Platforms_
NVIDIA H100 vs H200 vs B200: What Changes for Deployment
A field engineer's guide to the physical deployment differences between NVIDIA H100, H200, and B200 GPU platforms, covering power, cooling, cabling, and rack density changes that affect data center build and commissioning.
Key facts
- H100 SXM and H200 SXM both use the same HGX baseboard form factor and 700 W TDP per GPU, making power and cooling infrastructure identical between them.
- B200 SXM increases TDP to 1000 W per GPU, requiring liquid cooling (direct-to-chip or immersion) for sustained operation; air cooling is not viable at that density.
- H100 and H200 use NVLink 3.0 (600 GB/s per GPU) over the copper NVSwitch backplane inside the rack; B200 uses NVLink 4.0 (900 GB/s per GPU) with a new NVSwitch tray design that changes cable routing.
- H200 doubles HBM3e capacity to 141 GB per GPU versus H100's 80 GB, but this does not change physical cabling or power delivery—only firmware and memory stack.
- B200 introduces a 1.8 TB/s NVLink bandwidth per GPU, requiring higher-quality MPO trunk cables (OS2 single-mode for longer reaches) in the scale-out network to avoid signal degradation.
- The B200 HGX baseboard has a different power connector pinout (12VHPWR vs. previous generation) and requires updated PDU busbars or cable whips; mixing with H100/H200 PSUs causes incompatibility.
- All three platforms use the same 400 Gbps OSFP or QSFP-DD ports for InfiniBand/Ethernet scale-out, but B200's higher compute density may require more switch ports per rack, increasing MPO trunk cable counts.
Power Delivery: From 700 W to 1000 W per GPU
The H100 and H200 SXM modules share a 700 W TDP and identical HGX baseboard power delivery. This means the same 48 V busbars, the same PDU ratings, and the same cable whip gauges work for both. If your site is wired for H100, you can swap in H200 without touching power infrastructure—the only change is firmware and memory. The B200, however, jumps to 1000 W per GPU. That extra 300 W per GPU adds up fast: an 8-GPU B200 HGX baseboard draws 8 kW just for the GPUs, plus the NVSwitch and CPU. You must verify your PDU per-outlet rating, busbar ampacity, and cable whip gauge against the OEM spec. The B200 also uses a 12VHPWR connector on the baseboard, not the previous 8-pin EPS12V. Do not assume backward compatibility—check the pinout diagram before plugging in.
For the rack-level power chain, the B200's higher density means you may need to reduce the number of GPUs per rack or upgrade to higher-current PDUs. In a standard 42U rack, an H100/H200 system typically fits 4-6 HGX baseboards (32-48 GPUs) with air cooling. A B200 system with liquid cooling can fit 8 baseboards (64 GPUs) but only if the PDU can deliver 64 kW+ per rack. Plan your power budget per rack before ordering busbars. Leviathan Systems has seen sites where the PDU was rated for 60 kW but the B200 rack required 72 kW—that mismatch halts deployment until the PDU is swapped.
Cooling: Air vs. Liquid for B200
H100 and H200 SXM modules are air-cooled with a standard heatsink and fan tray. The thermal design power of 700 W is manageable with high-CFM fans and proper airflow management—typical rack inlet temperatures of 20-25°C work fine. The B200 at 1000 W cannot be air-cooled in a dense rack environment. The OEM specifies liquid cooling (direct-to-chip cold plates or immersion) for sustained operation. This changes the deployment workflow: you must install a coolant distribution unit (CDU), run coolant supply/return hoses to each GPU tray, and commission the loop with the correct coolant mixture (typically deionized water with a glycol additive).
Field experience shows that the most common failure in B200 liquid cooling is air entrapment in the cold plates during initial fill. This causes hot spots and GPU throttling. The fix is a slow, controlled fill with the rack tilted slightly forward (if the OEM allows) and using a vacuum-assisted fill kit. Also, the coolant flow rate per GPU must be verified against the OEM spec—too low and the GPU overheats; too high and you risk erosion of the cold plate microchannels. Always use a calibrated flow meter during commissioning. For H100/H200, the cooling check is simpler: verify fan speeds and inlet/outlet temps via nvidia-smi. For B200, you must also check coolant temperature, flow rate, and pressure at each tray.
Cabling: NVLink and Scale-Out Network Differences
The NVLink domain is copper inside the rack for all three platforms. H100 and H200 use NVLink 3.0 over the NVSwitch backplane—no fiber or MPO involved. The cabling work is limited to the power and management cables for the NVSwitch tray. B200 uses NVLink 4.0 with a redesigned NVSwitch tray that has a different cable routing pattern. The copper cables between the GPU baseboard and the NVSwitch tray are shorter and have a higher density of connectors. You must follow the OEM cable routing guide exactly—crossing cables or exceeding the minimum bend radius will cause signal integrity failures. The NVLink status check via nvidia-smi will show all links up only if the copper cables are properly seated and not damaged.
The scale-out network (InfiniBand or Ethernet) uses MPO fiber trunk cables for all three platforms. The port count per GPU is the same (typically 8 ports per GPU for H100/H200, 8 for B200), but B200's higher compute density may require more switch ports per rack. This increases the number of MPO trunk cables from the GPU trays to the leaf switches. For H100/H200, a typical rack uses 12-16 MPO-24 trunks. For B200, that can jump to 20-24 trunks. The cable routing in the overhead tray must account for this increased density—do not exceed the tray weight limit. Also, B200's higher bandwidth (1.8 TB/s NVLink) means the scale-out network may need OS2 single-mode fiber for longer reaches (over 100 m) to avoid signal loss. Verify the transceiver type and fiber mode against the OEM spec before ordering.
Rack Density and Weight Distribution
H100 and H200 systems weigh approximately 150-200 kg per fully populated rack (with 4-6 HGX baseboards, switches, and PDUs). The weight is distributed evenly across the rack rails. B200 systems with liquid cooling add the weight of the CDU, coolant hoses, and cold plates—typically 250-300 kg per rack. This exceeds the standard 42U rack weight limit of 250 kg for some models. You must verify the rack's static load rating and use a reinforced rack if needed. Also, the liquid cooling loop adds a center-of-gravity shift: the CDU is often mounted at the bottom, making the rack bottom-heavy. This is fine for stability but complicates forklift moves—always secure the CDU before transport.
For the floor loading, a B200 rack at 300 kg over a 0.6 m x 1.2 m footprint gives a point load of about 4.2 kN/m². Most data center raised floors are rated for 12 kN/m², so this is safe. But if you stack multiple B200 racks in a row, the cumulative load on the floor tiles can exceed the rating. Check the floor load map with the facility team before deployment. Leviathan Systems has seen a case where a row of 10 B200 racks exceeded the floor rating by 15%, requiring a structural reinforcement that delayed the project by two weeks.
Common Failure Modes in the Field
The most frequent failure when deploying H100 vs H200 vs B200 is power connector mismatch. Engineers assume the same PDU cable works for all, but B200's 12VHPWR connector is physically different from H100/H200's EPS12V. Always verify the connector type on the baseboard before plugging in. A forced connection can damage the pins and cause arcing. Second, liquid cooling leaks in B200 systems: the quick-disconnect fittings on the coolant hoses are the weak point. Always pressure-test the loop at 1.5x the operating pressure before powering on the GPUs. Use a leak detector mat under each GPU tray during commissioning.
Third, MPO trunk cable damage during routing: the higher density of cables in B200 racks increases the chance of kinking or exceeding the bend radius. Always use a cable management arm and a bend-radius guide. Inspect each MPO connector with a fiber microscope before plugging in—a single dirty or damaged ferrule can cause link errors that are hard to trace. Fourth, NVLink copper cable seating: the high-density connectors on the B200 NVSwitch tray are easy to misalign. Use the OEM insertion tool and listen for the click. After seating, run nvidia-smi nvlink -s to verify all links are up. If any link shows 'down', reseat the cable. Finally, firmware mismatch: H200 requires a different firmware version than H100 on the same HGX baseboard. If you swap GPUs without updating the baseboard firmware, the system may not boot. Always check the OEM compatibility matrix before mixing GPU generations.
Commissioning Checklist Differences
For H100 and H200, the commissioning checklist is similar: power on the rack, verify PDU voltages, check fan speeds, run nvidia-smi to confirm all GPUs are detected, and run a stress test (e.g., nvidia-smi -pl 700 && stress --gpu 8). The NVLink check is quick: nvidia-smi nvlink -s should show all links up. For B200, the checklist adds liquid cooling steps: verify coolant flow rate per GPU (use the OEM's recommended range), check coolant temperature at the CDU outlet, and run a thermal soak test for 30 minutes at full load. Also, verify the 12VHPWR connector temperatures with a thermal camera—if any connector exceeds the OEM spec, shut down and reseat.
The B200 also requires a network bandwidth test at the scale-out layer because the higher NVLink bandwidth can expose bottlenecks in the InfiniBand/Ethernet fabric. Run a multi-node MPI benchmark (e.g., OSU bandwidth test) to confirm the fabric is not oversubscribed. For H100/H200, this is optional but recommended. Finally, document the firmware versions of all components (GPUs, NVSwitch, baseboard) and compare against the OEM's validated stack. A mismatch can cause intermittent errors that are hard to diagnose later.
Standards referenced: IEC 60320 (power connector ratings) · TIA-568.3-D (fiber optic cabling) · TIA-942 (data center infrastructure) · IEEE 802.3bs (400 Gbps Ethernet) · InfiniBand Trade Association specifications (for NDR/NDR200) · OEM-specific NVLink and HGX baseboard specifications (refer to vendor documentation)
Frequently asked_
Can I use the same power cables for H100 and B200 in the same rack?
No. H100 and H200 use EPS12V 8-pin connectors on the HGX baseboard, while B200 uses 12VHPWR connectors. The pinout and current rating differ. Using an H100 cable on a B200 baseboard will not fit and may damage the connector. You must replace the PDU cable whips or use adapter cables rated for 12VHPWR. Always check the OEM's power connector diagram for your specific baseboard revision.
Does the B200 require a different type of fiber for the scale-out network?
Not necessarily, but the higher bandwidth may push you to OS2 single-mode fiber for longer reaches. For distances under 100 m, OM4 or OM5 multimode fiber with 400 Gbps transceivers works fine. For distances over 100 m, OS2 single-mode is required to avoid signal loss. The transceiver type (SR4, DR4, FR4) determines the fiber mode. Check the OEM's network topology guide for your specific switch and GPU configuration.
Can I air-cool a B200 if I reduce the rack density?
The OEM specifies liquid cooling for B200 at 1000 W TDP. Air cooling is not supported for sustained operation, even at lower density. The heatsink and fan tray are not designed for that thermal load. Attempting air cooling will cause thermal throttling within minutes under load. If you must use air, consider the H200 instead, which is air-cooled at 700 W.
What is the most common mistake when deploying B200 for the first time?
The most common mistake is not pressure-testing the liquid cooling loop before powering on the GPUs. Quick-disconnect fittings can leak if not fully seated. Always perform a pressure test at 1.5x operating pressure for 30 minutes. Also, many teams forget to update the baseboard firmware to support B200—the firmware is different from H100/H200. Check the OEM's compatibility matrix before installation.
How do I verify NVLink is working correctly on B200?
Use nvidia-smi nvlink -s to check the status of all NVLink links. Each GPU should show 18 links up (for NVLink 4.0). If any link shows 'down', reseat the copper cable at both ends. Also, run nvidia-smi nvlink -g 0 -r to check the link error counters. Any non-zero error count indicates a signal integrity issue—check the cable bend radius and connector seating. For a full validation, run a multi-GPU bandwidth test like nvbandwidth.