Daily Agentic, the front page of the machine age
Today
Hardware / Nvidiathe 17th of November 2026, 09:30 UTC

Nvidia Vera Rubin trains AI 2 times faster

Nvidia said its new Vera Rubin chips trained AI up to two times faster per chip than its last top system. The results came from MLPerf, a shared industry speed test, on the 16th of November 2026.

Nvidia Vera Rubin speed results

Nvidia said its new Vera Rubin chips trained AI up to two times faster per chip than its last top system. The results came from MLPerf, a shared industry speed test, on the 16th of November 2026.

MLPerf is run by MLCommons, a group of tech firms and universities. This round, called Training v6.1, was the first for Vera Rubin.

Training speed

Vera Rubin NVL72 is a rack of linked chips. It beat the older GB300 NVL72 by up to two times per chip, Nvidia said.

Nvidia was the only maker with results on every test this round. Twenty partner groups also sent results on Nvidia chips.

A test with more chips

On a test that trains DeepSeek-V3, a large open model, Nvidia grew from 256 chips to 1,024 chips. Training time fell from 16 minutes to 4.14 minutes, Nvidia's technical blog said.

Agent work

In a newer test called MLPerf Endpoints, Vera Rubin gave over 10 times more output per watt of power than GB300. That test ran Kimi K3, an open model built for AI agents.

Nvidia said that Kimi result is provisional and still awaits checks by MLCommons.

Below we break down the scaling

Four times the chips made training about 3.9 times faster, by our arithmetic. That fits the 97 percent efficiency Nvidia gave.

Vera Rubin: DeepSeek training time sixteen minutes on two hundred fifty six chips and four point one four minutes on one thousand twenty four chipsVera Rubin: DeepSeek training time sixteen minutes on two hundred fifty six chips and four point one four minutes on one thousand twenty four chips

What is confirmed and what is not

The training results are posted by MLCommons. The 10 times power figure is Nvidia's own and still provisional.

Key facts

Nvidia shared the MLPerf results on the 16th of November 2026.

Vera Rubin was up to two times faster per chip than GB300.

DeepSeek-V3 training fell from 16 minutes to 4.14 minutes.

Nvidia reports its quarterly results on the 17th of November.

Source: NVIDIA Blog: NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance