A specification sheet can change without users immediately noticing their applications slowing down. The case of Vera Rubin, Nvidia’s platform for artificial intelligence data centres, shows why memory capacity, communication bandwidth and the results achieved by software need to be read separately.
The manufacturer’s current documentation lists 288 GB of HBM4 memory and a bandwidth of 19.2 TB/s for a Rubin GPU. It also gives 3 TB/s for NVLink, the GPU-to-GPU communication system. These are two different specifications: they cannot be added together to give an “AI speed”.
What has been revised, according to published comparisons
In an article dated 28 October 2026, Frandroid reports that the previously announced figures were 22 TB/s for memory and 3.6 TB/s for NVLink, per GPU. The differences amount to approximately 12.7% and 16.7%, respectively. Nvidia’s current page allows readers to check the figures currently listed; on its own, it does not preserve the history of all its changes.
This distinction avoids two shortcuts. The first would be to present an earlier promise as the current specification. The second would be to turn a difference between two bandwidth figures directly into a measured slowdown in a service. The reported change is worth noting, but assessing its effect on a task requires a different kind of evidence.
Capacity and bandwidth answer different questions
The 288 GB indicates an amount of memory. The 19.2 TB/s describes a bandwidth: an amount of data that can be transferred per unit of time under the conditions associated with the specification. One figure tells us about available space; the other, about the flow of data.
A simple comparison helps explain the difference. Enlarging a warehouse is not enough to speed up the loading of its goods. Widening its access road does not create more space inside, either. A computer system’s memory is obviously not a warehouse, but the image helps avoid confusing the two dimensions.
So when examining a configuration, two separate questions need to be asked: does the task have the capacity it needs, and how quickly can it access the data? An impressive figure on the second line does not automatically answer the first question.
NVLink is for communication between processors
Nvidia’s description of NVLink presents it as a direct interconnect between GPUs. Its sixth generation is specified at 3 TB/s per GPU, with bidirectional bandwidth. The manufacturer says that the NVL72 configuration connects 72 GPUs and lists an aggregate bandwidth of 216 TB/s.
The word “aggregate” matters. A figure calculated for a group does not become the individual bandwidth of each of its components. Likewise, bandwidth between GPUs should not be confused with HBM4 memory bandwidth. The two data pathways serve distinct exchanges.
When an application distributes work across multiple processors, coordinating them is part of the problem to be solved. Having many GPUs therefore does not tell us everything about how they will be able to work together. That is precisely why a platform specification sheet has several communication entries rather than a single indicator.
Why the percentages do not predict an equivalent slowdown
The calculations of 12.7% and 16.7% compare announced specifications. They measure neither the time needed to train a model, nor an assistant’s response time, nor the actual cost of a request.
A purely hypothetical example helps illustrate the difference. If a task is waiting mainly for data to be transferred, changing that transfer can have a substantial effect on its duration. If it spends most of its time on another operation, the same change will not necessarily have the same effect. Without measuring the task as a whole, we do not know what proportion of its time is affected.
The same work would therefore need to be compared under clearly stated conditions: configuration, software, model, data volume and processing parameters. An isolated percentage may catch the eye; it cannot replace this testing protocol.
What businesses should look at before comparing two offers
The first point is scope. Are we comparing a GPU, a card, a server or a complete rack? Next come the units: GB for capacity, TB/s for bandwidth, operations per second for another specification. Moving from one entry to another without keeping these distinctions in mind can produce a misleading comparison.
The useful result also depends on the question being asked. For an interactive service, waiting time matters; for high-volume production, the number of tasks completed in a given period may be more relevant. These criteria should be stated before choosing the figure that seems most favourable.
Nvidia accompanies several of the performance figures on its page with comparison conditions and caveats. Reading these conditions is part of assessing the product. A commercial claim remains information attributed to the manufacturer until an independent test allows its significance to be assessed.
The most useful way to read Vera Rubin is therefore to retain the current figures, identify precisely what they describe and ask for results on the intended use. The specification sheet provides reference points; the decision depends on the work actually done.
Illustrative photograph of consumer computer components; it does not show the Vera Rubin platform.
