Milestones
- Helios previewed at Advancing AI12 Jun 2025Complete.
- OpenAI 6 GW agreement6 Oct 2025Complete.
- Oracle 50,000-GPU MI450 plan announced14 Oct 2025Complete.
- Helios launched at Advancing AI 2026Jul 2026Complete.
- First OpenAI Helios capacity onlineTarget Q4 2026Current milestone.
- Instinct MI500 generationTarget 2027Not yet reached.
Most important updates
- Jul 2026
- 14 Oct 2025
- 6 Oct 2025
- 31 Mar 2025
Upcoming
- Q4 2026First OpenAI Helios capacity online (next)
- 2027Instinct MI500 generation
Current obstacles
- Scale-up fabric maturityLinking GPUs over Ethernet is newer than NVLink. It must prove it stays fast under heavy AI traffic.
- HBM4 supplyAMD competes with NVIDIA for limited HBM4 output from SK hynix, Samsung and Micron.
Physics limits
- Huge currents at low voltageGPUs run below 1 volt, so a chip drawing over 1,000 W pulls more than 1,000 amps. Delivering that without big losses needs thick copper and voltage conversion right beside the chip.
- Memory density grows slowlyMemory per GPU grows only as fast as DRAM density and stack height, and DRAM cells have shrunk slowly in recent generations. Bigger models therefore need more GPUs, not just bigger ones.
- Heat per rackA rack of 72 top GPUs dissipates well over 100 kW. Liquid cooling can carry it away, but hot spots on each chip are approaching the limit of what cold plates can remove.
How it works
GPU
Chiplet GPU with HBM4
Each MI455X joins compute chiplets with 12 HBM4 memory stacks, more memory than rival GPUs.
Scale-up
UALink over Ethernet
GPUs in the rack talk over standard Ethernet switches using the open UALink protocol.
Scale-out
Pensando NICs
AMD's Pensando network cards link racks into clusters of many thousands of GPUs.
Spec sheet
| Spec | Instinct MI450 / Helios | MI355X | Change |
|---|---|---|---|
| HBM4 per GPU | 432 GBR (reported) | 288 GB HBM3E | ▲ 144 GB |
| Memory bandwidth per GPU | 19.6 TB/sR (reported) | 8 TB/s | ▲ 11.6 TB/s |
| FP4 compute per GPU | 40 PFLOPSR (reported) | — | —no comparable change |
| GPUs per rack | 72R (reported) | — | —no comparable change |
| Rack FP4 compute | 2.9 exaflopsR (reported) | — | —no comparable change |
| Rack HBM4 | 31 TBR (reported) | — | —no comparable change |
R reported by the company
Papers & demos
- Sep 2025codeROCm 7Big software release adding MI350 support and faster inference.
Update log
Jul 2026
- Major: PressHardware
Tue 14 Oct 2025
- Minor: PressCommercial
Mon 6 Oct 2025
- Major: PressCommercial
Mon 31 Mar 2025
- Minor: PressOther
About AMD
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
- Listed
- AMD (opens google.com)NASDAQ
- Valuation
- public
- Staff
- ~28,000Dec 2024
- Coverage
- 2 programs · 10 updateslatest 1 Jul 2026checked 25 Sep
- Partners
- Sanmina (opens sanmina.com)
- Suppliers
- TSMC
- People
- Lisa Su (opens en.wikipedia.org)Chair and CEOMark Papermaster (opens en.wikipedia.org)CTO