HomeFull speed ahead. Keep track of advancements in every frontier
Instinct MI450 / HeliosAMD—Instinct MI450 / HeliosAMD—
Instinct MI300 and MI350
Photo: 极客湾Geekerwan · CC BY 3.0 (opens commons.wikimedia.org)

Instinct MI300 and MI350

by AMDUS

ScalingStage 5 of 5

MI355X shipping to clouds and AI labs since mid-2025.

Updated 12 Jun 2025Checked 25 Sep0 updates this week

Milestones

No announced next step
  1. MI300X launched6 Dec 2023Complete.
  2. MI325XOct 2024Complete.
  3. MI350X and MI355X launched12 Jun 2025Complete.

Most important updates

  • 12 Jun 2025
  • Apr 2025
  • 18 Nov 2024
  • Aug 2024
  • 6 Dec 2023

Current obstacles

  • Export restrictionsUS rules on MI308 sales to China led AMD to take a charge of about $800M in 2025.

Physics limits

  • Crossing between chiplets costs energySplitting a GPU into chiplets improves yield, but every bit that crosses from one die to another costs more energy and time than a wire on a single die. That limits how finely a design can be split.
  • Stacking traps heatCompute dies sit on top of I/O dies, so heat from the upper layer must pass through or around the lower one. Stacked designs must run cooler or slower than flat ones.
  • Big dies rarely come out flawlessDefects land randomly on wafers, so the chance a die has none falls exponentially with its area. That statistic, not taste, is why large GPUs are built from chiplets.

How it works

3 parts
An HPE Cray blade from El Capitan: under each copper cold plate sits an MI300A, CPU and GPU chiplets sharing one package and memory
An HPE Cray blade from El Capitan: under each copper cold plate sits an MI300A, CPU and GPU chiplets sharing one package and memoryPhoto: Geekerwan (极客湾) · CC BY 3.0 (opens commons.wikimedia.org)
Chiplets

Compute stacked on I/O

Compute chiplets are bonded on top of I/O dies, and eight HBM stacks surround them. Small chiplets yield better than one giant die.

Memory

More memory per GPU

MI300X carries 192 GB and MI355X 288 GB of HBM, so a large model fits on fewer GPUs.

APU

CPU and GPU in one package

MI300A swaps some GPU chiplets for CPU cores sharing the same memory. It powers the El Capitan supercomputer.

Papers & demos

1 item
  1. Aug 2024talk
    MI300X architecture (Hot Chips 2024)Explained AMD's design of GPUs built from stacked chiplets.

Update log

5 updates

Thu 12 Jun 2025

  • Major: PressHardware

Apr 2025

  • Major: FilingRegulatory

Mon 18 Nov 2024

  • Major: PressHardware

Aug 2024

  • Minor: BlogTest

Wed 6 Dec 2023

  • Major: PressHardware

About AMD

The team behind Instinct MI300 and MI350

AMD

Instinct GPUs, Helios racks, EPYC CPUs

AMD designs CPUs and GPUs and is NVIDIA's main rival in AI chips. It pioneered building chips from smaller pieces (chiplets). Its Helios rack, backed by a 6 GW OpenAI deal, is its first full-rack AI system.

  • 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.
Founded
196957 yrs
Headquarters
United States
Status
Public
Valuation
public
Staff
~28,000Dec 2024
Works in
ComputingAI chipsAI
Coverage
2 programs · 10 updateslatest 1 Jul 2026checked 25 Sep
Customers
OpenAIOracleMeta
Suppliers
TSMC
Primer