The proven Hopper-generation workhorse. The H100 SXM5 pairs 80GB of HBM3 with fourth-generation Tensor Cores and NVLink, and remains the most widely deployed accelerator for large-model training and high-throughput inference — with a deep secondary market that keeps lead times and pricing competitive.
Key specifications for the NVIDIA H100 Tensor Core GPU. Hard specs mirror the live catalog; confirm exact revision and configuration at quote.
| Specification | H100 |
|---|---|
| Architecture | Hopper (GH100, TSMC 4N) |
| GPUs per unit | 1 (8 per HGX/DGX server) |
| Memory | 80 GB HBM3 |
| Memory bandwidth | 3.4 TB/s |
| FP8 tensor | 3,958 TFLOPS (with sparsity) |
| FP16 / BF16 tensor | 1,979 TFLOPS (with sparsity) |
| NVLink | 900 GB/s (4th-gen NVLink) |
| Board power (TDP) | 700 W |
| Form factor | SXM5 |
| Cooling | Air or direct-to-chip liquid |
| Typical lead time | 10–14 wk |
| Indicative price | From $26,500 / unit |
The H100 is the default choice when you want a battle-tested accelerator with a broad software stack and predictable supply. Its 80GB of HBM3 comfortably holds most models up to the tens-of-billions of parameters range with tensor and pipeline parallelism, and FP8 Transformer Engine support roughly doubles throughput over A100-class hardware.
Scales across NVLink and 400G/800G fabrics into large training clusters. The proven choice for training runs where the software stack maturity matters.
80GB fits most production models; FP8 delivers strong tokens-per-second per dollar for served endpoints and batch inference.
Widely available new, used and refurbished — the pragmatic pick for teams standing up capacity on a budget and timeline.
Density decides the facility. Size the electrical capacity, heat rejection and cooling method before you rack a single H100 — our free tools turn the specs above into facility numbers.
Add it to your quote list — free, no commitment — and we confirm a firm, itemized price, availability and lead time after review, usually within a business day.
The H100 SXM5 has 80GB of HBM3 with about 3.4 TB/s of memory bandwidth. A PCIe H100 NVL pairs two boards for 188GB combined; the newer H200 raises a single module to 141GB of faster HBM3e.
The H100 SXM5 has a rated board power (TDP) of 700W. A full 8-GPU HGX/DGX H100 server draws roughly 10kW once CPUs, memory, NICs, fans and PSU losses are added — size cooling and power to that figure, not to raw GPU TDP.
They share the same Hopper GH100 silicon and identical compute (FP8/FP16 TFLOPS). The H200 upgrades memory to 141GB HBM3e at 4.8 TB/s, versus 80GB HBM3 at 3.4 TB/s on the H100 — a large win for memory-bound inference and long-context workloads, at the same 700W.
Indicative pricing starts around From $26,500 / unit, with a typical lead time of 10–14 wk — confirmed per line at quote. A deep secondary market means used and refurbished units are often available faster and cheaper.
Our solutions engineers size compute, power and cooling together and confirm real lead times — no payment, no commitment, quotes back in about one business day.
Our engineers scope power, cooling and compute together. No payment, and quotes come back in about a business day.