NVIDIA HGX B300 8-GPU Platform

The NVIDIA HGX B300 is the Blackwell Ultra generation of the eight-GPU HGX baseboard, built for the reasoning models and agentic workloads that now dominate inference demand. Its eight Blackwell Ultra SXM GPUs present 2.1 TB of HBM3e as a single NVLink domain, dense FP4 throughput rises to 108 PFLOPS against 72 PFLOPS on HGX B200, and attention-layer performance is doubled. ConnectX-8 SuperNICs are integrated directly on the baseboard, lifting networking bandwidth to 1.6 TB/s so that multi-node clusters scale without a separate NIC tray. FP64 throughput is deliberately reduced, which makes the B300 an inference and training part rather than a classical HPC accelerator.

Typical customers are AI cloud operators, sovereign AI programmes, telecom operators building inference-as-a-service and enterprises that need to serve long-context models to thousands of concurrent users. HGX B300 is shipped inside OEM chassis from Supermicro, Dell, HPE, Lenovo, GIGABYTE and ASUS, usually in 8U or 10U form with air or direct-liquid cooling, and is often ordered in multiples of eight nodes to populate a full NVLink-connected scalable unit.

InTechnologies FZCO arranges HGX B300 systems for business customers in the UAE, the wider GCC and Central Asia. Because this is controlled data-centre hardware, the contract and invoice are issued by one of our licensed partners, and our own role covers OEM sourcing, consolidation at our Dubai hub and DDP or DDU delivery to the customer site. Blackwell Ultra GPUs fall under US export regulations, so we complete end-user verification for every destination before confirming an order.