Yandex publishes attention kernels that train faster on Nvidia's Blackwell chips

Yandex Research released FLAMIA, a family of attention kernels for Blackwell GPUs that train in MXFP8 and mixed NVFP4/FP8 low-precision formats, with safeguards so shared scaling factors do not leak information from future tokens. It reports forward-plus-backward attention speedups of 1.10 to 1.40 times on B200 and up to 1.42 times on B300 over the fastest baseline it tested, and stable Llama-style pretraining. SourcesA

Sources

• AFLAMIA: FLAsh MIcroscaling Attention for fast and causally safe training on Blackwell (arXiv 2611.03344, via Hugging Face)huggingface.co

This story ran in the Morning Brief of November 3rd 2026.