NVIDIA HGX B300 and H200 Servers — Gigabyte

HGX B300 is based on NVIDIA's Blackwell Ultra generation, while HGX H200 is based on the Hopper generation. According to NVIDIA, B300 has 288 GB of HBM3e memory per GPU and H200 has 141 GB; GPU-to-GPU NVLink bandwidth is 1.8 TB/s per GPU on B300 and 900 GB/s on H200. The right choice depends on model size, workload type, power and cooling capacity, and your existing infrastructure.

For HGX B300 and B200, Gigabyte offers 8U air-cooled models (G894, G893) and 4U direct liquid-cooled models (G4L4, G4L3); for H200, this range includes a 5U DLC model (G593). Air cooling fits more easily into existing data centers but takes up more rack space. DLC fits more GPUs into the same rack space and handles high power density, but requires facility infrastructure such as coolant distribution units (CDUs). We recommend models based on your data center's kW-per-rack capacity.

In HGX servers, eight SXM GPUs sit on a single baseboard and are directly interconnected through NVLink/NVSwitch; this design targets large-model training, where GPU-to-GPU communication is intensive. In PCIe GPU servers, GPU cards are installed in PCIe slots and GPU-to-GPU communication typically runs over the PCIe bus. PCIe servers are a more flexible option for workloads such as inference, fine-tuning, visualization and VDI.

HGX-based servers are supplied per project and built to order according to the selected configuration; lead time varies with GPU supply and configuration. Request a quote for the current status.

TopTel is a distributor; we deliver Gigabyte HGX servers through a network of authorized partners (resellers, system integrators, solution partners). If you are a partner, you can request a configuration, lead time and project design quote via the Request a Quote button on the model page or the contact form. If you are an end user, we will direct you to a partner that fits your needs.