AMD introduces Instinct MI455X AI accelerator with 432 GB HBM4 and up to 40 PFLOPS performance

AMD introduced the Instinct MI455X AI accelerators for data centers, built on the CDNA 5 architecture, targeting large-scale AI training, inference, and rack-scale systems. Each accelerator packs 432 GB of HBM4 memory with a peak bandwidth of 23.3 TB/s. The GPU contains 320 billion transistors and uses a multi-chiplet design: eight accelerator complex dies (ACD) fabricated on TSMC N2 (2nm) process, two fabric chiplets, and two I/O chiplets on TSMC N3P (3nm) process. It features 256 Work Group Processor (WGP) compute units, 192 MB of L2 cache split into two 96 MB blocks, 12 HBM4 memory stacks via a 192-channel interface, and supports PCIe Gen 6 and UALink networking. Performance reaches 20.13 PFLOPS in MXFP8/MXFP6 and 40.26 PFLOPS in MXFP4. Compared to the Instinct MI355X, the MI455X offers 1.5x more memory, 2.9x higher memory bandwidth, and 4x higher MXFP8/MXFP4 performance. New CDNA 5 features include Tensor Data Mover for asynchronous data transfer, work-group clustering, multicast, data prefetching, split barriers, and a command processor for lower latency. The accelerators support spatial partitioning into one, two, or eight partitions (NPS1 and NPS2 modes). AMD also announced the Instinct MI430X for scientific computing, sovereign AI, and AI-HPC workloads, claiming 288 TFLOPS FP64 performance but not detailing memory configuration.