AMD MI400 Detailed at Hot Chips: 40.26 PFLOPS MXFP4 and 432GB of HBM4 Per GPU

ServeTheHome·high signalXBlueskyLinkedInCopy link

AMD disclosed MI400 internals at Hot Chips 2026: 40.26 petaflops peak MXFP4 (4x MI355X), 20.13 petaflops each at MXFP6 and MXFP8, 315 teraflops vector FP16 and matrix FP32, and 432GB of HBM4 at 23.3 TB/s per GPU, up from 288GB of HBM3E. The part is eight accelerator complex dies on TSMC N2 with fabric, cache and I/O dies on N3P, 256 work group processors and 192MB of global L2. AMD claims 20 TB/s measured MLA decode bandwidth in FP8 for 3.8x MI355X throughput and about 2.4x energy efficiency, with the Helios rack hitting 1.7 PB/s across 72 GPUs.