NVIDIA H100 specifications (web)
Citation
NVIDIA. “NVIDIA H100 GPU.” Product web page with specifications, date not stated. https://www.nvidia.com/en-us/data-center/h100/. Read 2026-11-09.
Bottom line
- The H100 is NVIDIA’s Hopper-generation data center GPU. The SXM version has 80GB of memory, 3.35TB/s of memory bandwidth and up to 700W of power.
- NVIDIA claims up to 4X faster training than the prior generation on GPT-3 (175B).
Background
NVIDIA makes GPUs for AI and high-performance computing (HPC). This is its product page for the H100, with performance claims and a specifications table for the H100 SXM and H100 NVL. The page gives no price.
Key claims
- “H100 features fourth-generation Tensor Cores and a Transformer Engine with FP8 precision that provides up to 4X faster training over the prior generation for GPT-3 (175B) models.” (Transformational AI Training)
- Specifications, SXM then NVL: ”| GPU Memory | 80GB | 94GB |”, ”| GPU Memory Bandwidth | 3.35TB/s | 3.9TB/s |” (Product Specifications).
- ”| Max Thermal Design Power (TDP) | Up to 700W (configurable) | 350-400W (configurable) |” (Product Specifications).
- ”| FP8 Tensor Core* | 3,958 teraFLOPS | 3,341 teraFLOPS |”, where ”* With sparsity” (Product Specifications).
- ”| Server Options | NVIDIA HGX H100 Partner and NVIDIA- Certified Systems™ with 4 or 8 GPUs NVIDIA DGX H100 with 8 GPUs |” (Product Specifications).
- “NVIDIA H100 NVL comes with a five-year NVIDIA AI Enterprise subscription.” (Enterprise-Ready)
Evidence and method
A vendor’s product page. Performance claims are NVIDIA’s projections. The page says “Projected performance subject to change.” The table gives rated figures.
Limits
- No price or cost of ownership.
- Epoch AI GPU clusters (web) measures clusters in “H100 equivalents”. This page does not define that unit.