AMD

Helios

Rack-scale AI on open standards: 72 AMD Instinct GPUs in a single liquid-cooled rack, with pooled HBM4 memory and Ethernet-based networking.

Best for

Frontier model training, large mixture-of-experts models, and long-context, high-concurrency inference.

Talk to Us About Helios

GPUs per rack

Up to2.9

exaFLOPS FP4

31 TB

HBM4

Helios highlights

432 GB of HBM4 per GPU

Fits larger models and longer contexts on fewer GPUs.

UALink over Ethernet

Links all 72 GPUs at up to 260 TB/s, with no proprietary interconnect.

AMD ROCm

Supported in PyTorch, vLLM and SGLang.

AMD Helios configuration

Spec

Corvex build

GPUs

72 AMD Instinct MI455X per rack

CPUs

18 AMD EPYC "Venice"

GPU memory

Up to 31 TB HBM4, up to 1.7 PB/s

Compute

Up to 2.9 exaFLOPS FP4; 1.4 exaFLOPS FP8

Interconnect

Up to 260 TB/s scale-up; up to 43 TB/s scale-out

Deployment

Bare metal, Kubernetes, VMs or Slurm

Performance figures are AMD specifications.

Why Corvex

Deployed your way.

Bare metal, Kubernetes, VMs or Slurm, with physically isolated clusters available.

Secure by default.

SOC 2 and HIPAA compliant, in Tier III+ U.S. data centers we operate.

Run by operators.

Solutions engineering and 24/7 support, with on-site spares and proactive monitoring.

Which system is right for you?

Our engineers will help you match infrastructure to your workload.

Email sales