Nvidia Vera CPU Gains 39% With Software Optimization at SC26
Nvidia's new Vera CPU, using proprietary Olympus cores, delivers up to 39% higher performance with optimized software, achieving 67% higher per-core performance than Grace in matrix benchmarks.

Nvidia's Vera CPU prioritizes single-thread performance and high per-core efficiency over maximizing core count. The processor uses the company's proprietary Olympus cores, featuring a 10-wide decoder and six 128-bit SIMD pipelines. Designed as both a standalone data-center CPU and the host for Vera Rubin NVL72 servers, Vera marks a shift from the Arm Neoverse V2 cores used in Nvidia's Grace CPU. The architecture employs 88 high-performance cores with spatial multithreading for 176 total threads, focusing on instructions per clock.
Hannah Coutand, who leads Vera CPU marketing at Nvidia, explained that CPU design is a game of trade-offs. Maximizing for core count leaves less space to maximize for IPC, she said. Vera's Olympus cores were built for high per-core performance and aggregate throughput. They include a neural branch predictor, a deep out-of-order execution engine, six vector execution units, four load pipelines, two store pipelines, large private caches, and advanced prefetch mechanisms.
Performance metrics and IPC gains
Nvidia presented performance figures at the Supercomputing 2026 conference. Tests used double-precision general matrix multiplication. With an OpenBLAS kernel optimized for Grace's Neoverse V2 cores, Vera achieved 5.43 instructions per clock and 15.69 FP64 operations per cycle. This was a 20% improvement over Grace running the same kernel. When using a kernel specifically optimized for the Olympus architecture, Vera's performance jumped to 7.53 IPC and 21.75 FP64 operations per cycle. This represents a nearly 67% higher per-core, per-clock performance than Grace in this workload.
The table below compares the core architectures and benchmark results for Grace and Vera CPUs.
| Processor | Core Architecture | SIMD Pipelines | FP64 Ops/Cycle (Peak) | IPC (OpenBLAS) | FP64 Ops/Cycle (OpenBLAS) | IPC (Optimized) | FP64 Ops/Cycle (Optimized) |
|---|---|---|---|---|---|---|---|
| Grace | Arm Neoverse V2 | Four 128-bit | 16 | 4.51 | 13.03 | Not Applicable | Not Applicable |
| Vera | Nvidia Olympus | Six 128-bit | 24 | 5.43 | 15.69 | 7.53 | 21.75 |
AI workload suitability
Vera is not intended for general-purpose cloud workloads like web servers. Nvidia primarily positions the CPU for agentic AI and other AI-related compute tasks. Coutand stated that Vera's high per-core performance, high memory bandwidth, and low memory-access latency make it ideal for agentic AI workloads. She noted that other processor choices may be better suited for general-purpose cloud data center servers. The Olympus microarchitecture was originally conceived for data-processing applications before being targeted at AI.
Software optimization impact
Architecture-specific software optimizations are critical to unlocking Vera's potential. Nvidia claims these optimizations can improve performance by up to 39% in some compute-intensive workloads. In the SC26 tests, the kernel optimized for Grace's Neoverse V2 cores used 81.4% of Grace's peak FP64 throughput but only about 66.5% of Vera's peak. In contrast, a kernel tuned for the Olympus architecture raised utilization of Vera's peak FP64 throughput to 90.6%. Changes to matrix tile organization, register allocation, and instruction scheduling drove this high utilization.
The presentation at SC26 concluded by highlighting the performance gains demonstrated through Olympus-tuned kernels.





