GB300 vs GB200: What Changes in Deployment?
The GB300 NVL72 brings Blackwell Ultra to the data center. Here is what is different from the infrastructure perspective compared to GB200.
The NVIDIA GB300 NVL72 represents the second generation of Blackwell-architecture rack-scale GPU systems. While the form factor and liquid cooling requirements remain similar to the GB200 NVL72, there are meaningful differences that affect infrastructure deployment.
The most significant change is the move to Blackwell Ultra GPUs with 288 GB HBM3e per GPU, up from 192 GB HBM3e on the GB200. This increases the per-GPU memory bandwidth and capacity, but does not significantly change the power or cooling envelope at the rack level.
NVLink 5.0 replaces NVLink 4.0 for intra-rack GPU communication, providing higher bandwidth between GPUs within the NVL72 domain. The cabling topology within the rack changes accordingly, requiring updated runbooks and cable routing plans.
InfiniBand XDR replaces NDR for inter-rack fabric connectivity, pushing network speeds to 800GbE. This requires updated fiber specifications and switch infrastructure, though the physical cabling approach (MPO/MTP trunks) remains the same.
From a deployment labor perspective, the assembly process is nearly identical. Teams experienced with GB200 NVL72 can transition to GB300 with updated runbook training. The liquid cooling, power distribution, and rack integration workflows carry over directly.