AMD Introduces AI Accelerator Instinct MI455X with 432 GB HBM4 and Up to 40 Petaflops Performance

AMD has introduced specialized AI accelerators Instinct MI455X for data centers, based on the CDNA 5 architecture. Each accelerator is equipped with 432 GB of HBM4 memory with a peak bandwidth of up to 23.3 TB/s. The GPU contains 320 billion transistors and uses a combination of process technologies: eight ACD chiplets on the TSMC N2 process form 256 WGP compute units, along with Fabric and I/O chiplets manufactured on the TSMC N3P 3nm process. The memory is organized into 12 HBM4 stacks with a 192-channel interface, and the global L2 cache is 192 MB. The I/O subsystem supports two PCIe Gen 6 channels or three AMD AI-NIC network adapters via UALink. Peak performance reaches 20.13 PFLOPS in MXFP8/MXFP6 operations and 40.26 PFLOPS in MXFP4. Compared to the Instinct MI355X, the new accelerator offers 1.5 times more memory, 2.9 times higher bandwidth, and four times higher performance. The CDNA 5 architecture adds a Tensor Data Mover block for direct asynchronous data transfer, workgroup clustering, multicast transmission, and other improvements. The accelerators support NPS1 (interleaving across all HBM4 stacks) and NPS2 (split into two domains of six stacks each) modes. Also announced is the Instinct MI430X accelerator for scientific computing and AI-HPC with a performance of 288 TFLOPS in FP64.