Key takeaways

  • Vera Rubin NVL72 and GB300 NVL72 each connect 72 GPUs and 36 CPUs, but Rubin moves to Vera CPUs, HBM4 and sixth-generation NVLink.
  • NVIDIA lists 20.7 TB of HBM4 for Rubin versus 20 TB of HBM3E for GB300. The company publishes more than one Rubin bandwidth figure, so buyers should record the source and revision rather than silently mixing them.
  • GB300 has 130 TB/s of rack-level NVLink bandwidth. NVIDIA currently shows 216 TB/s for Rubin in its product table and 260 TB/s in newer technical material.
  • NVIDIA does not publish a rack-power figure in either product specification. The widely repeated 190–230 kW Rubin and 132–142 kW GB300 ranges are third-party planning estimates, not vendor specifications.
  • Both platforms are liquid-cooled. Supermicro says its Rubin cooling design can capture more than 98% of heat and is sized for 227 kW per rack; that cooling capacity is not proof of steady electrical draw.

What is actually being compared

NVIDIA Vera Rubin NVL72 and NVIDIA GB300 NVL72 share the same rack-scale outline: 72 GPUs, 36 CPUs and one all-to-all NVLink domain. The similarity ends at the component generation.

GB300 pairs Blackwell Ultra GPUs with Grace CPUs and fifth-generation NVLink. Vera Rubin pairs Rubin GPUs with Vera CPUs and sixth-generation NVLink. That makes Rubin a change in compute, memory, host processing and scale-up fabric—not simply a faster GPU dropped into the GB300 rack.

This comparison uses NVIDIA product pages and technical posts as the specification baseline. Where those official pages disagree, both figures are shown. That matters more than false precision for teams budgeting power, cooling and network capacity.

Rubin HBM4 versus GB300 HBM3E

The Rubin product table lists 20.7 TB of HBM4 across the rack and 1,400 TB/s of aggregate GPU-memory bandwidth. Its per-GPU column lists 288 GB and 19.2 TB/s.

NVIDIA technical material also describes each Rubin GPU as reaching 22 TB/s, and the DGX Vera Rubin page has displayed an “up to 1,580 TB/s” rack figure. Those values are directionally consistent with a later or peak configuration, but they are not numerically identical to the main product table. Infrastructure plans should therefore cite the exact NVIDIA document and retrieval date instead of treating 1,400 and 1,580 TB/s as interchangeable.

GB300 uses HBM3E, not HBM4. NVIDIA lists 20 TB of GPU memory and up to 576 TB/s of rack bandwidth. Its Blackwell Ultra technical overview describes up to 288 GB of HBM3E per GPU. Rubin keeps roughly the same rack memory capacity while changing the memory generation and substantially increasing published bandwidth.

Published metricVera Rubin NVL72GB300 NVL72
GPUs / CPUs72 Rubin / 36 Vera72 Blackwell Ultra / 36 Grace
GPU memory20.7 TB HBM420 TB HBM3E
GPU-memory bandwidth1,400 TB/s in product table; up to 1,580 TB/s elsewhereUp to 576 TB/s
Rack NVLink bandwidth216 TB/s in product table; 260 TB/s in technical material130 TB/s
CoolingLiquid-cooledFully liquid-cooled
Official rack power figureNot listedNot listed

The GB300 specification gives 130 TB/s of rack-level NVLink bandwidth. NVIDIA describes fifth-generation NVLink at 1.8 TB/s per Blackwell Ultra GPU.

For Rubin, the current product table lists 216 TB/s per rack and 3 TB/s per GPU. A newer NVIDIA NVLink technical post states 3.6 TB/s per GPU and 260 TB/s across the 72-GPU rack. The arithmetic in each pair is internally consistent: 72 × 3 equals 216, while 72 × 3.6 is about 259.2.

The safe conclusion is not that one page must be wrong. NVIDIA labels platform specifications as subject to change, and product pages can describe a qualified configuration while architecture posts describe an “up to” capability. Procurement documents should name the expected configuration and require the vendor to confirm the applicable NVLink number.

Either official Rubin figure is a major step above GB300. What cannot be inferred from fabric bandwidth alone is application speedup: model topology, collective patterns, software and scale-out networking still determine how much of the rack fabric a workload uses.

Rack power: estimates are not specifications

The most repeated comparison says Vera Rubin draws 190–230 kW and GB300 draws 132–142 kW. Those ranges appear in ModulEdge analysis, which attributes the Rubin range to supply-chain analysis. NVIDIA does not list either range in the product specification pages cited above.

That distinction changes how the numbers should be used. A reported planning envelope can be useful for early facility design, but it should not be described as nameplate draw, typical consumption or a guaranteed maximum. Actual provisioning also depends on the rack vendor, power profile, workload and site design.

Supermicro says its DLC-2 design for Vera Rubin supports more than 98% heat capture and is sized for 227 kW per rack. The 227 kW number describes the cooling system capacity. It does not establish that every Rubin rack continuously consumes 227 kW.

GB300 is also a fully liquid-cooled rack. NVIDIA further documents programmable power caps, integrated energy storage and controlled ramp behavior for GB300, all intended to smooth what the grid sees during workload transitions. Cooling method and power smoothing are separate questions from total rack demand.

What NVIDIA means by better performance per watt

NVIDIA says Vera Rubin NVL72 can deliver up to four times better training performance per watt and up to ten times better inference performance per watt than Blackwell. Those are vendor “up to” comparisons for selected platform workloads, not a promise that every model will produce ten times as many tokens for the same electricity.

For capacity planning, the useful inputs are workload-specific throughput, service-level objective, measured wall power and cooling overhead. Peak FLOPS and aggregate bandwidth explain why Rubin may improve efficiency, but neither replaces a benchmark that matches the intended model and serving pattern.

Practical planning checklist

  1. Freeze the document revision. Record whether a design uses the 216 or 260 TB/s Rubin NVLink figure and the 1,400 or 1,580 TB/s memory-bandwidth figure.
  2. Separate electrical draw from cooling capacity. Do not turn Supermicro’s 227 kW cooling rating into an NVIDIA rack-power specification.
  3. Ask for the rack-vendor envelope. Confirm nameplate, capped and expected workload power for the exact OEM configuration.
  4. Model transitions, not only averages. Include ramp rate, power smoothing, coolant temperature and heat-rejection capacity.
  5. Benchmark the target workload. Treat NVIDIA performance-per-watt claims as a starting hypothesis until the model, precision, batch size and latency target match.

Bottom line

Vera Rubin NVL72 is a clear generational upgrade in memory technology and scale-up bandwidth over GB300 NVL72. The strongest official comparison is HBM4 versus HBM3E and sixth-generation versus fifth-generation NVLink.

Power is less settled. NVIDIA has not put a single rack-wattage number in the public product tables, and its own Rubin bandwidth pages currently show different qualified and “up to” values. The credible planning approach is to preserve those distinctions, obtain the exact OEM power envelope and benchmark the workload—rather than turning a supply-chain estimate into a specification.

Sources

This article was researched and fact-checked against the following sources: