Google unveiled ‘Ironwood’, a man-made intelligence (AI) accelerator, which improved the reasoning speed by 10 times in comparison with the previous generation. The strategy is to maximise cost efficiency through optimized design chips and to reply to the demand for AI -based AI.
Google Cloud introduced Ionwood, the following -generation tensor processing device (TPU) on the annual technology conference ‘Cloud Next’ held in Las Vegas, USA on the ninth (local time). Ionwood is Google’s seventh generation TPU, which has improved greater than 10 times in comparison with the previous generation Trillium.
Particularly, it’s the primary Google TPU specialized within the AI ​​model reasoning, providing as much as 42.5 exaflops within the pod with 9216 chips.
Power in comparison with power has doubled, and the capability of HBM is six times larger than before. The chip will probably be provided to Google Cloud customers in the shape of 256 or 9216 chip configuration to Google Cloud customers at the top of this yr.
“Ionwood is essentially the most powerful and energy -efficient TPU developed by Google,” said Amin Bach Dart, vp of Google Cloud.
This announcement is in keeping with the flow of the AI ​​industry’s central axis from modeling to reasoning, and Google is trying to distinguish itself from competitors through chips specialized in high -performance reasoning. Particularly, the possession of each its own design hardware and the AI ​​model is a robust competitiveness of Google Cloud.
The truth is, Google’s high -performance TPU is combined with the AI ​​model of the Deep Mind to create a high synergy and is the muse for providing a stable AI cloud service.
Meanwhile, Startup SSI, founded by Open AI co -founder Ilia Sutsu Kever, announced that it had chosen Google Cloud as its principal computing supplier.
Considering that Sushker is from Google Brain, he’s foreseen. Last yr, World Labs, founded by Professor Pay Pay Stanford University, a former senior scientist of Google, selected Google Cloud.
By Park Chan, reporter cpark@aitimes.com
