Product
NVIDIA H100 80GB SXM5
Part number NVIDIA H100 80GB SXM5
- Price
- USD $100
- Stock
- 1
Hopper-architecture SXM accelerator with 80 GB of HBM3 memory, fourth-generation NVLink, and a configurable TDP up to 700 W, built for dense HGX training and inference nodes.
- 80 GB HBM3 memory at 3.35 TB/s
- SXM5 form factor for NVIDIA HGX H100 systems
- NVLink at 900 GB/s plus PCIe Gen5
- Configurable TDP up to 700 W
- Up to seven MIG instances at 10 GB each
Specifications
| GPU Model | NVIDIA H100 80GB SXM5 |
|---|---|
| Architecture | NVIDIA Hopper |
| Form Factor | SXM5 |
| GPU Count | 1 |
| GPU Memory | 80 GB HBM3 |
| Memory Bandwidth | 3.35 TB/s |
| Interconnect | NVLink 900 GB/s; PCIe Gen5 128 GB/s |
| Max TDP | Up to 700 W (configurable) |
| Cooling | Host-cooled SXM module |
| FP64 | 34 TFLOPS |
| FP64 Tensor Core | 67 TFLOPS |
| FP32 | 67 TFLOPS |
| TF32 Tensor Core | 989 TFLOPS (with sparsity) |
| FP16 / BFLOAT16 Tensor Core | 1,979 TFLOPS (with sparsity) |
| FP8 Tensor Core | 3,958 TFLOPS (with sparsity) |
| INT8 Tensor Core | 3,958 TOPS (with sparsity) |
| Multi-Instance GPU | Up to 7 MIGs at 10 GB each |
| Media Engines | 7 NVDEC, 7 JPEG |
| Typical Host | NVIDIA HGX H100 (4- or 8-GPU) and NVIDIA DGX H100 |
Description
The NVIDIA H100 80GB SXM5 is a Hopper-architecture Tensor Core GPU supplied as an SXM5 module for dense AI servers. It is the high-bandwidth, NVLink-connected member of the H100 family, intended for multi-GPU training and large-model inference rather than as a standalone add-in card.
Each module carries 80 GB of HBM3 memory with 3.35 TB/s of memory bandwidth. Fourth-generation NVIDIA NVLink provides 900 GB/s of GPU-to-GPU bandwidth, with PCIe Gen5 at 128 GB/s to the host. In an eight-GPU HGX H100 board, NVSwitch links the GPUs so any pair can use the full NVLink rate. The same Hopper Transformer Engine and FP8 Tensor Cores used across the H100 family are available here at the higher SXM power envelope.
NVIDIA publishes a configurable thermal design power of up to 700 W for the SXM variant. The module itself has no card-level fan; cooling is provided by the host system, which may be air or liquid depending on the server design. Multi-Instance GPU can partition a device into as many as seven hardware-isolated instances at 10 GB each, which is useful when a node is shared across inference jobs.
Typical use is large-scale model training, fine-tuning, and high-throughput inference on HGX H100 or DGX H100 class systems. Workloads that need fewer GPUs, lower power, or a standard PCIe slot are better served by the H100 80GB PCIe card. Request a quote if you need this module on its own or as part of an eight-GPU host.
Related products
GIGABYTE G593-SD0-AAX1
P/N G593-SD0-AAX1
Price USD $100 Stock 1
5U dual-socket Intel Xeon server built around an NVIDIA HGX H100 eight-GPU SXM5 module, with 900 GB/s NVLink/NVSwitch bandwidth, twelve Gen5 x16 low-profile slots, and 4+2 3000 W Titanium power supplies.
- Model
- G593-SD0-AAX1
- GPU Model
- NVIDIA HGX H100
- GPU Count
- 8 SXM5
- Form Factor
- 5U rackmount
NVIDIA H100 80GB PCIe
P/N NVIDIA H100 80GB PCIe
Price USD $100 Stock 1
Hopper-architecture PCIe Gen5 accelerator with 80 GB of HBM2e memory and a 350 W envelope, suited to mainstream servers and one- or two-GPU inference nodes.
- GPU Model
- NVIDIA H100 80GB PCIe
- Form Factor
- PCIe Gen5 dual-slot
- GPU Count
- 1
- GPU Memory
- 80 GB HBM2e
Ready to specify a configuration?
Share the use case, the target build, quantity, destination country, and timeline. We will take it from the inquiry.