TLDR Gripe / Practical
New FP4 MoE Kernels Achieve Identical Results on Blackwell GPUs
RightNow AI researchers have created bit-for-bit identical FP4 and MXFP8 Mixture of Experts kernels for NVIDIA Blackwell GPUs, ensuring consistent output logits regardless of batch size. These optimized kernels, which manage K-chunk splits and data exchanges across 8 GPUs, deliver significant performance gains—up to 5.1x for NVFP4 and 7.5x for MXFP8—while maintaining batch independence and precise arithmetic.
Read the full story