Technology•

Nvidia's Vera Rubin Rack Trains a 671-Billion-Parameter Model in 4.14 Minutes in Its First MLPerf Training Round

Nvidia's Vera Rubin NVL72 led all seven MLPerf Training v6.1 benchmarks in its debut, with per-GPU performance up to twice that of the Blackwell Ultra system it replaces.

Written By :MINRK
Nvidia's Vera Rubin Rack Trains a 671-Billion-Parameter Model in 4.14 Minutes in Its First MLPerf Training Round

Photo by Cory Woodward on Unsplash

1. Nvidia Vera Rubin leads every test in its benchmark debut

Nvidia's Vera Rubin NVL72 rack-scale system made its first appearance in the MLPerf Training benchmark suite on November 16, 2026, leading all seven tests in version 6.1 and recording the fastest training time on six of them. The company reported per-GPU performance of up to two times that of the GB300 NVL72, the Blackwell Ultra system that Vera Rubin succeeds.

The results are the first public, standardized measurements of how the new platform handles the work of training AI models. Until now, the case for Vera Rubin as a training machine rested on Nvidia's own projections. The benchmark round puts numbers produced under a common set of rules next to those projections and allows a direct comparison with the previous generation of hardware.

The gains were spread across several kinds of workload, including large mixture-of-experts models, dense large language models and advanced recommendation systems. The breadth matters because AI developers do not train one type of model, and hardware that excels at a single task can disappoint on another.

2. What is Nvidia Vera Rubin NVL72?

Vera Rubin NVL72 is Nvidia's newest rack-scale AI system, a single cabinet that holds 72 graphics processing units (GPUs) and 36 central processing units (CPUs) linked by the sixth generation of the company's NVLink interconnect. It replaces the GB300 NVL72, which is built on the Blackwell Ultra generation of chips.

A rack-scale system is sold and operated as one integrated unit and not as a collection of separate servers. The processors, the networking that joins them and the supporting hardware are designed together so that the whole rack behaves like one very large computer. Nvidia has kept the same basic layout for Vera Rubin that it used before, with the 72-GPU, 36-CPU arrangement carried over from the earlier design.

NVLink is the high-speed connection that lets the GPUs inside a rack exchange data with one another. Training a large model means constantly passing information between processors, so the speed of that link has a direct effect on how much useful work each chip can do. A faster interconnect reduces the time GPUs spend waiting on one another.

The comparison point throughout the results is the GB300 NVL72. That system had already been heavily tuned for these benchmarks, which makes it a demanding baseline. A new platform that doubles the per-GPU output of a mature, well-optimized predecessor is showing an improvement that comes from the hardware generation itself and not from the older system having been left unoptimized.

3. What MLPerf Training measures

MLPerf is a family of benchmarks that hardware makers, cloud providers and server builders use to submit results under shared rules. The training suite asks a simple question of each system: how long does it take to train a given model to a defined level of quality? A lower time is a better result, and because every submitter must reach the same quality target, the figures can be compared across vendors and generations.

The suite is made up of several tests, each built around a different model and task. Version 6.1 contains seven of them. Leading all seven, as Vera Rubin did, means the platform was ahead on every workload in the round. Posting the fastest training time on six means that in all but one case, a Vera Rubin submission reached the quality target sooner than any other entry.

Raw training time depends heavily on how many processors a submitter uses, since a larger cluster will usually finish sooner. For that reason, per-GPU performance is the figure buyers tend to watch most closely. It indicates how much work each individual chip contributes, which translates into how many chips, and how much floor space and power, a customer needs to get a job done.

4. The DeepSeek-V3 result: 4.14 minutes on 1,024 GPUs

The most striking number in the round came from the DeepSeek-V3 671B benchmark, which is built around a model with 671 billion parameters. A cluster of 1,024 Rubin GPUs completed the training run in 4.14 minutes. At the same GPU count, that works out to roughly double the per-GPU training speed of the GB300 NVL72.

DeepSeek-V3 is a mixture-of-experts model, a design in which the network is divided into specialized sub-networks and only some of them are activated for any given input. The approach has become popular for very large models because it allows the total parameter count to grow without every parameter being used on every step. Models of this kind are demanding to train across many processors, which is why the benchmark is treated as a meaningful test of a system's networking as well as its chips.

Nvidia also reported how well the system holds up as it grows. When the DeepSeek-V3 run was scaled from 256 GPUs to 1,024 GPUs, Vera Rubin achieved about 97% scaling efficiency. In an ideal world, four times as many processors would deliver exactly four times the throughput. In practice, communication overhead eats into the gain as clusters get larger. A figure near 97% means almost none of the added hardware was wasted on that overhead.

Scaling efficiency is important because the largest AI developers do not train on a single rack. They join many racks into clusters, and the economics of doing so depend on each additional rack contributing nearly its full capacity. A system that scales poorly forces customers to buy more hardware than the headline per-GPU figure would suggest.

5. Gains across three kinds of model

The improvement was not confined to one benchmark. Nvidia said the gains showed up across large mixture-of-experts models, dense large language models and advanced recommendation systems, three categories that place different demands on hardware.

Dense large language models are the conventional architecture behind many chatbots and coding assistants. Every parameter in the network takes part in processing each input, which makes the workload heavy on raw computation and memory. Mixture-of-experts models, as with DeepSeek-V3, route each input through a subset of the network and stress the system's ability to coordinate work across processors.

Recommendation systems are a different business altogether. They are the models that decide which products, videos or posts to show a user, and they run at the core of large consumer internet services. They work with enormous tables of data about users and items, so they test how quickly a system can move and look up information as much as how quickly it can calculate.

A platform that leads in all three areas is suited to the range of customers Nvidia sells to, from AI laboratories building frontier language models to internet companies tuning the systems that drive their feeds and storefronts.

6. CoreWeave reports a 2.7x speedup

Independent confirmation of the generational jump came from CoreWeave, the cloud provider that rents Nvidia hardware to AI developers. The company submitted its own multi-rack Vera Rubin results and said that on the DeepSeek-V3 benchmark at 256 GPUs, the new system ran 2.7 times faster than its previous GB200 NVL72 setup.

The CoreWeave figure uses a different baseline from Nvidia's headline number. Nvidia's comparison of up to two times is measured against the GB300 NVL72, the Blackwell Ultra system. CoreWeave's comparison is against the GB200 NVL72, an earlier Blackwell configuration. The two results therefore describe different steps and should not be read as conflicting.

Submissions from cloud operators carry particular weight for customers, because they reflect hardware deployed in a commercial data center and not only in the manufacturer's own laboratory. A multi-rack result also shows the system working across the connections between racks, which is closer to how large customers will use it.

7. Why training speed matters beyond the benchmark table

Training time is a cost. Every hour a cluster spends on a training run is an hour of electricity, cooling and capital tied up in that job. Hardware that finishes the same work in half the time per GPU allows a developer to train more models with the same equipment, or the same number of models with fewer chips.

It also affects the pace of research. AI developers improve their models by running many experiments, and the number of experiments they can complete in a given period is limited by how long each one takes. Shorter training runs mean more iterations, which is one reason the largest AI companies compete for early access to each new hardware generation.

For investors, benchmark rounds serve as one of the few neutral checkpoints on a product cycle. Nvidia's revenue depends on customers deciding that each new platform is worth the cost of upgrading. Standardized results that show a clear per-GPU advantage over the system being replaced support the argument for making that move.

Power is an increasingly central part of that calculation. The amount of electricity a data center can draw is fixed by its connection to the grid, and operators cannot simply add racks once that ceiling is reached. Getting more training work out of each GPU is one of the few ways to raise the output of a facility whose power supply is already fully committed, which is why per-GPU figures of the kind reported in this round attract attention from the companies that build and finance AI data centers.

Benchmarks still have limits. They measure specific models under specific rules, and a customer's own workload may behave differently. The results show what the hardware can do on the tests in the suite, and buyers generally treat them as a starting point for their own evaluation.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY