AMD Helios

Overview

AMD's first rack-scale AI system: 72 Instinct MI455X GPUs in 18 Open Rack Wide compute trays, 18 6th Gen EPYC Venice CPUs, Pensando Vulcano 800 AI NICs, UALink-over-Ethernet scale-up and Ethernet scale-out, running ROCm. A rack is specified at up to 2.9 exaflops FP4, 1.4 exaflops FP8, 31 TB of HBM4 and 1.7 PB/s of memory bandwidth. AMD claims up to 30% more inference tokens per dollar, 15% more peak FP4, and 50% more HBM than Nvidia's Vera Rubin NVL72.

Declared in full production on July 23, 2026; volume shipments targeted for late Q3 into Q4 2026. Named deployers include OpenAI, Anthropic, Meta, Microsoft, Oracle, HUMAIN, Tensorwave, Vultr and Cirrascale. OEM/ODM partners include HPE, Lenovo, Supermicro, Bull, Sanmina and Wiwynn.

Key Specs

GPUs per rack

72x Instinct MI455X plus 18x EPYC Venice CPUs

Peak rack performance

2.9 EF FP4 / 1.4 EF FP8; 31 TB HBM4

Claim vs Vera Rubin NVL72

Up to 30% more tokens per dollar; 50% more HBM capacity; 15% more peak FP4 (AMD Performance Labs, Jul 2026)