In the space of just a few months, AMD has put together a new MI300X GPU for its Instinct platform that comes with a bevy of welcome upgrades. MI325X uses the same CDNA 3 architecture as its predecessor, MI300X, but features like HBM3E as well as improvements to compute performance make all the difference. Of course, these changes also make AMD more competitive in AI and HPC versus Nvidia with its H200 GPU.
Switching over to HBM3E nets MI325X a memory bandwidth of 6TB/s, besting MI300’s 5.3TB/s and H200’s 4.8TB/s. AMD isn’t stopping at speed, though, as it equips its new GPU with 288GB of memory, a 50% uplift compared to MI300.
Of course, these GPUs are designed to work in clusters of eight. As such, the AMD Instinct MI325X platform provides 10.4 petaflops (PF) of BP16/FP16 performance and up to 2.3TB of HBM3E memory, all working together via AMD Infinity Fabric with a staggering 896GB/s of bandwidth.
MI325X will be available Q4 2024. Looking to the future, though, AMD also shared some details about its plans for future architectures, namely ‘CDNA 4’ and ‘CDNA Next’, which should arrive in a new annual cadence.
So, in 2025, all things being well, CDNA 4 hits the scene. AMD plans to use a 3nm process node to build GPUs with this architecture, and naturally continue using HBM3E memory with 288GB capacity per GPU. CDNA 4 will also be the first architecture to support FP4 and FP6 datatypes, with continued support for FP8 and FP16. Briefly, while CDNA Next should hit the market in 2026, AMD had little to say about it other than it will be a “next-gen architecture.”
If you’re hungry for more Computex news, we’ve got everything AMD covered, and more. Check out our thoughts on the new Ryzen 9000 series desktop processors and 300 AI series mobile chips.