Comparisons_
NVIDIA B200 vs GB200: HGX vs Rack-Scale, and What Changes to Deploy
A field engineer's guide to the deployment differences between NVIDIA HGX B200 and GB200 NVL72 systems, covering rack assembly, cabling, liquid cooling, and commissioning, with emphasis on the shift from air-cooled HGX to liquid-cooled rack-scale GB200.
Key facts
- HGX B200 uses an 8-GPU baseboard with NVLink 4.0 over a copper backplane; GB200 NVL72 integrates 72 GPUs across 18 compute trays with NVLink 5.0 over a copper spine inside the rack.
- GB200 NVL72 requires liquid cooling (direct-to-chip) for all compute and NVSwitch trays; HGX B200 can be air-cooled with rear-door heat exchangers or liquid-cooled via cold plates.
- HGX B200 rack integration uses standard 19-inch EIA-310 racks with 4U or 8U chassis; GB200 NVL72 uses a custom 42U rack with integrated power, cooling, and networking backplane.
- GB200 NVL72's copper NVLink spine is factory-terminated and tested; field work involves routing and connecting MPO trunk cables for the scale-out InfiniBand/Ethernet network, not NVLink.
- Liquid cooling for GB200 NVL72 requires a CDU (Coolant Distribution Unit) sized per OEM spec for flow rate and temperature range; HGX B200 liquid cooling can use facility water or a separate CDU.
- Commissioning GB200 NVL72 includes verifying liquid cooling loop integrity via pressure test and flow verification before power-on; HGX B200 commissioning focuses on GPU-to-GPU NVLink bandwidth validation via nvidia-smi.
- GB200 NVL72's power density per rack can exceed 100 kW, requiring dedicated 480V 3-phase power distribution and high-capacity busways; HGX B200 racks typically draw 30–40 kW per rack when populated with multiple chassis.
Rack Architecture: HGX B200 vs GB200 NVL72
HGX B200 is a modular GPU baseboard system designed for standard 19-inch EIA-310 racks. Each HGX baseboard holds 8 GPUs interconnected via NVLink 4.0 over a copper backplane, and the baseboard is installed into a 4U or 8U chassis that also houses the CPU host nodes, networking NICs, and power supplies. Rack integration is straightforward: mount the chassis, connect power (typically 200–240 V or 277 V AC), and run fiber MPO trunk cables from the chassis's InfiniBand/Ethernet ports to the top-of-rack or leaf switches. The NVLink domain is entirely within the baseboard, so no external NVLink cabling is required.
GB200 NVL72 is a rack-scale system where 72 GPUs are distributed across 18 compute trays (each containing a Grace CPU and two B200 GPUs) plus 9 NVSwitch trays that provide NVLink 5.0 connectivity over a copper spine embedded in the rack's backplane. The rack is a custom 42U enclosure with integrated power distribution, liquid cooling manifolds, and a networking backplane. The copper NVLink spine is factory-terminated and tested; field deployment focuses on mounting the compute and switch trays, connecting the liquid cooling quick-disconnects, routing the MPO trunk cables for the scale-out network (InfiniBand or Ethernet), and terminating power cables to the rack's busway. Unlike HGX B200, the NVLink in GB200 NVL72 spans the entire rack and is not field-serviceable.
Cooling Infrastructure: Air vs Liquid
HGX B200 can be deployed with air cooling using rear-door heat exchangers or facility air handling, provided the data center has sufficient cooling capacity per rack (typically 30–40 kW per rack). For higher density, liquid cooling is an option via cold plates attached to the GPU baseboard, with facility water or a separate CDU providing the coolant. The liquid cooling loop for HGX B200 is usually a secondary loop with a CDU that isolates the facility water from the electronics. The field engineer must ensure cold plates are torqued per OEM spec, tubing is routed without kinks, and quick-disconnects are leak-tested before power-on. Coolant quality (deionized water or dielectric fluid) must be maintained per the manufacturer's guidelines.
GB200 NVL72 requires liquid cooling for all compute and NVSwitch trays. The rack has integrated manifolds that distribute coolant to each tray via quick-disconnect couplings. The CDU is external to the rack and must be sized to handle the rack's total heat load, which can exceed 100 kW per rack. The field engineer must connect supply and return hoses from the CDU to the rack's manifolds, verify flow rates and temperatures per OEM spec, and perform a pressure test of the entire loop before filling. The cooling loop must be free of debris and properly deionized to prevent corrosion. Unlike HGX B200, there is no air-cooled fallback for GB200 NVL72; if liquid cooling fails, the system must shut down immediately.
Cabling and Networking: Scale-Out vs NVLink Separation
In both HGX B200 and GB200 NVL72, the NVLink domain is copper and internal to the chassis or rack. The field engineer's cabling work is exclusively for the scale-out network (InfiniBand or Ethernet) and management network. For HGX B200, this means running MPO trunk cables from the chassis's NIC ports to the leaf switches, typically using 12-fiber or 24-fiber MPO trunks with LC or MPO connectors at the switch side. The cabling must follow the structured cabling plan, with proper bend radius (per TIA-568.3 or OEM spec) and slack management. Polarity must be correct (Type A, B, or C per TIA-568.3) and all connectors inspected with a fiber microscope before mating.
For GB200 NVL72, the scale-out network is distributed across the compute trays, each with multiple NICs that connect to the rack's networking backplane. The backplane has MPO ports that aggregate network traffic to top-of-rack switches. The field engineer routes MPO trunk cables from the rack's backplane ports to the switches, ensuring polarity and cleanliness. The management network (Ethernet for BMC/IPMI) is typically copper Cat6A or Cat8, terminated with RJ45 connectors and tested with a cable certifier. The critical rule: never confuse NVLink with the scale-out network; nvidia-smi 'NVLink up' status depends on the copper spine, not the fiber MPO cables.
Commissioning Sequence: Power, Cooling, and Validation
Commissioning GB200 NVL72 must follow a strict sequence: first, verify the liquid cooling loop integrity by performing a pressure test (typically at 1.5x the operating pressure per OEM spec) and then fill the loop with coolant. Next, power up the CDU and confirm flow rates and temperatures are within spec. Only then can you power on the rack's PDU and the compute trays. After power-on, use the BMC interface to check for hardware errors, then run nvidia-smi to verify all 72 GPUs are detected and NVLink status shows 'Active' for all links. Finally, run a GPU stress test (e.g., NVIDIA's DCGM diagnostic) to validate performance under load. Ensure the CDU flow meters and temperature sensors are calibrated before commissioning.
For HGX B200, the sequence is simpler: mount the chassis, connect power and networking, then power on. After boot, use nvidia-smi to verify all 8 GPUs are present and NVLink is active. Run a GPU stress test to confirm thermal performance under air or liquid cooling. If liquid cooling is used, the pressure test and flow verification must be done before power-on, similar to GB200 NVL72. The key difference is that HGX B200's cooling loop is per-chassis, while GB200 NVL72's loop is per-rack, so the commissioning scope is larger for GB200. Always document the coolant type and flow rates per OEM spec.
Common Failure Modes in the Field
One frequent failure in GB200 NVL72 deployment is improper engagement of liquid cooling quick-disconnects. If a tray is not fully seated, the quick-disconnect may not open, starving the tray of coolant and triggering thermal shutdown. Always verify that the tray is fully inserted and the latch is engaged, and monitor the CDU flow meter for each branch. Another failure is coolant leaks at quick-disconnect couplings; perform a leak check with a pressure test before filling and use leak detection tape or sensors at vulnerable points. A third common issue is MPO connector contamination: dust or oil on the fiber endface causes link errors or link down. Inspect and clean every MPO connector with a one-click cleaner or lint-free wipes before mating, and use a fiber microscope to verify endface quality per IEC 61300-3-35.
For HGX B200, a common issue is GPU-to-GPU NVLink bandwidth degradation due to a loose baseboard connection. Catch this by running nvidia-smi topo -m and checking for unexpected NUMA node distances. Another failure is power supply imbalance: if the chassis's PSUs are not evenly loaded, the system may throttle. Use the BMC to check power readings and ensure all PSUs are active. In both systems, a frequent oversight is not labeling cables properly, leading to troubleshooting delays. Use a label printer with heat-shrink labels per TIA-606-B standards. Always document the topology and cable paths during installation.
Power and Thermal Density Planning
GB200 NVL72 racks have a power density that can exceed 100 kW per rack, requiring dedicated 480 V 3-phase power distribution with busways rated for that load. The field engineer must ensure the facility's power infrastructure can support the rack's peak draw, including inrush current during startup. Soft-start or staggered power-on sequencing may be necessary. The CDU must be sized to handle the rack's heat load, with coolant flow rates typically in the range of 30–50 liters per minute per rack per OEM spec. Facility water temperature must be within the CDU's operating range (e.g., 18–32°C) to avoid condensation or insufficient cooling. Water quality (conductivity, pH, and particulate level) must meet CDU requirements.
HGX B200 racks typically draw 10–15 kW per chassis, so multiple chassis can be installed in a single rack, but total rack power should not exceed the facility's per-rack limit (often 30–40 kW for air-cooled racks). For liquid-cooled HGX B200, a shared CDU across multiple racks is possible, but the piping layout must avoid long hose runs that increase pressure drop. In both cases, thermal design must account for hot-aisle/cold-aisle containment if air cooling is used, or for liquid cooling, ensure the CDU is located within the recommended distance from the rack to minimize pump head. Always verify facility power and cooling capacity before rack installation.
Standards referenced: TIA-568.3 (Fiber Optic Cabling Standard) · TIA-606-B (Administration Standard for Telecommunications Infrastructure) · IEC 61300-3-35 (Fiber Optic Connector Endface Inspection) · EIA-310 (Rack Enclosure Standard)
Frequently asked_
Can I use the same CDU for both HGX B200 and GB200 NVL72 racks?
Yes, but you must ensure the CDU's flow rate and temperature range match the OEM spec for each system. GB200 NVL72 typically requires higher flow rates (30–50 L/min per rack) than a single HGX B200 chassis (5–10 L/min). If the CDU is shared, size it for the total load and use separate branch circuits with flow control valves. Leviathan Systems often deploys dedicated CDUs per GB200 rack to simplify commissioning and avoid cross-contamination.
Do I need to clean MPO connectors before connecting them to the GB200 NVL72 rack backplane?
Absolutely. The backplane MPO ports are factory-terminated and should be clean, but the trunk cables you bring in may have dust caps that don't guarantee cleanliness. Always inspect the endface with a fiber microscope (per IEC 61300-3-35) and clean with a one-click cleaner or lint-free wipe if any contamination is visible. A single dirty connector can cause link errors that are hard to diagnose later.
What happens if a compute tray in GB200 NVL72 is not fully seated?
The tray's liquid cooling quick-disconnects will not open, so the tray will not receive coolant. The BMC will report a thermal event, and the system may throttle or shut down the tray. You must reseat the tray firmly until the latch clicks, then verify coolant flow via the CDU's branch flow meter. Leviathan Systems always checks tray seating during installation and before power-on.
Is the NVLink in GB200 NVL72 field-serviceable?
No. The NVLink 5.0 copper spine is integrated into the rack's backplane and is factory-terminated and tested. If an NVLink link fails, the entire rack may need to be replaced or sent back to the manufacturer. Field engineers only work on the scale-out network (fiber MPO) and the liquid cooling loop. Always verify NVLink status via nvidia-smi during commissioning to catch any factory defects.
Can I air-cool a GB200 NVL72 rack?
No. GB200 NVL72 is designed exclusively for liquid cooling. The compute and switch trays have no fans for GPU cooling; they rely entirely on the liquid cooling loop. Attempting to run them without liquid cooling will cause immediate thermal damage. The rack must be connected to a CDU with the correct coolant and flow rate before power-on.