# llama.cpp adds CUDA tweak for Volta GPUs, boosting V100 speed

> A two‑line change routes Volta (sm_70) to Turing’s MMVQ table, giving about a 3% speed gain on a Tesla V100 while keeping results identical.

Oossa · 2026-10-02 · https://oossa.com/en/llama-cpp-adds-cuda-tweak-for-volta-gpus-boosting-v100-speed

The llama.cpp project released a small CUDA update on 2026-10-02. It makes Volta GPUs (sm_70) use the same MMVQ tuning table as Turing GPUs. On a Tesla V100 32 GB the change added roughly 3 % throughput for the Qwen3.8‑27B model. The output quality, measured by perplexity, stayed the same. The patch was contributed by anyei and co‑authors.

## The facts

- Measured +3.17 % speed increase on a Tesla V100 32 GB (CUDA 12.0.140).
- Release posted on Fri Oct 02 2026.

## Why it matters

Users running llama.cpp on V100 GPUs can run models slightly faster without losing accuracy.

## Sources & references

1. [ggml-org/llama.cpp b11335](https://github.com/ggml-org/llama.cpp/releases/tag/b11335) – llama.cpp, 2026-10-02

Last updated: 2026-10-02