H100 SXM specs
NVIDIA · Shipping now · first available Oct 2022. Dense compute, memory, bandwidth and power from the vendor's documents; each precision stays separate.
All accelerators · Compare with GH200 (96 GB)
Specifications
| Dense BF16 · FP32 accumulate | 989 TFLOPS |
|---|---|
| Dense FP8 · FP32 accumulate | 1,979 TFLOPS |
| Dense FP4 · FP32 accumulate | Not published |
| Memory | 80 GB · HBM |
| Memory bandwidth | 3.35 TB/s |
| Maximum chip power | 700 W |
| Interconnect | NVLink 4 · 900 GB/s |
| Process node | TSMC 4N |
| First availability | Oct 2022 |
| Availability | Shipping |
| Est. purchase price | ~$27,500est — estimated / secondary-source figureMidpoint of an analyst range |
Defined the 2023–25 training buildout; the market’s reference chip.
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
$3.20/GPU-hrCheapest on-demand · Hyperstack · observed
10 providers: AWS, Azure, Google Cloud, Hyperstack, Jarvislabs, Lambda, Nebius, Oracle OCI, Together AI, Verda. Quotes checked through .
H100 rental prices from every provider →Value ratios
| $ per dense BF16 PFLOP-hour | $3.24 |
|---|---|
| Dense BF16 TFLOPS per watt | 1.41 |
| Bandwidth bytes per 1,000 BF16 FLOP | 3.39 |
| $ per dense FP8 PFLOP-hour | $1.62 |
| Dense FP8 TFLOPS per watt | 2.83 |
| Bandwidth bytes per 1,000 FP8 FLOP | 1.69 |
| $ per GB of memory per hour | $0.040 |
Prices divide the cheapest on-demand list price; theoretical ratios do not measure application speed.
Sources
Specifications reviewed . SXM column; sparse tensor figures divided by two, with existing whole-TFLOPS rounding retained.
First availability: NVIDIA Q3 FY2023 results ↗ (2022-11-16): “Began shipping the NVIDIA H100 Tensor Core GPU … with first systems available now.”
Purchase price: CNBC, citing Raymond James ↗ (2024-01-18): “Raymond James estimate Nvidia is selling the H100 for $25,000 to $30,000”
Rental price: Hyperstack published pricing ↗ (NVIDIA H100 SXM 80GB; maximum per GPU: 24 pCPUs and 240 GB RAM; actual allocation not specified; published per GPU-hour).