A100 SXM 80GB specs
NVIDIA · Legacy · first available Nov 2020 (estimated). Dense compute, memory, bandwidth and power from the vendor's documents; each precision stays separate.
All accelerators · Compare with H100 SXM
Specifications
| Dense BF16 · FP32 accumulate | 312 TFLOPS |
|---|---|
| Dense FP8 · FP32 accumulate | Not published |
| Dense FP4 · FP32 accumulate | Not published |
| Memory | 80 GB · HBM |
| Memory bandwidth | 2.04 TB/s |
| Maximum chip power | 400 W |
| Interconnect | NVLink 3 · 600 GB/s |
| Process node | TSMC N7 |
| First availability | Nov 2020est — estimated / secondary-source figure |
| Availability | Legacy |
| Est. purchase price | ~$10,000est — estimated / secondary-source figureReported price in February 2023, not at launch |
Workhorse of the GPT-3→GPT-4 era.
Dense tensor rates with FP32 accumulation, never sparse or FP16-accumulate headline figures. Memory is in decimal GB. The est marker flags a value its vendor-document review did not confirm.
Rent it
$1.35/GPU-hrCheapest on-demand · Hyperstack · observed
7 providers: AWS, Google Cloud, Hyperstack, Jarvislabs, Lambda, Oracle OCI, Verda. Quotes checked through .
A100 80GB rental prices from every provider →Value ratios
| $ per dense BF16 PFLOP-hour | $4.33 |
|---|---|
| Dense BF16 TFLOPS per watt | 0.78 |
| Bandwidth bytes per 1,000 BF16 FLOP | 6.54 |
| $ per dense FP8 PFLOP-hour | No dense FP8 rate |
| Dense FP8 TFLOPS per watt | No dense FP8 rate |
| Bandwidth bytes per 1,000 FP8 FLOP | No dense FP8 rate |
| $ per GB of memory per hour | $0.017 |
Prices divide the cheapest on-demand list price; theoretical ratios do not measure application speed.
Sources
Specifications reviewed . 80GB SXM column; process TSMC N7.
Purchase price: CNBC ↗ (2023-02-23): “a roughly $10,000 chip”
Rental price: Hyperstack published pricing ↗ (NVIDIA A100 80GB; maximum per GPU: 28 pCPUs and 120 GB RAM; actual allocation not specified; published per GPU-hour).