130 parts · 1,202 companies
Prices as of2026-10-25
Search

AMD Helios (MI455X): every supplier in the rack

72x AMD Instinct MI455X GPUs, EPYC "Venice" CPUs and Pensando networking on a double-wide OCP Open Rack Wide chassis — AMD's own reference numbers are 2.9 exaFLOPS FP4, 1.4 exaFLOPS FP8, 31 TB of HBM4, 260 TB/s of scale-up bandwidth and 43 TB/s of scale-out bandwidth. Unlike NVL72, Helios is a reference design built on open standards (OCP, UALink, Ultra Ethernet) rather than a single vendor's proprietary rack — AMD, HPE, Celestica, Dell, Lenovo, Supermicro, Sanmina and Wiwynn all build or sell systems on it.

Unmarked companies come from AMD's own Helios materials (amd.com, AMD Newsroom press releases); reported companies come from named trade press whose article body was read in full, not just the headline. This is a map of who plays where, not a confirmed bill of materials — Helios only began production in mid-2026 and most component-level suppliers below the rack-integration layer haven't been named publicly yet.

  1. Rack architecture and reference design (built on Meta's Open Rack Wide standard, submitted to OCP)
    Rack-scale systems 4.01
    AMD
  2. Rack integration, ODM and OEM systems
    Rack-scale systems 4.01 · EMS, system integration & rack-level test 4.02
    HPE (turnkey Helios systems worldwide in 2026), Celestica (scale-up switch R&D, design and manufacturing), Dell*, Lenovo*, Supermicro*, Sanmina*, Wiwynn*
  3. 18x compute trays (4x MI455X GPUs + 1x EPYC "Venice" CPU each)
    GPU baseboards & compute trays 4.03
    AMD
  4. MI455X GPU die (CDNA 5; compute chiplets on TSMC N2, fabric/cache and I/O chiplets on TSMC N3P)
    Merchant GPUs / accelerators 1.01 · Leading-edge foundry 1.09 · 2.5D/3D advanced packaging 1.11
    AMD, TSMC*
  5. HBM4 (432 GB per GPU, 31 TB per rack)
    HBM 2.01
    Samsung Electronics* (primary supplier as production began; AMD hasn't disclosed its full HBM4 supplier mix)
  6. EPYC "Venice" host CPU (Zen 6, up to 256 cores, 1.6 TB/s memory bandwidth)
    Server CPUs 1.05
    AMD
  7. Pensando "Vulcano" AI NIC (800 Gbps scale-out, PCIe Gen 6, OCP form factor, UALink to CPU/GPU)
    NICs / SuperNICs 1.06
    AMD
  8. Pensando Salina DPU (P4-programmable, 16 Arm N1 cores; front-end networking, storage and security offload)
    DPUs 1.07
    AMD
  9. Scale-up switch tray (2x Broadcom Tomahawk 6 ASICs per tray, UALink-over-Ethernet, 21.6 TB/s bidirectional per switch)
    Scale-up fabric 3.01 · Motherboards, CPU trays, switch trays 4.04
    Broadcom*, Celestica (R&D, design and manufacturing of the reference-design switch) · HPE's own Juniper-built switch variant (same Tomahawk 6 silicon, separate product)
  10. Switch-tray management (AMD Ryzen CPU + BMC, liquid cooled)
    BMC & platform firmware 4.10
    AMD

AMD's own materials describe a vertical-busbar power shelf and a quick-disconnect liquid cooling manifold inside the rack, but hadn't named those component suppliers publicly as of this research — outside the rack, facility-side power and cooling sit in Layers 6–7 same as for NVL72 — see Layer 6 and Layer 7. Compare with the GB300 NVL72 teardown.