AMD Instinct MI430X (CDNA 5, HPC / Sovereign AI)
Product Overview
AMD Instinct MI430X is the MI400 series HPC / sovereign AI SKU announced at Advancing AI 2026 (July), targeting high-precision scientific computing and national-scale compute projects. It shares the CDNA 5 architecture and the 432GB HBM4 / 23.3TB/s memory configuration with the MI455X; the biggest difference is native hardware-level FP64: 288 TFLOPS (8.7x that of the NVIDIA Vera Rubin).
The MI430X has been locked in for two national supercomputing systems:
- 🇺🇸 US ORNL "Discovery" (Oak Ridge National Laboratory, deployment in 2028)
- 🇪🇺 France's "Alice Recoque" (built jointly by GENCI + CEA)
AMD officially confirmed on the August 4 earnings call that the MI430X is already running in the ORNL and Jean Zay (France) exascale supercomputers. Shipping 2027 H1.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | CDNA 5 |
| Process | TSMC N2 (2nm) compute chiplet + 3nm auxiliary chiplet (same as MI455X) |
| Memory | 432 GB HBM4 |
| Memory Bandwidth | 23.3 TB/s |
| FP64 Compute (hardware) | 288 TFLOPS (8.7x that of Vera Rubin) |
| FP32 Compute | ~250 TFLOPS (estimated) |
| MXFP4 Compute | ~40 PFLOPS (estimated, same as MI455X) |
| Positioning | HPC / sovereign AI (high-precision scientific computing first) |
| TDP | Estimated 1,200-1,500 W |
| Availability | Shipping 2027 H1 |
📌 The MI430X is a "precision-first" design: it sacrifices part of the AI throughput in exchange for FP64 double-precision performance, complementing the MI455X (AI throughput first) within the same Helios ecosystem.
MI430X vs MI455X (Within the MI400 Series)
| Metric | MI430X | MI455X |
|---|---|---|
| Positioning | HPC / sovereign AI | Frontier training + inference |
| Architecture | CDNA 5 | CDNA 5 |
| Memory | 432 GB HBM4 | 432 GB HBM4 |
| Memory Bandwidth | 23.3 TB/s | 23.3 TB/s |
| FP64 | 288 TFLOPS (hardware) | 33 TFLOPS (estimated) |
| MXFP4 | ~40 PFLOPS (estimated) | 40 PFLOPS (dense) |
| Supercomputer deployments | ORNL Discovery / Alice Recoque / Jean Zay | Helios racks |
| Shipping | 2027 H1 | From 2026 H2 |
Typical Applications
- Flagship scientific computing: molecular dynamics, climate simulation, astrophysics
- Sovereign AI: national-scale compute infrastructure construction
- High-precision HPC: FP64 double-precision simulation (CFD / structural analysis)
- CPU+GPU hybrid computing: heterogeneous supercomputing nodes
Use Cases
- ✅ Scientific computing / HPC (288 TFLOPS FP64 as the core advantage)
- ✅ Sovereign AI supercomputers (national-level procurement)
- ✅ High-precision simulation (finance / energy / defense)
- ❌ Pure AI training throughput-first scenarios (choose the MI455X)
- ❌ Enterprise installed-base rack upgrades (choose the air-cooled MI350P)
Vendor Information
| Item | Details |
|---|---|
| Vendor | AMD Corporation |
| Official Announcement | Advancing AI 2026 (2026-07) |
| Product Page | https://www.amd.com/en/products/accelerators/instinct.html |
| Shipping | 2027 H1 |
Related Products
- AMD MI455X - MI400 series flagship (AI training)
- AMD MI350P - MI400 series enterprise variant (air-cooled PCIe)
- AMD MI400 - Series overview
- NVIDIA Rubin R200 - Same-generation competitor
- Full comparison table