HPE Compute XD690 — 8x NVIDIA Blackwell Ultra B300 Air-Cooled AI Server

Specifications

Overview

The HPE Compute XD690 is an air-cooled eight-way GPU server carrying eight NVIDIA Blackwell Ultra B300 HGX GPUs. It is the practical on-ramp to Blackwell Ultra for operators who are not ready to commit to a rack-scale NVL72 deployment or a liquid-cooled facility: the node fits standard enterprise racks and existing power and airflow budgets, while delivering the same B300 silicon that anchors the top of the NVIDIA stack.

HPE quotes a 50% increase in FP4 throughput and double the GPU-to-CPU bandwidth versus the previous Blackwell generation, which is the pair of numbers that decides whether a large MoE model can be served from a single node or must be sharded across several. That bandwidth matters disproportionately for reasoning and long-context workloads, where the bottleneck is frequently moving activations and expert routing data rather than raw matrix throughput.

What HPE brings beyond the GPU count is cluster operations. HPE Performance Cluster Manager handles provisioning, management and monitoring for clusters that scale to 100,000 nodes, and the vendor pairs the hardware with dedicated deployment engineers, project managers and HPE Tech Care reactive support. For buyers assembling a hundred-node AI cluster, that operational layer is often the deciding factor rather than a per-GPU specification.

The XD690 sits alongside the HPE Cray XD670, the previous-generation node built on eight NVIDIA H200 SXM GPUs with dual 5th Gen Intel Xeon Scalable processors, which remains in production for buyers optimising cost per token on mature models. QS Compute quotes both, together with the matching NVMe, fabric and power infrastructure for a complete AI node.

Key Benefits

Eight B300 HGX GPUs in an air-cooled node avoid the liquid-cooling retrofit that rack-scale systems demand. 50% higher FP4 and 2x GPU-CPU bandwidth directly increase served context length and tokens per second. HPE Performance Cluster Manager to 100,000 nodes turns individual nodes into an operable AI factory. Deployment engineers and HPE Tech Care shorten the path from delivery to production. Generation-over-generation continuity with the XD670 lets an existing HPE AI fleet scale without a new operational model.

Applications

Large language model training, fine-tuning and inference at scale; mixture-of-experts serving and reasoning workloads; enterprise and sovereign AI factories; research HPC clusters with AI workloads; model tuning and evaluation infrastructure; inference capacity added next to existing air-cooled HPE fleets.

Related Products