HomeGoogle Cloud introduced a fifth generation of tensor processors for AI training

Google Cloud introduced a fifth generation of tensor processors for AI training

Published on


At the next annual Cloud Next conference, Google announced the release of the fifth generation of its tensor processing units (TPUs) for the training of neural networks. Google announced the previous, fourth version of these processors in 2021, but they could only be made available by 2022.

Photo Source: Frederic Lardinois/TechCrunch

The company observes that, when he created the new chip, its efficiency was a focus on the new version of that chip. A compared to the previous generation, the new processor promises 2x improvement in performance per dollar for training a wide language model (LLM) and generative neural networks, as well as 2.5x improvement of performance at the same time as the inference system.

This is the most cost-effective and affordable cloud computing platform available today, said Mark Lohmeyer, vice president and general manager of compute and ML infrastructure at Google Cloud, in a press conference.

Image source: Google.

Gleichzeitig emphasized that the company didn’t skimp on the technical characteristics of the TPU v5e under the assumption of profitability. Clusters can consist of up to 256 TPU v5e chips, united by high-speed obv transmission and runput of more than 400 Tbit/s. To calculate a platform such as this, there are 100 Pops in INT8.

We give our customers the ability to scale their AI models beyond the physical boundaries of a single TPU module or a single TPU cluster. In other words, a large AI workload can now be distributed across multiple physical TPU clusters, ranging from 10 to 100 percent of the chips. Similarly, the task is going to be handled cost-effectively. As such, leveraging cloud GPUs and cloud TPUs, we give our customers more options, flexibility and capabilities to meet the broad and broad set of AI workloads that we see that continue to proliferate, says he.

Google also announced that it would allow the A3 virtual clusters powered by custom Nvidia H100 GPUs to be available next month.

If there’s something you’re missing, pick it up with the mouse and press CTRL + ENTER.

Latest articles

More like this

A new teaser for GTA 6? – Rockstar is a subject of conversation

The release of Grand Theft Auto 6 is arguably one of the...

Zhou Guanyu watched how he played CS2 alongside his teammates in another gaming crossover

A post on Instagram revealed that the Chinese driver Zhou Guanyu was...