Published :8/18/2020 2:16:46 AM
Click Count:2103
JP Morgan analysts pointed out that Google's fourth-generation TPU will be manufactured using the 7nm process, and the fifth-generation TPU has already been co-designed, and it is expected to use a 5nm transistor design.
Google claims that the number of floating-point operations per second provided by its fourth-generation TPU is more than twice that of the third-generation TPU. The third-generation TPU’s matrix multiplication per second is equivalent to 1 trillion floating-point operations.
In terms of memory bandwidth, there has also been a significant increase. The speed at which the chip obtains data from the memory for processing and the ability to perform specialized calculations has improved. Google said that overall, the performance of the fourth-generation TPU was 2.7 times higher than that of the third-generation TPU in last year's MLPerf benchmark test.
TPU is a special neural network chip launched by Google in 2015. It is built to optimize its own TensorFlow machine learning framework. Unlike GPU, Google TPU is an ASIC chip solution, which is a specially customized chip with extremely high research and development costs.
Naveen Kumar of Google’s AI department said that it would take more than 3 weeks to train AI models with the most advanced hardware accelerators at the time in 2015. Now using Google’s latest TPU supercomputer to train the same AI model, the speed is almost 10 times faster than that of the year. TPU is used to process multiple Google services, such as Google Search, Google Photos, Google Translate, Google Assistant, Gmail, etc.
In addition to Google, Broadcom also cooperates with other customers including Facebook, Microsoft (Microsoft) and AT&T to build special application chips (ASIC). With the support of many major customers, Broadcom's ASIC business has grown rapidly in revenue in the past four years. The US$50 million in 2016 surged 14 times and is expected to reach US$750 million in 2020.
