NVIDIA Vera CPU Gains 7.4% GEMM Throughput via Linux Kernel Scheduler Patches

NVIDIA engineers have submitted new Linux kernel patches to fix performance bottlenecks in their Vera CPU, a move that directly improves multi-core efficiency for users running dual-socket systems. These scheduler updates address a significant latency penalty that previously slowed down single-threaded tasks when switching between Simultaneous Multithreading (SMT) cores. The optimization matters because it unlocks higher compute throughput for workloads that rely on consistent physical core allocation.

Kernel updates reduce SMT latency and boost dual-socket compute performance

The changes target the Vera CPU architecture, which NVIDIA designs for high-performance computing and data center applications. Engineer Andrea Righi led the submission of patches that refine how the Linux kernel handles thread scheduling on these chips. The updates build upon previous optimizations, including Linux 7.3 NOHZ improvements, to create a more stable execution environment for demanding applications.

The most notable technical fix addresses a latency penalty of approximately 10,000 clock cycles that occurred when cores transitioned back to single-threaded mode. The new scheduling strategy prioritizes physical core PE0, ensuring that tasks remain on the same processing unit to minimize context-switching overhead. This approach reduces the idle time that previously degraded performance during mixed workloads.

These software optimizations have yielded measurable gains in benchmark tests. GEMM throughput on dual-socket Vera systems increased from 9.4 TFLOP/s to 10.1 TFLOP/s, representing a 7.4% improvement in the latest update. Cumulative gains since the initial release stand at 52%, with throughput rising from 6.2 TFLOP/s to 9.2 TFLOP/s after earlier scheduler and NOHZ adjustments. NVIDIA continues to refine the Vera CPU through iterative kernel patches rather than hardware revisions.

We looked at Unraid OS 7.3.0 Adds Internal Boot earlier while tracking Nvidia launches. These kernel updates demonstrate NVIDIA's commitment to maximizing Vera CPU performance through software-level tuning. The latest patches provide a clearer path for users seeking stable, high-throughput computing on dual-socket Vera platforms.