NVIDIA AI Updates: May 27, 2026

1. NVIDIA Vera CPU posts 1.6x geomean over Grace and beats leading x86 chips on developer workloads in Phoronix tests

NVIDIA. NVIDIA highlighted independent Phoronix benchmark results for its Arm-based Vera CPU, the chip designed for agentic AI workloads in next-generation Rubin systems. Vera ships 88 custom Olympus cores and 1.2 TB/s of memory bandwidth, and recorded a 1.6x geometric-mean uplift over the prior Grace generation across the Phoronix Test Suite. The tests showed Vera outperforming leading x86 server CPUs on developer-heavy workloads, including Linux kernel compilation, marking the first time NVIDIA’s general-purpose CPU has cleanly led that category in third-party testing. Source

2. NVIDIA CompileIQ auto-tunes compiler flags to squeeze more performance out of GPU kernels

NVIDIA. NVIDIA introduced CompileIQ, an auto-tuning tool that searches large compiler-option spaces to extract additional GPU kernel performance beyond what manual flag tuning typically achieves. The accompanying developer blog frames the technique as targeting one of the harder problems in performance engineering, especially for teams hand-optimizing LLM inference pipelines where small kernel-level gains compound at serving scale. CompileIQ automates the discovery of the compiler configurations that best match a specific workload, replacing trial-and-error sweeps with a guided search. Source