Platforms_
B300 Deployment Guide: The Physical-Layer Step Up from B200
Compares the power distribution, liquid cooling, and cabling changes required when moving from B200 to B300 racks, with explicit steps and decision criteria for field crews performing GB300-class deployments.
Key facts
- B300 racks draw higher total power than B200 racks, requiring separate compute and networking power feeds per OEM rack documentation.
- Liquid cooling for B300 uses higher-flow CDUs and larger-diameter manifold connections than B200 loops to maintain inlet temperatures within ASHRAE TC 9.9 limits.
- Scale-out networking remains MPO-terminated fiber for InfiniBand or Ethernet between racks; NVLink stays on the internal copper backplane.
- TIA-942-B specifies minimum bend radius and pathway separation for power and data cables in AI halls.
- Factory-polished MPO trunks must be inspected with a calibrated scope before mating; field termination of MPO ferrules is not performed.
- Sequential power-up of B300 domains begins with networking PDUs before GPU domains to allow fabric bring-up.
- Leak testing of B300 manifolds uses nitrogen pressure hold followed by helium sniff at the OEM-specified duration.
Power distribution changes from B200 to B300
B300 racks require additional or higher-capacity PDUs because total rack power increases and the OEM separates compute and networking power domains. Crews must verify that each domain has its own feed from the busway or floor PDU before any connections are made. This separation prevents a networking fault from taking down GPU power and allows independent metering.
Run new power whips sized to the B300 nameplate before removing B200 hardware. Confirm phase balance across all three legs on each feed with a calibrated meter. Do not energize the GPU domain until the networking domain shows stable voltage and the fabric switches report link-up.
Liquid cooling loop upgrades
B300 cold plates and quick-disconnects are sized for higher flow than B200, so existing CDUs often need replacement or parallel addition. Manifold connections move to larger-diameter hoses; crews must confirm the correct barb or dry-break fitting before routing. Inlet temperature must stay inside the ASHRAE TC 9.9 envelope for the higher TDP parts.
Route supply and return lines with the supply on the left side of the rack when facing the rear, matching the B300 labeling. Pressure-test the entire loop with nitrogen to the OEM hold time, then perform a helium sniff at every fitting. Only after both tests pass should the loop be filled with the specified coolant mix.
Scale-out cabling differences
MPO trunks carry the InfiniBand or Ethernet scale-out fabric between B300 racks; the internal NVLink fabric remains copper on the rack spine. Trunk lengths must be calculated from the actual patch-panel locations rather than rack-row drawings, because B300 racks are deeper. Maintain minimum bend radius per the cable manufacturer and keep power cables on the opposite side of the rack from fiber pathways per TIA-942-B.
Clean every MPO connector with the OEM-approved cassette before insertion and inspect with a scope. Record the polarity and length of each trunk in the as-built documentation. Do not rely on nvidia-smi NVLink status to validate fiber links; those are separate domains.
Rack integration and power-up sequence
Position the B300 rack, land the power whips, and connect the liquid-cooling manifolds before any networking cables are routed. This order prevents fiber damage during mechanical alignment. Once manifolds are leak-checked, connect the networking domain power first so switches can boot and form the fabric.
Apply GPU-domain power only after the networking fabric reports all expected links. Run the OEM bring-up script that exercises both domains before handing the rack to software teams. Leviathan Systems crews document each step with timestamped photos for the commissioning package.
Common failure modes and how to catch them
The most frequent B300 issues are reversed supply/return cooling lines and insufficient PDU capacity on one domain. Both produce immediate thermal or power alarms during commissioning. Verify manifold orientation against the rack label and measure current on each PDU leg before GPU power-up.
MPO polarity errors appear as missing links between switches; an OTDR or continuity tester on the trunk before final dressing catches these. Over-torqued quick-disconnects leak after thermal cycling; use the torque wrench specified in the OEM termination kit and re-check after the first 24 hours of operation.
Validation and documentation
After integration, run the full power, thermal, and fabric test suite provided by the OEM. Capture inlet/outlet temperatures, PDU currents, and switch port status. Any parameter outside the published range must be resolved before the rack is released.
Update the DCIM system with the new rack power and cooling capacities. Store MPO inspection images and pressure-test results with the rack serial number. This package is required for the next maintenance window or capacity audit.
Standards referenced: ASHRAE TC 9.9 · TIA-942-B · OEM B300 rack power and cooling specifications
Frequently asked_
Can we reuse B200 PDUs and CDUs for a B300 rack?
No. B300 power draw exceeds B200 ratings on both domains, and the CDU flow rate is insufficient. New PDUs and at least one additional or larger CDU are required. Measure actual load after installation rather than relying on nameplate comparison alone.
How do we confirm MPO polarity on B300 scale-out trunks?
Use a calibrated MPO continuity tester before dressing the cables into the rack. Record the result against the trunk label. An OTDR trace after installation provides a second confirmation and baseline for future troubleshooting.
What is the correct order for connecting power domains on first power-up?
Networking PDUs must be energized and switches booted before GPU-domain power is applied. This allows the fabric to form so that any GPU-domain fault can still be reported over the network. The OEM bring-up script enforces this sequence.
Who typically performs the B300 manifold leak test on site?
Leviathan Systems field crews perform the nitrogen hold and helium sniff test after manifold connection but before coolant fill. Results are logged with timestamps and photos for the commissioning record.
How should crews verify NVLink versus scale-out links during bring-up?
NVLink status comes from the internal copper spine and appears in nvidia-smi output. Scale-out links are validated separately with the fabric management tools over the MPO fiber trunks. Confirm both before declaring the rack ready.