HPE's air-cooled eight-way NVIDIA server. It holds eight Blackwell Ultra accelerators on an HGX baseboard and is intended for organisations training, tuning and serving large models without moving to liquid cooling.
No payment is taken online. Every request is an enquiry answered by an engineer.
Where does this sit?
Enterprise. Four to eight enterprise accelerators in a rack chassis, sized for production load rather than experiments. Typically bought by an enterprise ai or platform team. Datacentre or a well-provisioned rack.
Who it's for
What it runs
When it is the wrong answer
Every figure below is explained in plain English. Switch to the technical view for the bare specification.
Rack-mounted air-cooled eight-way GPU server
2x Intel Xeon 6 — exact SKU range awaiting verification
8x NVIDIA Blackwell Ultra (B300 HGX)
MeansHGX is a factory-fitted baseboard of eight linked accelerators, not eight separate cards.
MattersThe internal NVLink between them is what makes large-model training practical.
DDR5 — maximum capacity awaiting verification
NVMe SFF — exact count awaiting verification
NVIDIA ConnectX and BlueField NICs, OCP options — exact port speeds awaiting verification
Air-cooled, per HPE product page
High-wattage redundant PSUs — exact rating awaiting verification
Verification
Manufacturer documented — HPE Compute XD690 product page and QuickSpecs a50009220enw. Reviewed 2026-08-31.
Manufacturer sourceThe things that catch people out after the hardware has already been ordered.
What else will I need?
Relationships documented by the manufacturer, or by us during a deployment.
Ethernet AI fabric switching up to 800GbE, supported within the HPE portfolio.
200Gb/s InfiniBand switching and adapters for HPC and AI clusters.
Scale-out object storage built for AI data pipelines and RAG.