GeekRadar / The comparison desk

COMPARE.

GPU

NVIDIA

NVIDIA A100 SXM4 40 GB

Graphics processor

Video memory: 40 GB^*
Memory bus: 5,120 bit^*
Rated memory bandwidth: 1,560 GB/s^*

No matched benchmark results are available for this entry yet.

Processing & platform

Architecture: Ampere^*
Process node: 7 nm^*
GPU cores: 6,912^*
Reported GPU frequency: 1,095 MHz^*
Graphics interface: PCIe 4.0 x16^*
Reported chip die area: 826 mm²^*
Chip/package TDP: 400 W^*
Transistors: 54,200 million^*
GPU L2 cache: 40 MB^*
Graphics processor: GA100^*
Texture mapping units: 432^*
Raster output units: 160^*
Streaming multiprocessors: 108^*
Tensor cores: 432^*
Chip foundry: TSMC^*
Card or module format: SXM Module^*
Declared GPU TDP: 400 W^*
Suggested system PSU: 800 W^*
External power connectors: None^*
OpenCL: 3.0^*
CUDA compute capability: 8.0^*

Memory & storage

Memory options: 40^*
Video memory: 40 GB^*
Video memory type: HBM2e^*
Memory bus: 5,120 bit^*
Rated memory bandwidth: 1,560 GB/s^*

Power specifications

Thermal design power: 400 W^*

System configuration

Detailed specifications

Specifications 1 · 20 fields

GPU architecture — NVIDIA A100 40GB SXM — 40 GB HBM2: Ampere^*
Theoretical peak: BF16 tensor dense per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 312^* TFLOPS
Theoretical peak: FP16 tensor dense per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 312^* TFLOPS
Theoretical peak: FP32 non-tensor dense per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 19.5^* TFLOPS
Native hardware support: FP4 tensor — NVIDIA A100 40GB SXM — 40 GB HBM2: No^*
Theoretical peak: FP64 non-tensor dense per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 9.7^* TFLOPS
Theoretical peak: FP64 tensor dense per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 19.5^* TFLOPS
Native hardware support: FP6 tensor — NVIDIA A100 40GB SXM — 40 GB HBM2: No^*
Native hardware support: FP8 tensor — NVIDIA A100 40GB SXM — 40 GB HBM2: No^*
Theoretical peak: INT8 tensor dense per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 624^* TOPS
Theoretical peak: TF32 tensor dense per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 156^* TFLOPS
Physical configuration — NVIDIA A100 40GB SXM — 40 GB HBM2: SXM4^*
GPU interconnect bandwidth (NVLink 3) — NVIDIA A100 40GB SXM — 40 GB HBM2: 600^* GB/s
PCIe interface declaration — NVIDIA A100 40GB SXM — 40 GB HBM2: PCIe Gen4^*
Maximum board power — NVIDIA A100 40GB SXM — 40 GB HBM2: 400^* W
Peak memory bandwidth per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 1,555^* GB/s
Memory interface width per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 5,120^* bit
Memory capacity per GPU — NVIDIA A100 40GB SXM — 40 GB HBM2: 40^* GB
Memory technology — NVIDIA A100 40GB SXM — 40 GB HBM2: HBM2^*
GPU model — NVIDIA A100 40GB SXM — 40 GB HBM2: NVIDIA A100 40GB SXM^*

Specifications 2 · 13 fields

Foundry: TSMC^*
Vendor: NVIDIA^*
FP16 GFLOPS: 77,970^* GFLOPS
FP32 GFLOPS: 19,490^* GFLOPS
FP64 GFLOPS: 9,746^* GFLOPS
Product: NVIDIA A100 SXM4 40 GB^*
Type: GPU^*
Release Date: 2020-05-14^*
Process Size (nm): 7^* nm
TDP (W): 400^* W
Die Size (mm^2): 826^* mm²
Transistors (million): 54,200^*
Freq (MHz): 1,095^* MHz

Specifications 3 · 31 fields

proc transistors: 54,200 million^*
proc die size: 826^* mm²
cores: 6,912^*
tensor cores: 432^*
sm count: 108^*
base clock: 1,095^*
boost clock: 1,410^*
memory clock: 1,215^*
ram: 40^*
name: NVIDIA A100 SXM4 40 GB^*
ram type: HBM2e^*
bus width: 5,120^*
bus interface: PCIe 4.0 x16^*
ram bandwidth: 1.56 TB/s^*
fp16 performance: 77.97^*
fp32 performance: 19.49^*
fp64 performance: 9.746^*
bf16 performance: 311.84^*
tf32 performance: 155.92^*
tdp: 400^* W
suggested psu: 800^* W
slot width: SXM Module^*
api cuda: 8^*
is retail board: No^*
gpi value: 22.11^*
vendor: nvd^*
manufacturer: NVIDIA^*
arch name: Ampere^*
card release date: 2020-05-14^*
proc foundry: TSMC^*
proc process size: 7^* nm

Specifications 4 · 37 fields

CUDA compute capability

8.0^*

OpenCL

3.0^*

Architecture

Ampere^*

Reported theoretical BF16 peak

311.84^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Graphics interface

PCIe 4.0 x16^*

Memory bus

5,120^* bit

Reported shader processors

6,912^*

Reported theoretical FP16 peak

77.97^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Reported theoretical FP32 peak

19.49^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Reported theoretical FP64 peak

9.746^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Declared L1 cache configuration

192 KB (per SM)^*

Declared GPU L2 cache

40 MB^*

Reported display outputs

No outputs^*

Theoretical pixel fill rate

225.6 GPixel/s^*

External power connectors

None^*

Reported chip die area

826 mm²^*

Chip foundry

TSMC^*

Process node declaration

7 nm^*

Transistors

54,200 million^*

Graphics processor

GA100^*

Reported video memory

40^* GB

Rated memory bandwidth

1.56 TB/s^*

Video memory type

HBM2e^*

Raster output units

160^*

Reported shading units

6,912^*

Card or module format

SXM Module^*

Streaming multiprocessors

108^*

Suggested system PSU

800 W^*

2:4 tensor sparsity

Yes^*

Declared GPU TDP

400 W^*

Tensor cores

432^*

Theoretical texture fill rate

609.1 GTexel/s^*

Reported theoretical TF32 peak

155.92^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Texture mapping units

432^*

GPU cores

6,912^*

Video memory

40^* GB

Video memory type

HBM2e^*

* Provisional data