NVIDIA H100 PCIe 80GB — Full Datasheet Specifications
All figures from NVIDIA H100 PCIe datasheet (DS-10167-001). Tensor Core TFLOPS with sparsity assumes 2:4 structured sparse format. Peak rates based on GPU Boost Clock.
Source: NVIDIA H100 PCIe Datasheet DS-10167-001. All trademarks are the property of NVIDIA Corporation.
H100 vs H200 vs A100 PCIe
Source: NVIDIA datasheets DS-10167-001 (H100 PCIe), DS-10581-001_v01 (H200 NVL PCIe), DS-10010-001 (A100 PCIe).
H100 PCIe Key Workloads
The H100 PCIe is the first data-centre GPU to feature the Transformer Engine — purpose-built for the generation of AI models that underpins GPT, BERT, diffusion, and multimodal architectures.
Large Language Model Training
FP8 Transformer Engine with automatic mixed precision delivers up to 30x higher LLM training throughput versus A100 on BERT, GPT-style, and diffusion models. 80GB VRAM enables training models up to 30B parameters on a single card.
AI Inference Serving
The H100 PCIe is the reference GPU for production LLM inference with frameworks such as TensorRT-LLM, vLLM, and NVIDIA Triton. FP8 precision doubles throughput vs FP16 for inference at equivalent perplexity.
Multi-Tenant GPU Partitioning
MIG partitions one H100 PCIe into up to 7 isolated 10GB GPU instances — enabling guaranteed QoS for multiple concurrent workloads, tenants, or microservices on a single physical card.
Scientific & HPC Simulation
Third-generation NVLink (600 GB/s bridge between paired H100 NVL) and PCIe 5.0 x16 reduce inter-GPU and CPU-GPU data movement bottlenecks in distributed simulation, climate modelling, and molecular dynamics.
Generative AI (Image & Video)
Stable Diffusion XL, Sora-class video generation, and image-to-3D models benefit from the H100 Transformer Engine and 2 TB/s bandwidth during the denoising attention computation passes.
RAG & Vector Search at Scale
NVIDIA cuVS and FAISS-GPU leverage H100 HBM bandwidth for billion-scale approximate nearest-neighbour (ANN) vector search — critical for production retrieval-augmented generation (RAG) pipelines.
Compare & Related Products
Need NVIDIA H100 PCIe for AI or HPC?
We advise on PCIe vs SXM, confirm server PSU and slot compatibility (350W card requires 80+ Platinum supply and PCIe 5.0 slot), and provide UK lead time, availability, and quote.