olympus Cores Delivering Great Performance

Summary

NVIDIA's Vera data center CPU isn't ramping up until later this year but I recently had the opportunity to try out this new ARM-based CPU designed for agentic AI workloads. NVIDIA's Vera CPU with its in-house-designed Olympus CPU cores ends up packing a heavy-hitting punch with competitiveness to Intel/AMD x86_64 CPUs that I have never seen out of any other ARM or non-x86_64 processors. Continue on with these early benchmarks of the NVIDIA Vera CPU on Linux. Vera is NVIDIA's next-gen data center CPU designed for agentic AI and similar modern data center workloads. Vera most notably will be found with NVIDIA NVL72 Vera Rubin as the host CPU for powering these powerful racks for AI while it will also be found standalone for CPU racks. Unlike NVIDIA's Grace that uses Arm Neoverse-V2 cores, Vera makes use of NVIDIA's in-house "Olympus" core design. Vera features 88 Olympus cores that claim to deliver 2x the performance of its predecessor with leading energy efficiency too. Olympus is compatible with the Armv9.2 ISA and supports FP8 precision, 176 threads in total via spatial multi-threading, and is paired with LPDDR5X memory for delivering up to 1.2TB/s of memory bandwidth. Over Grace, Vera also has double the L2 cache at 2MB per core, a larger unified L3 cache at 164MB, and supports PCIe Gen 6 as well as CXL 3.1 connectivity. The Vera CPU as tested for this initial benchmarking had a peak 450 Watt socket TDP. With the LPDDR5X memory it's around 50 Watts or less of power consumption. NVIDIA Vera data center CPUs remain on track for shipping in the second half of the year, but ahead of the ramp, NVIDIA invited me to their Santa Clara headquarters to run some of the first public benchmarks of this new CPU with their Olympus cores. In this article are those very initial results. But before talking performance, it's first important to note the level of Linux support. With Vera not officially out there yet I really didn't know what to expect for the upstream Linux kernel sup...