SERVCHIP
Interactive view
ARCHITECTURE
Hopper
SERIES
H100 NVL
Memory
188GB HBM3 (2x 94GB)
Bandwidth
7.9 TB/s (2x 3.95 TB/s)
TDP
600W
In StockH100 NVL·Best for: Dual-GPU LLM Inference
NVIDIA H100 NVL Tensor Core GPU
Hopper
|Data Center GPUsOverview
The NVIDIA H100 NVL Tensor Core GPU is a dual-GPU data center accelerator featuring two Hopper-architecture H100 GPUs on a single board, each with 94GB of HBM3 memory for a total of 188GB. Connected via NVLink, the dual-GPU design delivers 3,958 FP8 TFLOPS of combined AI performance with 7.9 TB/s aggregate memory bandwidth. Purpose-built for running and fine-tuning large language models, the H100 NVL provides up to 4.5x the inference performance of a single H100 at a fraction of the per-GPU power cost, making it ideal for high-throughput LLM serving in enterprise and cloud environments.
Key Features
- 188GB HBM3 memory (2x 94GB) with 7.9 TB/s aggregate bandwidth
- Dual Hopper GPUs with NVLink inter-GPU connection
- 3,958 FP8 TFLOPS combined peak AI performance
- Multi-Instance GPU (MIG) per GPU for multi-tenant inference
- Optimized for large language model inference and fine-tuning
Ideal Use Cases
ai inferenceai traininghpc
Full Specifications
Complete technical specifications for the H100 NVL.
Memory
- Memory
- 188GB HBM3 (2x 94GB)
- Memory Bandwidth
- 7.9 TB/s (2x 3.95 TB/s)
Compute
- Tensor Cores
- 1056 (4th-gen, 2x 528)
- CUDA Cores
- 36864 (2x 18432)
- FP8 TFLOPS
- 3958 (2x 1979)
- FP16 TFLOPS
- 1978 (2x 989)
- TF32 TFLOPS
- 989 (2x 494.5)
- FP64 TFLOPS
- 120 (2x 60)
Interconnect
- Interconnect
- NVLink 4.0, 900GB/s per GPU
Physical
- TDP
- 600W
- Form Factor
- SXM5 Dual Module
- Cooling
- Liquid Cooled
Manufacturing
- Launch Date
- Q4 2023
- Manufacturing Process
- 4nm
