AGUASTO

NVIDIA / Individual accelerator

NVIDIA H100 SXM

A documented commercial SXM accelerator; this record does not establish supplier stock or an Aguasto allocation.

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

Manufacturer specification reference

Architecture
HopperNVIDIA H100 GPU specifications (external source) · HGX H100, H200 and B200 reference architecture components (external source) · CUDA GPU Compute Capability (external source)
Category
Individual acceleratorNVIDIA H100 GPU specifications (external source) · HGX H100, H200 and B200 reference architecture components (external source) · CUDA GPU Compute Capability (external source)
Accelerator memory
80 GBPer H100 SXM accelerator, excluding host memory; not H100 NVL.NVIDIA H100 GPU specifications (external source)
Memory type
HBM3HGX H100, H200 and B200 reference architecture components (external source)
Memory bandwidth
3.35 TB/sPeak GPU memory bandwidth per SXM accelerator, not NVLink bandwidth.NVIDIA H100 GPU specifications (external source)
Power
700 WMaximum configurable GPU TDP, not server input or facility power.NVIDIA H100 GPU specifications (external source)
Precision formats
FP64, FP32, TF32, BF16, FP16, FP8, INT8NVIDIA H100 GPU specifications (external source)
Cooling / integration
SXM server integration; verify the OEM server cooling design. H100 NVL PCIe air-cooling specifications must not be applied to this SXM record.NVIDIA H100 GPU specifications (external source) · HGX H100, H200 and B200 reference architecture components (external source) · CUDA GPU Compute Capability (external source)

Configuration boundaries

  • The headline fields select the SXM variant. H100 NVL has a different capacity, power range and form factor.
  • Manufacturer peak Tensor Core throughput values may include structured sparsity; do not treat them as measured tokens per second.

Published arithmetic throughput

Manufacturer peaks, not Aguasto measurements or tokens per second
SpecificationPrecisionPublished peakBoundary & source
FP8 peak computeFP83,958 TFLOPSOne H100 SXM; manufacturer peak Tensor Core specification with sparsityManufacturer source ↗

Deployment considerations

  • Reserve accelerator memory for KV cache, activations and serving overhead in addition to model weights.
  • Check server topology and collective communication behavior before choosing tensor or pipeline parallelism.

Software ecosystem

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

Continue the research

vLLMSGLangTensorRT-LLMModel 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. NVIDIA H100 GPU specifications ↗

    Hopper architecture; SXM memory 80 GB; SXM bandwidth 3.35 TB/s; SXM maximum configurable TDP 700 W; precision support; SXM versus NVL variant distinction. Publication: date not stated in the reviewed source. Verified: 2026-10-12.

  2. HGX H100, H200 and B200 reference architecture components ↗

    H100 SXM HBM3 memory type. Publication: date not stated in the reviewed source. Verified: 2026-10-12.

  3. CUDA GPU Compute Capability ↗

    CUDA product identification for H100, H200, B200, B300, GB200 and GB300. Publication: date not stated in the reviewed source. Verified: 2026-10-12.