NVIDIA HGX B300 buyer’s guide
Air vs liquid cooling for HGX B300
Choose cooling for facility kW and training density—then configure Elite APEX hosts with fabric and contract-ready BOMs.
Buyer’s guide
Size HGX B300 to the facility—not brochure TFLOPS
NTS configures Elite APEX HGX B300 systems in Fremont with fabric, power, and contract-ready BOMs for large-model training and multi-node generative AI.
Use air when CRAH/CRAC capacity and rack density allow faster deploy without CDU plant. Move to DLC for sustained 8-GPU training, high kW/rack, and clock stability under continuous load.
TCO is dominated by silicon, power, cooling plant, and time-to-science—NTS models node kW, fabric, and air vs DLC CapEx/OpEx before PO.
Decide & configure
Cooling paths, platforms, and procurement
Pick air or DLC, then continue into Elite APEX SKUs, liquid hubs, staging, and contract vehicles.
- Power + airflow
- Faster CDU-free deploy
- Watch sustained clocks
- Clock stability
- High kW/rack
- Facility worksheets
- CDU redundancy
- Flow & QDs
- kW worksheets
- B300 8-GPU
- NDR / RoCE
- ETO BOMs
- Thermal soak
- Serial manifests
- Nationwide rack-stack
- Vehicle selection
- CLIN structure
- Fallback paths
- Rack design
- Fabric attach
- Acceptance tests
- Air vs DLC call
- Fabric map
- Contract quoting
Frequently asked questions
Choose B300 for large-model training and multi-node generative AI where GPU-to-GPU NVLink/NVSwitch bandwidth matters, and programs standardize on Blackwell Ultra / B300 tooling with engineer-to-order EPYC or Xeon hosts.
- 8-GPU HGX density
- Documented PCIe/NIC topology
- Federal and enterprise ETO paths
Air suits facilities with adequate CRAH/CRAC and moderate density. DLC is recommended for sustained 8-GPU training, high kW/rack, and continuous-load clock stability—NTS includes CDU/manifold worksheets in architecture review.
- Air: simpler, faster when plant allows
- DLC: density and thermal headroom
- Facility decision as much as silicon
Silicon, power, cooling plant CapEx/OpEx, fabric so GPUs are not starved, and time-to-science. NTS models node kW aggregation (including NICs/storage) and air vs DLC tradeoffs at your GPU count.
- Node and rack kW
- Air vs DLC facility cost
- NDR InfiniBand or high-speed Ethernet
Systems quote on SEWP V, ITES-4H, and GSA MAS with BOM validation. Call-for-price catalog SKUs receive CLIN-aligned quotes rather than public Offer schema prices.
- Multi-vehicle mapping
- Staging on the same vehicle when possible
- Architecture review before PO
