Google announced a new generation of AI supercomputer, saying it is faster and more energy-efficient than NVIDIA A100

News on April 5th, local time on Tuesday, Alphabet’s Google disclosed some new details, showing the super machine used to train artificial intelligence models. computer, saying it is faster and more energy-efficient than systems based on Nvidia's A100 chip.

Google has designed its own custom chips, called Tensor Processing Units (TPUs), and uses these chips in more than 90% of its artificial intelligence training work. This process trains the data through a model to improve its usefulness in tasks like human text response or image generation.

Currently, Google TPU has entered the fourth generation. Google released a scientific paper on Tuesday detailing how it uses custom optical switches it developed to connect more than 4,000 chips into a supercomputer.

For companies building AI supercomputers, improving these connections has become a key point in the competition, as so-called large-scale language models continue to grow in size and cannot be stored on a single chip. Powering technologies like Google’s Bard or OpenAI’s ChatGPT.

Such models must be distributed across thousands of chips, which then work together over weeks or longer to train the model. Google's PaLM model, the largest publicly disclosed language model to date, was trained for 50 days spread across two 4,000-chip supercomputers.

Google says its supercomputers can easily and dynamically reconfigure connections between chips, helping avoid failures and optimize performance.

"Circuit switching allows us to easily bypass failed components," Google Senior Researcher Norm Jouppi and Google Distinguished Engineer David Patterson wrote in a blog post about the system. "This flexibility even allows us to change the topology of the supercomputer interconnection network to accelerate the performance of ML (machine learning) models."

While Google has not disclosed details of its supercomputers until now, It has been operating out of a data center in Mayes County, Oklahoma since 2020. Google said startup Midjourney used the system to train its model, which can generate new images after being fed a few words of text.

Google said in the paper that its supercomputer is 1.7 times faster and 1.9 times more energy efficient than a system based on Nvidia's A100 chip, compared with systems of the same size. It is understood that the A100 chip was launched earlier than the fourth generation TPU.

Google said it did not compare the fourth-generation TPU with Nvidia’s current flagship H100 chip, because the H100 was launched after Google’s application of this chip and uses newer technology.

Google hinted that it is developing a new TPU that will compete with Nvidia H100, but did not provide details. Jouppi told the media that Google has "ample reserves of future chips."

The above is the detailed content of Google announced a new generation of AI supercomputer, saying it is faster and more energy-efficient than NVIDIA A100. For more information, please follow other related articles on the PHP Chinese website!