NVIDIA did not double the number of GPUs in each tray when it moved from GB200 NVL72 to GB300 NVL72. Both DGX rack systems use the same high-level layout: 18 compute trays, each with two Grace CPUs and four GPUs, connected through nine NVLink switch trays. The upgrade is inside that footprint—Blackwell Ultra GPUs, more HBM3e and faster scale-out network adapters.
That distinction matters to buyers. GB300 is not a denser 72-GPU rack; it is a newer 72-GPU rack with a larger memory pool, twice the per-port compute-network rate and a somewhat higher facility power requirement.
Cover: AI-generated conceptual illustration, not a photograph or product rendering.
Key differences at a glance
| Specification | DGX GB200 NVL72 | DGX GB300 NVL72 |
|---|---|---|
| GPU generation | 72 Blackwell GPUs | 72 Blackwell Ultra GPUs |
| Grace CPUs | 36 | 36 |
| Compute trays | 18; four GPUs and two CPUs per tray | 18; four GPUs and two CPUs per tray |
| GPU memory | Up to 13.4 TB HBM3e | 20 TB HBM3e |
| Total fast memory | 30.2 TB | 37 TB |
| NVLink bandwidth | 130 TB/s | 130 TB/s |
| Compute-network adapters per tray | Four ConnectX-7, single-port 400G | Four ConnectX-8, single-port 800G |
| Rack power | Approximately 120 kW | Up to 142 kW |
The memory and product totals come from NVIDIA’s DGX GB200 specifications and GB300 NVL72 specifications. NVIDIA’s combined DGX GB rack hardware guide provides the tray, networking, cooling and power details.
The rack layout stays the same
NVIDIA’s hardware guide covers both DGX GB200 and DGX GB300. It specifies 18 one-rack-unit compute trays, nine one-rack-unit NVLink switch trays, two top-of-rack management switches and multiple power shelves. Each compute tray contains two Grace CPUs and four Blackwell-family GPUs.
That yields the same rack totals in both generations: 36 Grace CPUs and 72 GPUs. A GB200 Grace Blackwell Superchip combines one Grace CPU with two Blackwell GPUs. Each compute tray contains two of those CPU/GPU groupings, which produces the documented tray total of two CPUs and four GPUs.
The physical commonality has a practical implication: the GB300 comparison should start with component capability and facility headroom, not an assumed change in tray count. A buyer still has 18 compute nodes and one 72-GPU NVLink domain to deploy and manage.
GB300’s clearest gain is memory
DGX GB200 provides up to 13.4 TB of HBM3e across the rack. GB300 raises that to 20 TB while keeping the GPU count at 72. That is about 49% more aggregate GPU memory, calculated from NVIDIA’s published totals.
Total fast memory also rises from 30.2 TB to 37 TB. The larger HBM pool is relevant to long-context inference and reasoning workloads because it can hold larger model states and KV caches before an operator has to partition work differently or spill data outside GPU memory. It does not, by itself, guarantee a proportional application-speed gain; that depends on the model, precision, batching and software stack.
The fifth-generation NVLink fabric remains a 130 TB/s rack-scale interconnect in both products. GB300 therefore changes the compute and memory resources attached to the fabric, not the headline aggregate NVLink bandwidth.
Scale-out networking moves from 400G to 800G
The most concrete infrastructure change is at each compute tray’s cluster network. NVIDIA lists four single-port ConnectX-7 400G OSFP adapters for GB200 and four single-port ConnectX-8 800G OSFP adapters for GB300.
That doubles the nameplate rate of each compute-network port. It does not mean an application automatically doubles its multi-rack throughput: switch fabrics, optics, congestion control, collective-communication patterns and storage paths still determine delivered performance. It does mean a GB300 deployment has to be designed around a different adapter generation and potentially a higher-bandwidth scale-out fabric.
BlueField DPUs, a BMC management path and local NVMe storage remain part of the rack architecture. For a closer look at the GB300 data paths, see our GB300 NVL72 networking guide.
Power and cooling need more headroom
NVIDIA’s rack guide describes approximately 120 kW of rack power for the DGX GB platform. Its current GB300 enterprise reference architecture gives a more specific ceiling of up to 142 kW for a full GB300 rack. Those are planning figures, not a promise that every workload will continuously draw the maximum.
At the published limits, GB300 adds up to 22 kW over the approximately 120 kW GB200 reference—about an 18% increase. Operators should validate the exact ordered configuration and vendor integration before sizing power distribution, backup capacity or cooling.
The cooling description also needs precision. CPUs and GPUs are cooled by liquid through cold plates and rack manifolds, while networking, storage and other components remain air cooled. Calling either rack “fully liquid-cooled” hides that residual air-cooling requirement. For broader facility context, our GB300 power and cooling overview covers rack-side implications in more detail.
Operations are similar, but releases are product-specific
NVIDIA’s common control guide uses Base Command Manager to manage rack components and Mission Control to orchestrate power operations. The documented startup order is to establish the NVLink fabric by bringing up the NVSwitch trays before the compute trays. Firmware is released as a rack-level recipe rather than a collection of arbitrary component updates.
That does not make GB200 and GB300 firmware interchangeable. NVIDIA’s September 2026 GB300 release notes state that system software release 1.0.12 is only for DGX GB300 NVL72. They also require every component in the rack—including compute nodes and NVSwitch—to be on release 1.0.5 or later for NVLink Recovery stability. The operational lesson is straightforward: preserve the common rack workflow, but qualify software and firmware against the exact product generation.
Bottom line
GB300 NVL72 is best understood as an in-place generational upgrade to NVIDIA’s 72-GPU rack pattern. It keeps the 18-tray, 36-CPU, 72-GPU layout and 130 TB/s NVLink fabric. The material changes are 20 TB of GPU memory instead of 13.4 TB, ConnectX-8 800G adapters instead of ConnectX-7 400G adapters, Blackwell Ultra GPUs and a full-rack power requirement that can reach 142 kW.
For procurement, the decision is less “how many more GPUs fit in a rack?” and more “does the workload benefit enough from the larger memory pool and faster scale-out network to justify the newer platform and added facility headroom?”
Sources
This article was researched and fact-checked against the following sources:
- Hardware — NVIDIA DGX GB Rack Scale Systems User Guide (docs.nvidia.com)
- DGX GB200: AI Infrastructure for State-of-the-Art AI Models (nvidia.com)
- NVIDIA GB300 NVL72 (nvidia.com)
- System Hardware and Components — NVIDIA NVL72 AI Factory (docs.nvidia.com)
- System Control and Configuration — NVIDIA DGX GB Rack Scale Systems User Guide (docs.nvidia.com)
- Overview — NVIDIA DGX GB300 NVL72 Release Notes (docs.nvidia.com)
