HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

Instinct MI450 / HeliosAMD—Instinct MI450 / HeliosAMD—

AMD

Instinct GPUs, Helios racks, EPYC CPUs

USpublic~28,000 staffest. 1969amd.com (opens amd.com)Checked 25 Sep

The brief

Analyst view
  • OpenAI agreed to deploy 6 GW of AMD GPUs (Oct 2025) and got a warrant for up to 160M AMD shares.
  • Helios racks with 72 GPUs launched in 2026. First OpenAI capacity due online in Q4 2026.
  • Bought ZT Systems (closed March 2025) to learn how to build whole racks.

Technical approach

As reported
Chiplets

Modular chiplet GPUs

GPUs are built from small chips stacked on base chips. This keeps yields high and lets AMD scale up.

Memory

More HBM per GPU

AMD often leads on memory per GPU (192 GB, then 288 GB, then 432 GB), so big models fit on fewer GPUs.

Open stack

ROCm and open standards

ROCm is AMD's open-source rival to CUDA. It backs open standards to link GPUs together.

Programs

AMD, two programs

Updated 1 Jul 2026
Instinct MI450 / HeliosAMD's first full-rack AI system: 72 MI455X GPUs with HBM4 memory, AMD CPUs and networking. Built to compete with NVIDIA's Vera Rubin rack.PilotNextFirst OpenAI Helios capacity online · Q4 2026Updated 1 Jul 2026View program →Instinct MI300 and MI350AMD's 2023-2025 data-center GPUs, built from stacked chiplets and known for lots of memory. The MI355X has 288 GB of HBM3E.ScalingUpdated 12 Jun 2025View program →
Research · Proto · Pilot · Deploy · Scaling

Research & demos

2 items
  • Sep 2025
    codeROCm 7Instinct MI450 / Helios

    Big software release adding MI350 support and faster inference.

  • Aug 2024
    talkMI300X architecture (Hot Chips 2024)Instinct MI300 and MI350

    Explained AMD's design of GPUs built from stacked chiplets.

Open technical problems

Open · Active · Partial
  • Software parity with CUDAActive research

    ROCm is improving fast but still trails CUDA in tools and out-of-the-box speed on new models.

  • First-generation rack-scale systemsOpen

    Helios is AMD's first 72-GPU rack. Networking, cooling and reliability at huge scale are unproven.

Peers

AI chips · 4

Company & funding

Staff
~28,000as of 2024
HQ
US
Founded
1969

Partners & customers

3 customers · 1 supplier · 1 partner
  • Customer
    OpenAI

    6 GW multi-generation GPU agreement (Oct 2025).

  • Supplier
    TSMC

    Fabricates AMD chiplets.

  • Customer
    Oracle

    Planned 50,000 MI450 supercluster from Q3 2026.

  • Customer
    Meta

    Instinct GPU customer.

  • Partner
    Sanmina (opens sanmina.com)

    Bought ZT Systems' manufacturing arm and builds AMD racks.

People

Key people
Primer