Google AI Research Showcase The 4th-Gen TPU Chips for Machine Learning Acceleration

Please Don't Forget To Join Our ML Subreddit

Google introduces the preview of its latest machine-learning clusters at its I/O conference, aiming for nine exaflops of peak performance while using 90% carbon-free electricity. It will be the largest publicly accessible machine learning hub in the world.

The TPU V4 Pod is at the heart of the new clusters. These tensor processing units were first unveiled at Google I/O last year, and AI teams from Meta, LG, and Salesforce have already gotten their hands on them. The V4 TPUs allow researchers to utilize any framework they want, including Tensorflow, JAX, and PyTorch and have already helped Google Research make breakthroughs in fields like language understanding, computer vision, and speech recognition.

The clusters, which are based in Google’s Oklahoma data center, are intended to chew through data in natural language processing, computer vision algorithms, and recommendation systems.

(Image credit: Google)

Slices of access to the clusters are available, ranging from four chips (one TPU VM) to thousands. Three-dimensional torus networks are used in portions with at least 64 fragments, offering more bandwidth for collective communication activities. The V4 chips can also access twice as much memory as the previous generation — 32GB instead of 16GB — and have double the acceleration speed while training large-scale models.

The release of Cloud TPU v4 is a significant milestone for both Google Research and our TRC program. 

Google’s commitment to sustainability means that, since 2017, the firm has been matching the energy use of its data centers with renewable energy purchases, intending to run the company entirely on renewable energy by 2030. The V4 TPU uses less fuel than previous versions, producing three times the FLOPS per Watt of the V3 chip.

All Google AI Cloud users will access Cloud TPU v4 Pods in evaluation (on-demand), preemptible, and committed usage discount (CUD) choices.

References:

  • https://www.tomshardware.com/news/google-unveils-next-gen-ai-chips
  • https://techcrunch.com/2022/05/11/google-launches-a-9-exaflop-cluster-of-cloud-tpu-v4-pods-into-public-preview/
  • https://www.hpcwire.com/2021/05/20/google-launches-tpu-v4-ai-chips/