AMD Instinct MI500 Series (CDNA 6)
Product Overview
The AMD Instinct MI500 series is the next-generation flagship AI accelerator after the MI400 series (CDNA 5). It was formally previewed by AMD CEO Lisa Su at Advancing AI 2026 (2026-07-23, San Francisco), with an official launch expected in 2027.
The defining trait of the MI500 is that it is jointly defined and developed by AMD and OpenAI: based on the real compute demands of its frontier large models, OpenAI feeds AMD operational data such as workload characteristics, memory capacity, interconnect bandwidth, and power efficiency, which in turn guide the chip architecture, memory specifications, and software stack adaptation. According to Lisa Su, from the 2023 MI300X to the 2027 MI500, AMD aims to achieve a thousandfold leap in AI performance within four years.
🔮 Status: forward-looking / in development. The specifications below are based on the architecture generation and year disclosed at Advancing AI 2026; concrete compute, memory capacity, TDP, etc. have not yet been announced — final figures are subject to AMD's official announcements.
Core Specifications (Disclosed at Advancing AI 2026, Pending Final Confirmation)
| Parameter | Value |
|---|---|
| Architecture | CDNA 6 (new generation; the MI400 is CDNA 5) |
| Process | TSMC 2nm |
| Memory Type | HBM4E (the MI400 uses HBM4) |
| Memory Capacity | To be confirmed (TBD) |
| Memory Bandwidth | To be confirmed (TBD) |
| FP4 Compute | To be confirmed (TBD) |
| FP8 Compute | To be confirmed (TBD) |
| FP16 Compute | To be confirmed (TBD) |
| TDP | To be confirmed (TBD) |
| Interconnect | To be confirmed (TBD; next-generation UALink expected) |
| Availability | 2027 (expected) |
MI500 vs MI455X (Generational Expectations)
| Metric | MI455X (CDNA 5, current) | MI500 (CDNA 6, forward-looking) | Change |
|---|---|---|---|
| Architecture | CDNA 5 | CDNA 6 | New generation |
| Process | TSMC N2 (2nm) | TSMC 2nm (advance on the same node) | Node optimization |
| Memory | 432 GB HBM4 | HBM4E | Next-generation memory |
| Performance target | Baseline | Thousandfold AI performance vs the MI300X | Within 4 years |
| Launch | 2026 H2 | 2027 (expected) | About 1 year later |
⚠️ For the MI500 in the table above, only the architecture / year / memory type are confirmed; concrete compute and capacity are subject to the official launch.
The OpenAI Joint Development Model
The MI500 is AMD's first chip product jointly defined with a leading AI lab (OpenAI):
- Since October 2025, OpenAI and AMD have had a multi-year, multi-generation strategic agreement that locks in a combined 12GW+ GPU purchase commitment
- OpenAI will deploy a total of 6GW of AMD Instinct GPUs, with the first 1GW based on the MI450 series (online in 2026 H2)
- AMD has issued OpenAI warrants for up to 160 million shares of common stock (strike price $0.01/share); if fully exercised, OpenAI would hold about 10% of AMD
- Based on real large-model workloads, OpenAI feeds back requirements on memory capacity, interconnect bandwidth, and power efficiency to guide MI500 architecture design
Vendor Information
| Item | Details |
|---|---|
| Vendor | AMD Corporation |
| First Preview | Advancing AI 2026 (2026-07-23, San Francisco) |
| Expected Launch | 2027 |
| Positioning | Successor flagship to the MI400 series |
| Joint Development | OpenAI |
| Product Page | https://www.amd.com/en/products/accelerators/instinct.html |
Related Products
- AMD MI400 series - Current generation (CDNA 5, Helios rack)
- AMD MI455X - MI400 flagship model
- NVIDIA Rubin R200 - Contemporary competitor
- NVIDIA Rubin Ultra - Same-generation successor preview
- Full comparison table