Share This Article
Google’s new graphics chips will be used for AI training
The American corporation Google regularly improves its technologies for training artificial intelligence. All processes are carried out on supercomputers using their own designed graphics chips.
The representatives of Google stated that their elements of the new generation demonstrate better results in terms of speed and energy efficiency than the products of Nvidia. Using the TPU chip for learning language models shows better efficiency than other chips. The new development allows the processing of arrays of data in short periods.
In March 2023, Google released an article describing creating a supercomputer. According to it, the company’s developers combined more than 4 thousand chips in their work. Thanks to this, they managed to get a superpowerful machine capable of simultaneously processing a huge number of requests. There were specially made optical switches to connect the elements, which allowed synchronization of the individual computers.
Testing
The new fourth-generation TPU graphics chips from Google were tested for productivity, compared with chips A100, produced by Nvidia. It turned out that when training AI, TPU works 1.7 times faster. In addition, the efficiency of these chips is 1.9 times higher than that of the A100. However, the products entered the market almost simultaneously.
Most of the performance testing of the new product from Google was carried out while the supercomputer was working. According to the technology giant, its supercomputer can analyze the productivity of chips and their interaction with each other. This approach avoids problems with elements and prevents the AI’s learning performance from declining.
Google’s plans
Google plans to use next-generation chips to organize data center servers in 2025. The creation of systems based on TPU will be carried out gradually. Some work can begin in the near future. The chip of this model was first released in 2015, and since then, developers have been improving the technology regularly. Initially, the chip was created exclusively for the development of artificial intelligence but gradually began to be used in solid-state drives and network adapters.
According to some experts, using a new generation of TPU for data centers will allow Google to become more competitive in the segment of cloud servers, where at the moment, the leadership belongs to Amazon. The latter uses the third generation of Graviton processors for its data centers, which are based on Arm architecture. Amazon’s development is characterized by high performance and efficiency.
So far, Google has used TPU chips for machine learning but is actively working on using them for its data centers worldwide.