AMD’s Instinct MI455X: Aiming for the Sun

Hello you fine Internet folks, here at AMD’s Advancing AI event we are looking at AMD’s brand new Instinct MI455X, replacing the older Instinct MI355X at the top of their AI stack. It is AMD’s first GPU designed for rack-scale AI deployments and is based on the new CDNA5 architecture with major changes to the compute unit and SoC, including enhancements to compute performance, improved memory bandwidth, larger memory capacity, and packaged using TSMC’s CoWoS-L.Instinct MI455X packageIt comes along with the new Helios rackscale solution, enabling scaling to 72 GPUs in a single pod, up from 8 GPUs for previous MI355X systems. It also features a new fault-tolerant scale-up networking architecture based on UALink over Ethernet (UAEoL) with single-hop all-to-all communication across the entire rack.AMD Helios rackA single MI455X contains 256 Work Group Processors (WGPs) across 8 Accelerator Complex Dies (XCDs), with a max “engine” clock of 2.4GHz. This enables peak compute figures ranging…