Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

AMD reveals the MI325X, a 288GB AI accelerator built to battle Nvidia's H200

Jun 03, 2024 - theregister.com
AMD's flagship AI accelerator, the Instinct MI325X, is set to receive a high-bandwidth boost when it arrives later this year. The MI325X, which is similar to Nvidia's H200, will be composed of eight compute, four I/O, and eight memory chiplets using a combination of 2.5D and 3D packaging technologies. The accelerator's capacity will be boosted to 288GB, more than twice that of the H200 and 50 percent more than Nvidia's Blackwell chips. However, AMD's focus on FP16, which requires twice as much memory per parameter as FP8, may limit the advantage of its large memory capacity.

In response to Nvidia's Blackwell parts, AMD is moving to a yearly release cadence for new Instinct accelerators. The next-gen CDNA 4 compute architecture will stick to the same 288GB of HBM3e config as MI325X, but move to a 3nm process node for the compute tiles, and add support for FP4 and FP6 data types. AMD's future plans also hint at significant architectural upgrades, potentially involving heterogeneous multi-die deployments or photonic memory expansion.

Key takeaways:

  • AMD's flagship AI accelerator, the Instinct MI325X, will receive a high-bandwidth boost when it arrives later this year, with a memory capacity of 288GB, more than twice that of Nvidia's H200.
  • The MI325X's memory bandwidth will increase to 6TB/sec, a boost from the 5.3TB/sec of the MI300X and 1.3x more than the H200.
  • AMD is moving to a yearly release cadence for new Instinct accelerators to better compete with Nvidia's Blackwell parts.
  • AMD's next-gen CDNA 4 compute architecture will stick to the same 288GB of HBM3e config as MI325X, but move to a 3nm process node for the compute tiles, and add support for FP4 and FP6 data types.
View Full Article

Comments (0)

Be the first to comment!