NVIDIA A100 (SXM 80GB)

Accelerators

Processing & platform

Architecture: Ampere^*

Memory & storage

Video memory: 80 GB^*
Video memory type: HBM2e^*
Rated memory bandwidth: 2,039 GB/s^*

System configuration

Detailed specifications

Specifications 1 · 19 fields

GPU architecture — NVIDIA A100 80GB SXM — 80 GB HBM2e: Ampere^*
Theoretical peak: BF16 tensor dense per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 312^* TFLOPS
Theoretical peak: FP16 tensor dense per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 312^* TFLOPS
Theoretical peak: FP32 non-tensor dense per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 19.5^* TFLOPS
Native hardware support: FP4 tensor — NVIDIA A100 80GB SXM — 80 GB HBM2e: No^*
Theoretical peak: FP64 non-tensor dense per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 9.7^* TFLOPS
Theoretical peak: FP64 tensor dense per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 19.5^* TFLOPS
Native hardware support: FP6 tensor — NVIDIA A100 80GB SXM — 80 GB HBM2e: No^*
Native hardware support: FP8 tensor — NVIDIA A100 80GB SXM — 80 GB HBM2e: No^*
Theoretical peak: INT8 tensor dense per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 624^* TOPS
Theoretical peak: TF32 tensor dense per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 156^* TFLOPS
Physical configuration — NVIDIA A100 80GB SXM — 80 GB HBM2e: SXM4^*
GPU interconnect bandwidth (NVLink 3) — NVIDIA A100 80GB SXM — 80 GB HBM2e: 600^* GB/s
PCIe interface declaration — NVIDIA A100 80GB SXM — 80 GB HBM2e: PCIe Gen4^*
Maximum board power — NVIDIA A100 80GB SXM — 80 GB HBM2e: 400^* W
Peak memory bandwidth per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 2,039^* GB/s
Memory capacity per GPU — NVIDIA A100 80GB SXM — 80 GB HBM2e: 80^* GB
Memory technology — NVIDIA A100 80GB SXM — 80 GB HBM2e: HBM2e^*
GPU model — NVIDIA A100 80GB SXM — 80 GB HBM2e: NVIDIA A100 80GB SXM^*

Specifications 2 · 16 fields

device name: NVIDIA A100 (SXM 80GB)^*
typical power w: 400^* W
on chip sram mb: 40^*
hbm memory gb: 80^* GB
memory bandwidth tb s: 2^*
headline metric: 624 TOPS INT8 dense (1248 sparse); 80 GB HBM2e; the pre-Hopper datacenter workhorse^*
source type: Vendor^*
vendor: NVIDIA^*
tops per watt calc: 1.56^*
compute paradigm: Electronic^*
device type: GPU^*
release year: 2020^*
maturity: Commercial^*
precision: INT8^*
throughput unit: TOPS^*
peak throughput: 624^*

* Provisional data