NVIDIA B300 (Blackwell Ultra) 8-GPU AI Server
| Brand | NVIDIA |
| Model | B300 (Blackwell Ultra) |
| Configuration | 8x B300 288GB HBM3E / dual Intel Xeon 6776P / DDR5 |
| Warranty | 3 months - 3 years |
Overview
The NVIDIA B300 (Blackwell Ultra) is a data-center GPU built for generative AI and large-scale inference. Compared with B200, B300 raises HBM3E capacity to 288GB per GPU (279GB usable), memory bandwidth to 8TB/s, and strengthens FP4 throughput to about 15 PFLOPS dense NVFP4. For trillion-parameter MoE models, long-context inference and multimodal generation, the value of B300 is simple: more memory and more FP4 throughput per GPU, which means fewer cards for the same model and a lower cost per generated token.
Key specifications
| Architecture | NVIDIA Blackwell Ultra |
| Memory | 288GB HBM3E (279GB usable) |
| Memory bandwidth | 8TB/s |
| Low-precision compute | ~15 PFLOPS NVFP4 (dense) |
| Interconnect | NVLink 5 |
| Form factor | SXM module / 8-GPU server |
| GPU power | ~1400W |
DGX B300 8-GPU reference configuration
Using the NVIDIA DGX B300 as a reference, an 8-GPU system provides about 2.1TB of total GPU memory, roughly 144 PFLOPS FP4 and 72 PFLOPS FP8, paired with dual Intel Xeon 6776P processors. CPU, memory, NVMe storage and InfiniBand or RoCE networking can be tailored to the workload.
| GPU | 8x NVIDIA B300 288GB HBM3E |
| Total GPU memory | ~2.1TB |
| FP4 compute | ~144 PFLOPS |
| FP8 compute | ~72 PFLOPS |
| CPU | 2x Intel Xeon 6776P |
| Networking | Optional InfiniBand / RoCE fabric |
Typical workloads
Trillion-parameter MoE inference and multi-expert deployment
Long-context RAG, document understanding and agent workloads
Multimodal and video generation, high-throughput image inference
Scientific computing and FP4 quantized inference acceleration
Supply and delivery
We supply through two channels: standard 8-GPU B300 systems are usually available from stock or with a short lead time, while custom CPU, memory, storage and fabric configurations are build-to-order. Every system is burn-in tested, GPU stress tested and memory checked before shipment. Complete servers, SXM modules and spare parts are all available.
Warranty and service
New complete servers carry manufacturer warranty; refurbished modules are covered for 3 to 12 months. We support rack installation, cluster fabric commissioning, driver and CUDA environment setup, and ongoing maintenance. Server pricing moves with the market, so please request a current quote.
FAQ
What is the difference between B300 and H200, and which should I choose?
How much memory does a B300 have, and what is the total for an 8-GPU system?
B300 draws 1400W per GPU — what power and cooling does the data center need?
Are B300 servers in stock or built to order, and what is the lead time?
For a configuration list or a current price, please send an inquiry and we will reply with recommendations and lead time during business hours.
