TL;DR
- Announced 2025 as AMD's next-generation accelerator family targeting 2026 deployment.
- Headline pitch: pod-scale 'Helios' rack architecture coupling MI400 GPUs with EPYC CPUs and Pensando networking.
- Positioned against NVIDIA Vera Rubin and GB300 NVL72 at the rack level.
- Public detail remained limited at time of writing; treat all specific numbers as preliminary.
Overview#
MI400 is AMD's roadmap answer to NVIDIA's NVL72-class rack architectures. The pitch is system-level: instead of competing on GPU SKU alone, AMD's 'Helios' platform couples MI400 GPUs with EPYC CPUs, Pensando DPUs, and a refreshed Infinity Fabric to deliver a coherent rack-scale unit.
At the time of writing, public detail on MI400 silicon and configurations remained limited. AMD has communicated the general direction — chiplet-based, HBM4, FP4-native — but production specifications will firm up through 2026.
Preliminary Specifications#
| Metric | MI400 (preliminary) |
|---|---|
| Architecture | CDNA-Next |
| Memory | HBM4 (capacity TBD) |
| Bandwidth | Multi-TB/s tier |
| FP4 (Matrix) | Multi-PFLOP target |
| System | Helios rack architecture |
| Form factor | OAM (likely) |
MI400 specifications are preliminary and subject to substantial revision before launch. Use this entry for direction-setting; revisit at launch for production numbers.
The Helios Platform#
Helios is AMD's name for the rack-scale system around MI400. The idea mirrors NVIDIA NVL72: bring CPU, GPU, fabric and management plane into a single integrated rack, eliminating the seams that show up when separately-sourced components have to coexist.
Pensando DPUs handle north-south networking and security offload; EPYC CPUs handle dataloading and orchestration; Infinity Fabric scales the GPU domain inside the rack. The competitive question is whether AMD can match NVIDIA's NVLink Switch System bandwidth and latency, not just the GPU peak FLOPS.
When MI400 Will Matter#
- Hyperscale operators seeking a credible second source at rack scale, not just per-GPU.
- Sovereign and regional AI builds where supply diversification is a strategic requirement.
- Workloads where ROCm has already matured to production parity for the specific stack.
- For 2026 production decisions, MI400 is a planning item — H200/B200/MI325X/MI355X remain the deployable choices today.
Pitfalls#
- Plans change: AMD roadmaps have historically slipped or rebinned, particularly on first-generation rack platforms.
- Software maturity: rack-level orchestration on AMD lacks the Mission Control / NVL72 management plane experience NVIDIA has invested in.
- Avoid betting production on pre-launch specifications — wait for shipping silicon before committing capacity.
Software Outlook#
MI400 will require ROCm 8 or later. AMD has been investing heavily in PyTorch, vLLM, SGLang and JAX ROCm backends; whether MI400 launches with parity FP4 performance against Blackwell will determine its competitive footprint through 2026-2027.
References
- AMD Data Center Roadmap · AMD