NVIDIA's opt-in MXFP8 kernel hits 2.5x in a Kimi K3 microbenchmark on one VR200 NVIDIA merged an opt-in TensorRT-LLM path on November 16 that runs eligible Kimi K3 shared experts with packed MXFP8 weights on SM107 hardware, with contributor-reported gains reaching 2.5x in an eager prefill microbenchmark on one VR200 GPU. The merged pull request benchmarks a single shared-expert layer, and NVIDIA says end-to-end latency and accuracy were not measured; the 2.5x figure is the best eager prefill result, at 8,192 tokens. NVIDIA's opt-in MXFP8 kernel hits 2.5x in a Kimi K3 microbenchmark on one VR200 The TensorRT-LLM change is off by default and targets replicated shared experts on SM107. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published NVIDIA merged an opt-in TensorRT-LLM path on November 16th that runs eligible Kimi K3 shared experts with packed MXFP8 weights on SM107 hardware, with contributor-reported gains reaching 2.5x in an eager prefill microbenchmark on one VR200 GPU. The merged pull request benchmarks a single shared-expert layer. NVIDIA says end-to-end latency and accuracy were not measured. The 2.5x figure is the best eager prefill result, at 8,192 tokens. What the kernel changes…