HGX B300

Overview

Data center GPUs are not deployed individually. NVIDIA packages them onto a standardized baseboard, the HGX platform, which server manufacturers build into complete systems. The HGX B300 carries eight Blackwell Ultra GPUs in SXM form factor, connected to each other over NVLink so they behave less like eight separate processors and more like one large one. It is the successor to the HGX B200 and the standard unit in which Blackwell Ultra capacity is bought, racked, and scaled.

How it works

The defining feature of the platform is what happens inside the node. Each GPU communicates with every other at 1.8 TB/s over fifth-generation NVLink, for 14.4 TB/s of total interconnect bandwidth, orders of magnitude faster than any external network. This matters because large models are split across GPUs, and the constant exchange of activations and gradients between them is often the real bottleneck. Blackwell Ultra also carries 288 GB of HBM3e per GPU, roughly two terabytes per system, so far more of a model stays in fast memory instead of being sharded across additional nodes. Beyond the node boundary, 1.6 TB/s of networking bandwidth connects each system to the wider fabric, which is how 64 systems combine into a single cluster.

Why it matters

The HGX system, not the individual GPU, is the real unit of AI infrastructure: workloads are scheduled in nodes, clusters are described in nodes, and performance depends on how well the node's internal fabric holds up under distributed training. A single HGX B300 delivers up to 144 petaflops of FP4 compute, and the generation's larger memory means fewer systems are needed to hold a given model, reducing the communication overhead that erodes scaling efficiency. When a cluster is described as 64 HGX B300 systems, that is a statement about its architecture, not just its size.