AMD Instinct MI400 Series & Helios Rack

⚠ Disclaimer: This entry may be incomplete, out of date, or inaccurate. It is AI-maintained on a best-effort basis. Do not rely on it as a sole source — verify claims independently using the sources listed below.

Summary

The MI400 series is AMD’s fifth-generation CDNA accelerator line, announced at CES 2026 with a target production ramp in H2 2026. Three variants cover a range of use cases: MI430X (cost-optimized), MI440X (balanced), and MI455X (flagship). The MI455X anchors AMD’s Helios rack-scale system — a double-wide rack housing 72 MI455X GPUs with 31 TB of HBM4 memory and a stated 2.9 exaFLOPS FP4 aggregate throughput. The MI400 generation moves to TSMC 2nm, adopts HBM4, and introduces UALink for scale-up interconnect alongside UltraEthernet for scale-out.

Key Facts

Architecture: CDNA 5 (5th Gen)
Process node: TSMC 2nm (first GPU family on 2nm)
Variants: MI430X (entry), MI440X (mid), MI455X (flagship)
Memory per GPU (MI455X): 432 GB HBM4; 19.6 TB/s bandwidth
Peak compute per GPU (MI455X): ~40 PFLOPS FP4; ~20 PFLOPS FP8 (AMD stated, not independently verified)
Scale-up interconnect: UALink (AMD, open standard co-developed with industry consortium)
Scale-out interconnect: UltraEthernet (UEC spec), AMD Pensando Pollara NICs
Helios rack — GPU count: 72 MI455X GPUs
Helios rack — aggregate memory: 31 TB HBM4
Helios rack — aggregate compute: 2.9 exaFLOPS FP4; 1.4 exaFLOPS FP8
Helios rack — scale-up bandwidth: 260 TB/s aggregate
Helios rack — scale-out bandwidth: 43 TB/s Ethernet-based
Helios rack — form factor: Double-wide rack (wider than standard 19-inch)
Host CPU: Zen 6-based EPYC Venice processors
Status: Announced; production ramp H2 2026 (Q3 2026 per AMD public statements)
Key customer: OpenAI (disclosed June 2025, described as anchor customer)

What It Is / How It Works

The Helios system represents AMD’s most aggressive move into rack-scale integrated AI infrastructure — analogous to NVIDIA’s NVL72/NVL144 GB200 rack systems. Rather than selling individual GPUs and letting OEMs build systems, AMD (with partners) delivers Helios as a complete rack unit with GPUs, networking, cooling, and host CPUs pre-integrated.

Form factor: Helios uses a double-wide rack — wider than the standard 600mm 19-inch rack — to accommodate the thermal and electrical requirements of 72 MI455X GPUs. AMD has noted this requires operator planning for physical footprint in advance.

Scale-up interconnect (UALink): The MI400 generation transitions from proprietary Infinity Fabric (for scale-up within a rack) to UALink, an open-standard developed under the UALink Consortium. UALink is AMD’s answer to NVIDIA’s proprietary NVLink/NVSwitch — it achieves comparable bandwidth but under an industry-consortium governance model, theoretically allowing other vendors to implement compatible switches. At rack scale, Helios delivers 260 TB/s of UALink-based scale-up bandwidth for 72 GPUs.

Scale-out interconnect (UltraEthernet): Between Helios racks, AMD uses the UltraEthernet Consortium (UEC) standard over AMD Pensando Pollara NICs. UltraEthernet adds RDMA semantics, congestion control, and loss-recovery optimized for AI all-reduce traffic patterns onto standard Ethernet physical layer — again positioned as a CUDA-free, open alternative to InfiniBand.

HBM4: The MI455X’s 432 GB HBM4 per chip represents roughly 2.25x the memory of the NVIDIA B200 (192 GB). At rack scale, 31 TB of aggregate HBM4 in a single Helios rack is AMD’s central capacity argument for very large model deployment.

Process: TSMC 2nm is the most advanced commercially available node as of mid-2026, giving AMD a process advantage if production yields are viable.

Notable Developments

2026-06: AMD and OpenAI announce OpenAI as Helios/MI400 anchor customer; Lisa Su and Sam Altman appear together at AMD Advancing AI event — AMD frames the deal as validation of its rack-scale approach
2026-05: AMD unveils full MI400 product lineup at Advancing AI 2025 event; announces MI500 roadmap with claimed 1000x AI performance improvement over multi-year horizon
2026-02: AMD confirms Helios and MI400 on track for H2 2026 ramp in quarterly investor communications
2026-01: CES 2026 — AMD discloses MI430X, MI440X, MI455X product family and Helios rack architecture with specifications; announces double-wide rack form factor; confirms UALink and UltraEthernet as interconnect standards

Competitive Context

Helios targets NVIDIA’s GB200 NVL72 and NVL144 rack-scale systems directly. The NVL72 packages 72 GB200 GPUs in a liquid-cooled rack; NVIDIA quotes ~1.4 exaFLOPS FP8 per NVL72 rack. AMD claims 1.4 exaFLOPS FP8 and 2.9 exaFLOPS FP4 for Helios — roughly comparable FP8 numbers with an advantage at FP4, though FP4 precision applicability varies by model and workload.

Key competitive differences:

Memory: Helios 31 TB vs. NVL72 approximately 13.5 TB HBM3e — AMD claims a substantial lead here
Interconnect: UALink (open) vs. NVLink (proprietary) — AMD’s open-standards argument is relevant for buyers concerned about vendor lock-in
Availability: Helios targets Q3 2026; NVL72/144 B200 systems are already shipping — NVIDIA has a timing lead
Software: CUDA ecosystem vs. ROCm — NVIDIA retains a significant software maturity advantage

Claim Verification

Claim: Helios delivers 2.9 exaFLOPS FP4 per rack

Status: Unverified — no independent benchmark yet (product not yet shipping as of June 2026)

Supporting sources:

Tom’s Hardware — CES 2026 MI400 announcement — AMD stated spec at announcement
TechPowerUp — MI400 preview — corroborates AMD announcement figures

Refuting / questioning sources:

No independent third-party benchmark — product announced but not yet shipping
Peak FLOPS figures assume FP4 precision with no accuracy degradation guarantees and ideal interconnect conditions

Summary: All figures are AMD announcements; independent verification will require post-shipment MLPerf or third-party benchmarks expected late 2026.

Claim: OpenAI is an anchor customer for Helios

Status: Partially verified