NeuroBrix uses V100 tensor cores via Triton kernels, enabled only where benchmarks show gains

Many V100 GPUs in universities and cloud instances go underused because Triton compiles matrix products on Volta to ordinary CUDA-core instructions, leaving the tensor cores idle. NeuroBrix's nbxtensor library adds hand-written kernels built on the mma.sync.m8n8k4 instruction for key operations such as matrix products, convolutions, flash attention and fused mixture-of-experts GEMMs. These kernels compete with the portable Triton kernels in a certification process that checks correctness against a float64 reference and times each candidate on the target GPU class. The engine then serves whichever path the certificate records as faster for each operation and shape. In a November 2026 benchmark on Tesla V100 hardware, the Triton engine outperformed PyTorch in 30 of 36 model-and-card pairs, reaching up to 5.24x on Janus-Pro-7B, while being slower in the remaining six.