AGUASTO

AMD / Individual accelerator

AMD Instinct MI355X

Launched accelerator with manufacturer-reported deployment/benchmark evidence; supplier stock and Aguasto availability are not verified.

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

Manufacturer specification reference

Architecture
CDNA 4AMD Instinct MI355X specifications (external source) · AMD Instinct GPUs: MLPerf Inference v5.1 (external source) · AMD Instinct MI355X GPU brochure (external source)
Category
Individual acceleratorAMD Instinct MI355X specifications (external source) · AMD Instinct GPUs: MLPerf Inference v5.1 (external source) · AMD Instinct MI355X GPU brochure (external source)
Accelerator memory
288 GBPer OAM accelerator; not the eight-GPU baseboard or host memory.AMD Instinct MI355X specifications (external source)
Memory type
HBM3eAMD Instinct MI355X specifications (external source) · AMD Instinct GPUs: MLPerf Inference v5.1 (external source) · AMD Instinct MI355X GPU brochure (external source)
Memory bandwidth
8 TB/sManufacturer peak theoretical GPU memory bandwidth per accelerator.AMD Instinct MI355X specifications (external source)
Power
1,400 WManufacturer maximum board-power specification, corroborated by brochure; not measured server draw.AMD Instinct MI355X specifications (external source)
Precision formats
FP64, FP32, FP16, BF16, INT8, OCP-FP8, MXFP8, MXFP6, MXFP4AMD Instinct MI355X specifications (external source)
Cooling / integration
Manufacturer lists passive and active cooling; verify the actual liquid-cooled/OEM system design.AMD Instinct MI355X specifications (external source) · AMD Instinct GPUs: MLPerf Inference v5.1 (external source) · AMD Instinct MI355X GPU brochure (external source)

Configuration boundaries

  • Published FLOPS are peak theoretical specifications; they are not model serving benchmarks.
  • Microscaling formats and OCP FP8 differ from earlier CDNA FP8 implementations; kernel and checkpoint compatibility must be verified.

Published arithmetic throughput

Manufacturer peaks, not Aguasto measurements or tokens per second
SpecificationPrecisionPublished peakBoundary & source
OCP-FP8 peak computeOCP-FP85 PFLOPSPer GPU; manufacturer peak matrix specification without structured-sparsity figureManufacturer source ↗
MXFP4 peak computeMXFP410.1 PFLOPSPer GPU; manufacturer peak microscaling matrix specificationManufacturer source ↗

Deployment considerations

  • Match ROCm, kernel, framework and model versions against the current support matrix.
  • Weight-memory estimates exclude KV cache and serving overhead; reserve runtime headroom.
  • Check collectives, scale-up topology and model parallelism on the intended OEM system.

Software ecosystem

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

Continue the research

AMD ROCmvLLMSGLangModel 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 MI355X specifications ↗

    Launch date 2025-06-12; CDNA 4 architecture; 288 GB HBM3e; 8 TB/s memory bandwidth; 1400 W module board-power specification; Precision support; Cooling and ROCm ecosystem. Publication: date not stated in the reviewed source. Verified: 2026-10-12.

  2. AMD Instinct GPUs: MLPerf Inference v5.1 ↗

    MI300X, MI325X and MI355X launched products with manufacturer-reported deployments or benchmark submissions. Publication: 2025-09-09. Verified: 2026-10-12.

  3. AMD Instinct MI355X GPU brochure ↗

    Maximum TBP 1400 W. Publication: date not stated in the reviewed source. Verified: 2026-10-12.