Google is developing two custom chips for servers
Google has been using its internally developed TPU (Tensor Processing Unit) since 2015, which is a processor for neural network training, mainly used for deep learning and AI computing, etc., and is now in its fourth generation. Google plugged it into the TensorFlow framework while continuing to use third-party CPUs and GPUs for some other key processes and processing tasks.
Google’s competitor in cloud services, Amazon, launched the Graviton series processors using the Arm architecture in 2018, which has advanced to the third generation and has impressive performance and energy efficiency performance on servers.
According to The Information, Google hopes to go further on the road of self-developed chips and use self-developed chips more widely in data center servers like Amazon. Currently, its chip development team is developing two 5nm chips based on the Arm architecture.
One of the SoCs, called “Cypress,” was designed by Google’s team in Israel. The other, code-named “Maple,” is based on Marvell’s design. Responsible for the design of these two chips is Uri Frank, an industry veteran with 25 years of experience in CPU design at Intel. Google hired Uri Frank in March 2021 to lead a new department dedicated to developing custom chips. It is understood that Uri Frank regards “Cypress” as Plan A, while “Maple” is Plan B. Both chips will be manufactured by TSMC.
It is reported that Google’s custom server chips will be mass-produced in the second half of 2024 and deployed to data centers as early as 2025.
