Google goes closer to the v. output of its tensor processing unit. Most notably, Nvidia is the head of jut.

Google is proud of the fourth generation of its Tensor Processing Unit (TPU). The TPU v4 was announced back in 2021 and finally appeared in the wild in late 2022. The AI ​​accelerator is intended to be used primarily in machine learning clusters in Google’s cloud. So the chip isn’t just sitting on the shelves, giving Google a chance to unpack with architectural details every now and then. This week’s paper from Cornell University claims that the TPU v4 is at least 1.2x faster and 1.3x more efficient than the Nvidia A100.

Nvidia founder and CEO Jensen Huang is not easily persuaded. In a blog post, Nvidia points out that the A100 is already three years old. The newer Nvidia H100 is said to be four times more powerful than the A100. That’s true, but Nvidia has lost sight of the fact that Google has also been integrating the TPU v4 into its own systems since 2020. Both accelerators are based on the same 7 nm technology.

Correct comparison?

A comparison with the Nvidia H100 was not possible during the study in 2022, it sounds, since the chip was announced but actually not available anywhere. The first concrete H100 systems did not appear until the end of the year.

The Nvidia H100 is now neatly integrated into existing cloud systems. That proves Google and Nvidia right. In the implementation study, Google actually made a relevant comparison and the TPU v4 performed well in the tests. Unfortunately for the company, the release of the research was delayed until April and the Nvidia H100 is ready to use today. Easter is just around the corner, but the newspaper still reads like figs afterwards.

The main thing to remember for those not involved is that both Google and Nvidia make good AI accelerators, and neither will they settle for a silver medal.