AMD Instinct MI440X (CDNA 5, Enterprise 8-GPU)
Product Overview
AMD Instinct MI440X is the enterprise SKU of the Instinct MI400 series, disclosed at CES 2026 (2026-01-06) alongside the MI400 series. It is positioned for standard 8-GPU local server deployments (paired with single/dual-socket AMD EPYC Venice / Zen 6 CPUs), complementing the flagship MI455X aimed at rack-scale Helios platforms.
As of August 2026, AMD has released neither an official specification sheet for the MI440X nor an official launch; the parameters below are extrapolated from public MI400 series information and the same-generation MI455X platform, marked as estimates, and will be updated after the official release.
Core Specifications
| Parameter | Value (estimated / pending official confirmation) |
|---|---|
| Architecture | CDNA 5 |
| Process | TSMC N2 (2nm) compute chiplet + N3 (3nm) auxiliary chiplet (same source as MI455X) |
| Memory | 432 GB HBM4 (inherited from the MI400 series platform) |
| Memory Bandwidth | Estimated ~19.6–23.3 TB/s |
| FP4 Compute | Estimated ~40 PFLOPS (MXFP4, same generation) |
| FP8 Compute | Estimated ~20 PFLOPS (MXFP8) |
| FP64 Compute | Estimated 5 TFLOPS (not optimized for sovereign AI, below the MI430X's 288 TFLOPS) |
| Form Factor | Standard 8-GPU server (OAM / PCIe, paired with EPYC Venice) |
| Interconnect | Infinity Fabric (to the EPYC Venice CPU, 256 GB/s bidirectional) + UALink scale-out |
| TDP | Not disclosed |
| First Disclosure | 2026-01-06 (CES 2026) |
| Launch Status | Disclosed, not officially launched, no official specification sheet |
📌 The MI440X belongs to the same CDNA 5 + 432GB HBM4 platform as the MI455X / MI430X; the differences are mainly in deployment form (8-GPU server vs Helios rack vs HPC/sovereign AI) and FP64 double-precision capability (the MI440X is not FP64-optimized, while the MI430X reaches 288 TFLOPS).
Positioning Within the MI400 Series
| SKU | Positioning | Deployment Form | FP64 | Status |
|---|---|---|---|---|
| MI455X | Frontier training + inference (flagship) | Helios rack (72 cards) | ~5 TFLOPS | Announced 2026-07, shipping H2 |
| MI440X | Enterprise training + inference | Standard 8-GPU server | ~5 TFLOPS | Disclosed at CES, not launched |
| MI430X | HPC / sovereign AI | Supercomputers / sovereign cloud | 288 TFLOPS | Shipping 2027 |
| MI350P | Enterprise air-cooled inference | Air-cooled PCIe (CDNA 4) | — | Released |
Use Cases
- ✅ Enterprise on-premises AI training / inference (standard 8-GPU server, no rack retrofit needed)
- ✅ Private cloud / government and enterprise intelligent computing (in tandem with EPYC Venice)
- ❌ Hyperscale rack-level deployments (choose MI455X + Helios)
- ❌ High-precision HPC / sovereign supercomputing (choose the MI430X)
Vendor Information
| Item | Details |
|---|---|
| Vendor | AMD Corporation |
| First Disclosure | 2026-01-06 (CES 2026) |
| Product Page | https://www.amd.com/en/products/accelerators/instinct.html |
| Launch Status | Disclosed, not officially launched |
Related Products
- AMD MI455X - MI400 series flagship (Helios rack)
- AMD MI430X - MI400 series HPC / sovereign AI
- AMD MI350P - MI400 series enterprise air-cooled variant (CDNA 4)
- AMD MI400 - Series overview
- NVIDIA Rubin R200 - Same-generation competitor
- Full comparison table