AGUASTO

AMD / Individual accelerator

AMD Instinct MI455X

AMD lists a July 2026 launch and published specifications. Reviewed Helios sources still describe partner volume deployments as expected in the second half of 2026; broad shipment availability is not independently established here.

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

Manufacturer specification reference

Architecture
CDNA 5AMD Instinct MI455X GPUs (external source) · AMD Instinct MI400 Series GPUs (external source) · AMD Helios rackscale reference design (external source)
Category
Individual acceleratorAMD Instinct MI455X GPUs (external source) · AMD Instinct MI400 Series GPUs (external source) · AMD Helios rackscale reference design (external source)
Accelerator memory
432 GBPer MI455X GPU, not the four-GPU compute tray or Helios rack.AMD Instinct MI455X GPUs (external source)
Memory type
HBM4AMD Instinct MI455X GPUs (external source) · AMD Instinct MI400 Series GPUs (external source) · AMD Helios rackscale reference design (external source)
Memory bandwidth
23.3 TB/sProduct-page peak GPU memory bandwidth per accelerator.AMD Instinct MI455X GPUs (external source)
Power
Not verified for this recordNo numeric module TDP/TBP was published in the reviewed MI455X product page or brochure.AMD Instinct MI455X GPUs (external source) · AMD Instinct MI400 Series GPUs (external source) · AMD Helios rackscale reference design (external source)
Precision formats
OCP-MXFP4, OCP-MXFP6, OCP-MXFP8, OCP-FP8, FP16, BF16, FP32, FP64, INT8AMD Instinct MI455X GPUs (external source)
Cooling / integration
Direct liquid cooling; Enhanced Accelerator Module (EAM).AMD Instinct MI455X GPUs (external source) · AMD Instinct MI400 Series GPUs (external source) · AMD Helios rackscale reference design (external source)

Configuration boundaries

  • The dedicated MI455X product specification supplies the headline bandwidth. Helios page contains different bandwidth text; retain source context and seek OEM confirmation.
  • No total-power estimate or claim of an operating Aguasto deployment is supplied.

Published arithmetic throughput

Manufacturer peaks, not Aguasto measurements or tokens per second
SpecificationPrecisionPublished peakBoundary & source
OCP-MXFP4 peak computeOCP-MXFP440.3 PFLOPSOne GPU; manufacturer peak matrix specificationManufacturer source ↗
OCP-FP8 peak computeOCP-FP820.1 PFLOPSOne GPU; manufacturer peak matrix specificationManufacturer source ↗

Deployment considerations

  • Confirm vendor shipment status and production software support before procurement planning.
  • Keep EAM GPU, compute-tray and Helios rack boundaries distinct in power and memory calculations.

Software ecosystem

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

Continue the research

AMD ROCmSGLangvLLMKubernetes GPU orchestration — NVIDIA device pluginModel weights versus runtime memoryCompare compatible hardware categoriesWhy modular infrastructure?

Sources & verification

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

  1. AMD Instinct MI455X GPUs ↗

    Launch date 2026-07-23; CDNA 5; 432 GB HBM4; 23.3 TB/s GPU memory bandwidth; Direct liquid cooling and EAM; Precision table; ROCm, HIP and listed SGLang support. Publication: date not stated in the reviewed source. Verified: 2026-10-12.

  2. AMD Instinct MI400 Series GPUs ↗

    Helios partner volume deployments expected in 2H 2026. Publication: date not stated in the reviewed source. Verified: 2026-10-12.

  3. AMD Helios rackscale reference design ↗

    Source inconsistency: compute tray paragraph states 19.6 TB/s while product highlight states 23.3 TB/s. Publication: date not stated in the reviewed source. Verified: 2026-10-12.