Product
NVIDIA B200 180GB SXM
Part number NVIDIA B200 180GB SXM
- Price
- USD $100
- Stock
- 1
Blackwell-architecture SXM accelerator with 180 GB of HBM3E memory, 7.7 TB/s of bandwidth, fifth-generation NVLink, and a configurable TDP up to 1,000 W, used in HGX B200 eight-GPU systems.
- 180 GB HBM3E memory at 7.7 TB/s
- SXM module for NVIDIA HGX B200 8-GPU systems
- Fifth-generation NVLink at 1.8 TB/s
- Configurable TDP up to 1,000 W
- FP4 Tensor performance up to 18 PFLOPS with sparsity, per NVIDIA’s Blackwell datasheet
Specifications
| GPU Model | NVIDIA B200 180GB SXM |
|---|---|
| Architecture | NVIDIA Blackwell |
| Form Factor | SXM |
| GPU Count | 1 |
| GPU Memory | 180 GB HBM3E |
| Memory Bandwidth | 7.7 TB/s |
| Interconnect | Fifth-generation NVLink 1.8 TB/s; PCIe Gen5 128 GB/s |
| Max TDP | Configurable up to 1,000 W |
| Cooling | Host-cooled SXM module |
| FP4 Tensor Core | 18 PFLOPS (with sparsity) |
| FP8 / FP6 Tensor Core | 9 PFLOPS (with sparsity) |
| INT8 Tensor Core | 9 POPS (with sparsity) |
| FP16 / BFLOAT16 Tensor Core | 4.5 PFLOPS (with sparsity) |
| TF32 Tensor Core | 2.2 PFLOPS (with sparsity) |
| FP32 | 75 TFLOPS |
| FP64 / FP64 Tensor Core | 37 TFLOPS |
| Multi-Instance GPU | Up to 7 MIG instances |
| Typical Host | NVIDIA HGX B200 partner and NVIDIA-Certified Systems with 8 GPUs |
| HGX Board Memory | 1.4 TB aggregate across 8 GPUs |
| HGX NVLink Bandwidth | 14.4 TB/s aggregate; 1.8 TB/s GPU-to-GPU |
Description
The NVIDIA B200 180GB SXM is the Blackwell-architecture GPU used on NVIDIA HGX B200 baseboards. Official NVIDIA specifications for the individual GPU list 180 GB of HBM3E memory with 7.7 TB/s of bandwidth, a configurable thermal design power of up to 1,000 W, and fifth-generation NVLink at 1.8 TB/s, with PCIe Gen5 at 128 GB/s to the host.
NVIDIA’s Blackwell datasheet publishes per-GPU Tensor rates that include 18 PFLOPS of FP4 (with sparsity), 9 PFLOPS of FP8/FP6, 4.5 PFLOPS of FP16/BF16, and 37 TFLOPS of FP64 / FP64 Tensor performance. Up to seven Multi-Instance GPU partitions are listed. An eight-GPU HGX B200 board aggregates 1.4 TB of GPU memory and 14.4 TB/s of NVLink bandwidth, with 1.8 TB/s of GPU-to-GPU bandwidth through NVLink Switch.
This is a training and large-scale inference accelerator, not a graphics or workstation card. The higher memory capacity and fifth-generation NVLink fabric are aimed at trillion-parameter-class models and tightly coupled eight-GPU nodes. Cooling is provided by the host system; OEM servers may be air- or liquid-cooled. Per-GPU TDP should be planned at the published 1,000 W ceiling unless a specific OEM SKU states a lower configured point.
Blackwell Ultra (HGX B300) is the follow-on eight-GPU platform, with a published aggregate memory of 2.1 TB. Request a quote if you need B200 modules or an HGX B200-class host.
Related products
NVIDIA Blackwell Ultra B300 SXM
P/N NVIDIA HGX B300
Price USD $100 Stock 1
Blackwell Ultra SXM GPU used in NVIDIA HGX B300 eight-GPU systems. NVIDIA publishes 2.1 TB of aggregate HBM memory, fifth-generation NVLink at 1.8 TB/s GPU-to-GPU, and up to 144 PFLOPS of sparse FP4 Tensor performance across the eight-GPU board.
- GPU Model
- NVIDIA Blackwell Ultra B300 SXM
- Form Factor
- SXM
- GPU Count
- 8 per HGX B300 board
- GPU Memory
- 2.1 TB aggregate (HGX B300, 8 GPUs)
GIGABYTE G894-SD3-AAX7
P/N G894-SD3-AAX7
Price USD $100 Stock 1
8U dual-socket Intel Xeon 6 server built around an NVIDIA HGX B300 eight-GPU SXM module, with 1.8 TB/s NVLink bandwidth, onboard ConnectX-8 networking, and twelve 3000 W Titanium power supplies.
- Model
- G894-SD3-AAX7
- GPU Model
- NVIDIA HGX B300
- GPU Count
- 8 SXM
- Form Factor
- 8U rackmount
Ready to specify a configuration?
Share the use case, the target build, quantity, destination country, and timeline. We will take it from the inquiry.