NEWHow AI models draw the world map · Gecko Tests
Beta
ChipsReading · ~3 min · 54 words deep

AMD MI400

AMD Instinct MI400 is the data-center GPU series AMD has said is planned for 2026, with HBM4 memory, following the MI300 and MI350 generations.

Text reviewed October 5, 2026

MI400 on hardware map
TL;DR

AMD Instinct MI400 is the data-center GPU series AMD has said is planned for 2026, with HBM4 memory, following the MI300 and MI350 generations.

Level 1

AMD MI400 is the successor to MI325X in AMD's data-center GPU roadmap. Uses HBM4 memory (vs HBM3e on MI325X), targets 2.5-3× FP8 throughput of MI325X. Expected to ship H2 2026. Major cloud partners: Microsoft Azure, Meta, Oracle all announced MI400 commitments. Software stack: ROCm 6.x+ with pytorch-level parity claimed.

Level 2

MI400 is AMD's attempt to close the gap with NVIDIA on AI training. Roadmap context: MI300X (2023), MI325X (2024), MI400 (2026). Each generation targets 2× perf vs predecessor. HBM4 brings 1.5-2 TB/s bandwidth per stack, making memory-bound workloads (LLM inference, training gradient ops) faster. Packaging: CoWoS-L advanced packaging to increase die + HBM count.

Level 3

ROCm software stack remains the adoption gate · MI400 viability depends on ROCm parity for major frameworks (PyTorch, JAX, DeepSpeed, vLLM). Infinity Fabric interconnect v4 scales to 256 GPUs in a rack with NVLink-equivalent bandwidth. Key customers: Meta, Microsoft, Oracle have placed large MI400 orders. MI400 will be the flagship AMD training chip through 2027.

Why this matters now

MI400 is AMD's best shot at breaking NVIDIA's AI training moat · H2 2026 shipment is the critical test.

The takeaway for you
If you are a
Curious · Normie
  • ·AMD's new powerful AI chip · competes with NVIDIA's best
  • ·Coming out in 2026
  • ·Bought by Microsoft and Meta to reduce dependence on NVIDIA
If you are a
Builder
  • ·Access via Azure, Oracle, Meta clouds (once GA)
  • ·Requires ROCm-compatible workload
  • ·Best if you want NVIDIA alternative for cost or supply reasons
If you are a
Investor
  • ·AMD's flagship AI training chip · competitive signal
  • ·ROCm parity still the gating factor
If you are a
Researcher
  • ·Successor to MI325X · HBM4 · CoWoS-L
  • ·2-3× FP8 throughput of MI325X
  • ·Infinity Fabric v4 · 256-GPU racks
Gecko's take

MI400 is AMD's make-or-break moment · ROCm parity determines whether it ships as compute or just as leverage.

H2 2026. Major clouds committed. Samples flowing to select partners.

Glossary

All terms · A-ZMore ChipsLearning paths

Explore live data

Models leaderboardBenchmarksAI Bubble Index

Cite or embed

MethodologyAPI
Next · MI300X