CARDINAL
EXACT VARIANT REFERENCE

A100 SXM4 80GB

Four core configurations differ by memory capacity and physical integration

SPECIFICATIONS

Memory per accelerator80 GB HBM2e
Memory bandwidth2039 GB/s
Form factorSXM4 module
Maximum accelerator power400 W

Manufacturer specifications describe the product, not the condition of an offered unit. Checked 2026-09-08

COMPUTE SPECIFICATIONS

Manufacturer peak arithmetic per accelerator. These values do not predict application throughput or latency.

PrecisionExecutionPeak rateSparsityReference
FP64Vector9.7 TFLOPSDenseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

FP64Tensor19.5 TFLOPSDenseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

FP32Vector19.5 TFLOPSDenseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

TF32Tensor156 TFLOPSDenseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

TF32Tensor312 TFLOPSStructured sparseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

BF16Tensor312 TFLOPSDenseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

BF16Tensor624 TFLOPSStructured sparseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

FP16Tensor312 TFLOPSDenseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

FP16Tensor624 TFLOPSStructured sparseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

INT8Tensor624 TOPSDenseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

INT8Tensor1248 TOPSStructured sparseSource ↗
Conditions

Manufacturer peak per GPU; not sustained workload performance

  • Peak arithmetic rates are not token throughput, latency or a guarantee of application performance
  • Structured-sparse rates require eligible sparsity and supported kernels; do not compare them directly with dense rates
  • The June 2021 manufacturer table explicitly spans all four memory/form-factor variants; equal listed compute rates do not imply equal memory bandwidth or power

PLATFORM AND SOFTWARE

  • 400 W is the standard configuration
  • NVIDIA lists up to 500 W for a separate custom thermal solution SKU; do not assume that rating for a standard offer
  • Confirm driver and framework support for the intended deployment image
  • Benchmark the exact precision and batch settings instead of comparing theoretical tensor figures

Manufacturer maximum accelerator rating; actual draw and supported limits depend on system

INTERCONNECT

NVLink

600 GB/s. Manufacturer aggregate bidirectional bandwidth; topology depends on platform

PCIe Gen4 x16

64 GB/s. Theoretical aggregate bidirectional interface bandwidth

MANUFACTURER SOURCES

NVIDIA A100 specifications ↗NVIDIA · Checked 2026-09-08NVIDIA A100 datasheet ↗NVIDIA · Checked 2026-09-08NVIDIA A100 four-variant datasheet, June 2021 ↗NVIDIA · Checked 2026-09-08

HOW CARDINAL HELPS

Share your quantity, destination and server model, if known. We help turn those requirements into a clearly specified request.

  • Help confirm the exact configuration and platform requirements
  • Request supplier records for condition, unit identifiers and available testing evidence
  • Clarify included accessories, delivery and proposed warranty terms before you decide

Any additional testing or guarantees must be agreed in the written offer.