AGUASTO

Research / Comparison

Compare the same category.

Manufacturer references with consistent units and explicit variant boundaries. A component is not a rack; arithmetic peaks are not model benchmarks.

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

Compare up to three references within one product category. Manufacturer peak figures can use different precisions and sparsity assumptions; they are not measured tokens per second.

Single-accelerator references only. Capacity, bandwidth and power retain each source's product boundary. These are not Aguasto inventory or benchmark results.
Reference fieldNVIDIA H100 SXMNVIDIA H200 SXM
Architecture / manufacturerHopperNVIDIAHopperNVIDIA
Product statuscommercial-productA documented commercial SXM accelerator; this record does not establish supplier stock or an Aguasto allocation.availableThe manufacturer product page explicitly says available. No supplier stock or Aguasto deployment is verified.
Memory capacity80 GBHBM3Per H100 SXM accelerator, excluding host memory; not H100 NVL.NVIDIA H100 GPU specifications (external source)141 GBHBM3ePer H200 SXM GPU; host memory excluded.NVIDIA H200 GPU specifications (external source)
Memory bandwidth3.35 TB/sPeak GPU memory bandwidth per SXM accelerator, not NVLink bandwidth.NVIDIA H100 GPU specifications (external source)4.8 TB/sPeak GPU memory bandwidth per accelerator.NVIDIA H200 GPU specifications (external source)
Manufacturer peak compute

FP8 peak compute: 3,958 TFLOPS (FP8)

One H100 SXM; manufacturer peak Tensor Core specification with sparsityNVIDIA H100 GPU specifications (external source)

FP8 peak compute: 3,958 TFLOPS (FP8)

One H200 SXM; manufacturer peak Tensor Core specification with sparsityNVIDIA H200 GPU specifications (external source)
Power boundary700 WMaximum configurable GPU TDP, not server input or facility power.NVIDIA H100 GPU specifications (external source)700 WMaximum configurable SXM GPU TDP; not H200 NVL or complete server draw.NVIDIA H200 GPU specifications (external source)
Documented precisionsFP64, FP32, TF32, BF16, FP16, FP8, INT8FP64, FP32, TF32, BF16, FP16, FP8, INT8
CoolingSXM server integration; verify the OEM server cooling design. H100 NVL PCIe air-cooling specifications must not be applied to this SXM record.SXM server thermal implementation depends on OEM system. The separately documented H200 NVL is a dual-slot air-cooled PCIe product.
Software referencesCUDAExact model, driver and framework compatibility requires separate validation.CUDAExact model, driver and framework compatibility requires separate validation.
Specification caveats

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.

NVIDIA retains a preliminary/specifications-may-change footnote on this page; preserve that qualification.

Headline power selects SXM. NVL has a different published power limit.

Engineering 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.

Larger GPU memory can accommodate more weights or cache, but capacity alone does not establish latency or concurrency.

Confirm the serving framework, GPU driver and selected model format together.

Primary sourcesNVIDIA H100 GPU specifications (external source) · HGX H100, H200 and B200 reference architecture components (external source) · CUDA GPU Compute Capability (external source)NVIDIA H200 GPU specifications (external source) · CUDA GPU Compute Capability (external source)