AMD / Rack-scale system

AMD Helios

AMD explicitly identifies Helios as a reference design rather than an AMD product for sale. The reviewed page expects OEM/ODM volume deployments in the second half of 2026.

Original Aguasto conceptual illustration. Not manufacturer geometry, an operating Aguasto asset, or a measured benchmark.

Manufacturer specification reference

Accelerator memory
31,000 GBPublished rounded aggregate up to 31 TB GPU HBM converted to decimal GB; CPU memory excluded.AMD Helios Rackscale Solution brochure (external source)
Memory bandwidth
1,670 TB/sBrochure's up to 1.67 PB/s aggregate GPU memory bandwidth converted to decimal TB/s; not scale-up interconnect bandwidth.AMD Helios Rackscale Solution brochure (external source)
Power
Not verified for this recordNo operating rack draw verified in reviewed Helios sources; power shelf description is not an operating-power measurement.AMD Helios rackscale reference design (external source) · AMD Helios Rackscale Solution brochure (external source)
Precision formats
OCP-MXFP4, OCP-MXFP6, OCP-MXFP8AMD Helios Rackscale Solution brochure (external source)

Configuration boundaries

  • Rounded manufacturer rack totals are retained, rather than presenting exact capacity derived from GPU count.
  • Bandwidth in this record is the brochure aggregate; the webpage separately quotes a per-GPU peak. Interconnect throughput is a different metric.
  • Manufacturer theoretical throughput is not a benchmark or a delivered token-rate claim.

Published arithmetic throughput

Manufacturer peaks, not Aguasto measurements or tokens per second
SpecificationPrecisionPublished peakBoundary & source
OCP-MXFP4 peak computeOCP-MXFP42.9 EFLOPSWhole reference rack; manufacturer peak theoretical specificationManufacturer source ↗
OCP-MXFP8 peak computeOCP-MXFP81.4 EFLOPSWhole reference rack; manufacturer peak theoretical specificationManufacturer source ↗

Deployment considerations

  • An OEM implementation needs its own electrical, thermal, networking and acceptance specification.
  • Rack memory is distributed across GPUs; compare the reference rack only with other complete rack configurations.
  • Track OEM availability separately from AMD's reference-design publication.

Software ecosystem

ROCm. Confirm the exact model, runtime release, server integration and vendor support matrix before evaluating a deployment.

Sources & verification

Reference revision 1 · Reviewed . Manufacturer claims and documentation are attributed; Aguasto has not run the cited products in production.

  1. AMD Helios rackscale reference design ↗

    72 MI455X GPUs; Helios is a reference design not an AMD product for sale; Volume deployments expected in 2H 2026; ROCm ecosystem; Liquid-cooled Open Rack Wide architecture. Publication: date not stated in the reviewed source. Verified: 2026-10-12.

  2. AMD Helios Rackscale Solution brochure ↗

    72 GPUs in 18 four-GPU compute trays; Up to 31 TB HBM4; Up to 1.67 PB/s aggregate GPU memory bandwidth; Liquid cooling; ROCm; Up to 2.9 EFLOPS OCP MXFP4 and 1.4 EFLOPS OCP MXFP6/MXFP8. Publication: date not stated in the reviewed source. Verified: 2026-10-12.