At its Advancing AI event, AMD unveiled the Instinct MI455X, its first GPU designed for rack-scale AI deployments, replacing the MI355X atop the AI stack. Built on the new CDNA5 architecture and packaged with TSMC's CoWoS-L, the MI455X features 256 Work Group Processors (WGPs) across 8 Accelerator Complex Dies, a 2.4 GHz max engine clock, and 432 GB of HBM4 memory at 23.3 TB/s bandwidth. Peak compute reaches 315 TFLOPS FP32/FP16 and 40.26 PFLOPS for OCP MXFP4.

Architectural shifts include Wave32 SIMD32 units (four per WGP), quadrupled VGPR capacity to 1,024 per wave, and matrix units capable of 65,536 FP4 ops/cycle/WGP. L1 Data Cache and LDS doubled to 64 KB and 320 KB respectively. A 192 MB Global L2 across two Fabric Cache Dies delivers 54 TB/s aggregate bandwidth. The CPU link moves from PCIe to a 16-lane Infinity Fabric with 256 GB/s bidirectional bandwidth.

The Helios rack-scale platform scales to 72 GPUs per rack using UALink over Ethernet (UAEoL) with 3.6 TB/s per GPU scale-up bandwidth. Six switch trays provide 260 TB/s aggregate scale-up bandwidth. Each of 18 compute trays holds four MI455X GPUs and a 96-core EPYC 9006 CPU. A full rack delivers up to 2.9 EFLOPS MXFP4 and 31 TB shared HBM4 at 1.7 PB/s. CDNA5 marks AMD's transition from the 15-year GCN lineage to an RDNA-based microarchitecture, unifying EPYC, Pensando, and Radeon IP.