NVIDIA DGX B200
Eight Blackwell GPUs, for training, fine-tuning and inference.
- GPUs
- 8x Blackwell
- GPU memory
- 1,440 GB
- Networking
- 400 Gb/s
Price confirmed on your quote
Configure
DGX systems
NVIDIA DGX B300
Eight Blackwell Ultra GPUs, for training and real-time inference.
- GPUs
- 8x Blackwell Ultra
- GPU memory
- 2.1 TB
- Networking
- 800 Gb/s
- Size
- 10 RU
Enquiry
Ask about the NVIDIA DGX B300
Six questions. Our team follows up, or you can pick a time.
Would rather just email? support@webparam.org
Configure this system
Spec it the way you would run it. Pricing comes back on your quote.
- Total GPU memory
- 2.1 TB
Pricing is confirmed on your quote.
Not sure? Talk to a specialist
Overview
The DGX B300 is NVIDIA's Blackwell Ultra system: eight GPUs, 2.1 TB of GPU memory and networking at up to 800 Gb/s per port. Compared with the DGX B200, NVIDIA quotes 1.5x the dense FP4 performance and 2x the attention performance.
It comes in two power configurations, AC power supplies or a DC busbar, so it can go into a traditional enterprise rack or an NVIDIA MGX-compatible one.
- Made by
- NVIDIA
- Category
- DGX systems
What it’s built for
Training and fine-tuning large language models
Real-time inference for reasoning models
Building block for a larger DGX BasePOD or SuperPOD cluster
Specifications
- GPUs8x Blackwell Ultra
- GPU memory2.1 TB
- Networking800 Gb/s
- Size10 RU
| GPUs | 8x NVIDIA Blackwell Ultra SXM |
|---|---|
| FP4 performance | 144 PFLOPS sparse, 108 PFLOPS dense |
| FP8/FP6 performance | 72 PFLOPS sparse, 36 PFLOPS dense |
| NVLink switch systems | 2x |
| NVLink bandwidth | 14.4 TB/s aggregate |
| GPU memory | 2.1 TB total |
|---|---|
| HBM3e bandwidth | 62 TB/s |
| System memory | 2 TB, configurable to 4 TB |
| Processors | Intel Xeon 6776P |
|---|
| Cluster | 8x OSFP ports serving 8x ConnectX-8 VPI, up to 800 Gb/s InfiniBand or Ethernet |
|---|---|
| Storage and in-band | 2x dual-port QSFP112 BlueField-3 DPU, up to 400 Gb/s InfiniBand or Ethernet |
| Management | 1 GbE onboard NIC with RJ45, and 1 GbE RJ45 host BMC |
| OS | 2x 1.9 TB NVMe M.2 |
|---|---|
| Internal | 8x 3.84 TB NVMe E1.S |
| Power | 14.5 kW (busbar), 15.1 kW (PSU) |
|---|---|
| Rack units | 10 RU |
| Dimensions (H x W x L) | 442 x 482.6 x 904.2 mm |
| Operating temperature | 10-35°C |
| Included | NVIDIA DGX OS |
|---|---|
| Recommended | NVIDIA AI Enterprise, NVIDIA Mission Control |
| Also supports | Red Hat Enterprise Linux, Rocky, Ubuntu |
| NVIDIA support tier | Three-year business-standard hardware and software support |
| Lead time | Confirmed when we talk |
|---|---|
| Warranty | Confirmed when we talk |
| Support | Confirmed when we talk |
| Installation | Confirmed when we talk |
Specifications come from the NVIDIA DGX B300 datasheet, October 2026 and the NVIDIA DGX B300 user guide. NVIDIA may change them.
Related systems
AI Server 8 GPU
Up to eight dual-slot GPUs in a 4U ASUS chassis, for enterprise AI and HPC.
- GPUs, up to
- 8× H200 NVL
- Memory, up to
- 1,536 GB
- Chassis
- 4U ASUS
Price confirmed on your quote
Configure
Frequently asked questions
What does the NVIDIA DGX B300 need on site?
NVIDIA lists power as 14.5 kW (busbar), 15.1 kW (PSU), rack space as 10 RU, operating temperature as 10-35°C. We can talk through power, cooling and rack space for your site.
How much does the NVIDIA DGX B300 cost?
Every system shows an indicative price in rand, with and without VAT, which changes as you configure it. Your final price is confirmed on your quote.
What are the lead time and warranty?
We confirm lead time, warranty, support and installation when we talk, because they depend on the configuration and your site.
Can you set up the software as well?
Yes. We can deploy open-weight models from our catalogue on your hardware, and build the agents, automation and integrations that run on them.