English 
| Product Name: | |
|---|---|
| Platform: | |
| GPU Architecture: | |
| GPU Type: | |
| GPU Interconnect: | |
| Memory Type: | |
| PCIe Version: | |
| Quantity: | |
B300 Server
NVIDIA
The NVIDIA B300 GPU server is built on the NVIDIA Blackwell Ultra architecture, designed for large-scale AI training, high-throughput inference, and LLM workloads in modern data center environments. Deployed via the HGX platform with 8-GPU configurations, B300 systems support sixth-generation NVLink fabric and increased memory capacity for trillion-parameter model workloads. Specifications below are sourced from NVIDIA official product documentation.
The Blackwell Ultra GPU features fifth-generation Tensor Cores with FP4 and FP8 acceleration. An 8-GPU HGX B300 configuration delivers 72 PFLOPS of FP8 tensor throughput and 144 PFLOPS of FP4 tensor throughput, supporting mixed-precision training and high-throughput inference pipelines.
B300 GPUs use the SXM form factor for maximum NVLink bandwidth. Each GPU integrates 288 GB of on-package HBM3e memory, with 8-GPU HGX systems providing 2.3 TB of total GPU memory across the baseboard.
Parameter | HGX B300 (8-GPU) |
Total GPU Memory | 2.3 TB HBM4 |
Aggregate Memory Bandwidth | 176 TB/s |
Per-GPU Memory | 288 GB HBM3e |
NVLink Generation | Sixth Generation |
NVSwitch Bandwidth | 28.8 TB/s |
The HGX B300 platform provides 3.6× more on-package memory per GPU compared with the H100 SXM configuration, enabling larger model sizes and reduced KV-cache offloading during inference workloads.
Sixth-generation NVLink technology delivers higher per-link bandwidth compared with Hopper-generation systems. Integrated NVSwitch on the HGX baseboard provides 28.8 TB/s of total switch bandwidth, enabling full-mesh all-to-all communication across all 8 GPUs within a single node.
For multi-node deployments, B300 systems integrate with next-generation InfiniBand and Ethernet networking solutions. The NVIDIA Quantum-X800 InfiniBand platform supports 800 Gb/s per link for large-scale distributed training clusters.
Increased memory capacity and bandwidth allow B300 servers to handle larger model partitions per GPU, reducing communication overhead in data-parallel and tensor-parallel training configurations. FP4 support enables higher throughput for pre-training workloads where precision requirements allow.
Higher per-GPU memory capacity supports larger KV caches and more concurrent inference sessions per server, improving throughput for generative AI services and LLM API deployments in commercial operations.
FP64 and TF32 compute capabilities support scientific simulation, computational biology, and climate modeling workloads. The Blackwell architecture includes dedicated acceleration for sparse computation and numerical methods.
HGX B300 servers are delivered as 8U rackmount systems through NVIDIA-Certified Systems partners. Deployments require liquid cooling infrastructure and high-wattage redundant power supplies compatible with modern AI factory data center designs.
B300 systems support the full NVIDIA software stack including CUDA, cuDNN, TensorRT, and all major AI frameworks. NVIDIA AI Enterprise subscriptions provide enterprise support, security patches, and managed software distributions for commercial procurement environments.
The B300 uses the Blackwell Ultra architecture with higher memory capacity (288 GB per GPU vs. 80 GB on H100 SXM) and sixth-generation NVLink. FP8 aggregate throughput for an 8-GPU HGX B300 system reaches 72 PFLOPS, compared with 32 PFLOPS aggregate for an 8-GPU HGX H100 configuration.
Each B300 Blackwell Ultra GPU integrates 288 GB of HBM3e on-package memory. At the HGX system level with 8 GPUs, the platform delivers 2.3 TB of total HBM4 memory with 176 TB/s of aggregate bandwidth.
Yes. The B300 supports the full NVIDIA CUDA software ecosystem and all major AI frameworks including PyTorch, TensorFlow, and JAX. Existing GPU-accelerated applications can run on Blackwell architecture with appropriate CUDA version updates.
