←29 Sep 2026 · Vinod Bijlani

AMD Instinct in six years: from 32GB on a GPU to 31TB in a rack

AMD's data centre GPUs have moved from a single accelerator to a rack-scale system. The system around the accelerator now matters as much as the chip.

GPUsAMDAI factories
Part of the series: My Thoughts on AI factory build-outs →

In 2020, AMD's data centre GPU had 32GB of memory. In 2026, AMD has 31TB in a single rack. That jump says a lot about where AI infrastructure is heading.

AMD Instinct in six years:

  • 2020: MI100 introduces CDNA for data centre compute.
  • 2021: MI200 takes AMD into exascale HPC.
  • 2023: MI300 brings chiplets, 3D packaging and up to 192GB of HBM.
  • 2024: MI325X pushes memory to 256GB of HBM3E.
  • 2025: MI350 and MI355X move to CDNA 4 and add FP4 and FP6.
  • 2026: MI400 moves into the rack-scale era.

That last step is the interesting one. With Helios, AMD brings together 72 MI455X GPUs, 31TB of HBM4, EPYC CPUs, Pensando networking, a scale-up interconnect and ROCm in one rack-scale system.

The progression is clear: GPU → node → rack → AI factory.

The accelerator still matters. But increasingly, the system around it matters just as much: memory, interconnect, networking, power, cooling and software.

NVIDIA established much of the blueprint for modern accelerated computing. AMD is now building a serious alternative across the AI infrastructure stack. For enterprises, neoclouds and service providers, that means more choice across architecture, supply and economics, and a much more interesting GPU market.

Read moreDiscuss on LinkedIn ↗
AMD Instinct data centre GPU evolution, from MI100 in 2020 to the MI400 series and the Helios rack in 2026
Infographic · tap to enlarge

Invite Vinod to speak, write or advise.

Keynotes, panels and executive briefings on AI factories, sovereign AI and AI economics. Send a message on LinkedIn.