HGX B200 rack design guide
Liquid-cooled B200 rows from CDU to acceptance
Design DLC-ready HGX B200 training rows—cold plates, manifolds, fabric, and Fremont soak—without changing facility standards mid-PO.
Rack design guide
NTS designs DLC-ready B200 rows from CDU to acceptance—compute, cooling, fabric, storage attach, and Fremont burn-in on one accountable BOM.
Design inputs include GPUs per rack, north/south NIC plan, facility supply temperature and flow, power density, weight limits, and contract vehicle for hardware plus integration CLINs.
Avoid common failure modes: under-sized CDU, NIC counts that starve collectives, and shipping without leak/flow evidence.
Design & configure
Move from reference architecture into GPU SKUs, liquid hubs, staging, and procurement.
GPUs per rack, NIC plan (ConnectX/BlueField), facility supply temperature and flow, rear-door vs direct-to-chip preference, kW/rack and weight limits, plus contract vehicle for hardware and integration.
Dual-socket HGX B200 compute, cold plates/manifolds/QDs/CDU, InfiniBand NDR or RoCE east-west with separate OOB, NVMe plus parallel FS or object tiers, and Fremont thermal soak with fabric loopback.
Under-sized or non-redundant CDU for 24/7 training, NIC counts that starve GPU collectives, and shipping without leak/flow evidence that causes on-site acceptance failure.
Quote compute, cooling, and L11/L12 on SEWP V, ITES-4H, or GSA MAS with BOM validation—preferably the same vehicle for hardware and integration.