It seems that Google has made important progress in the development of chips for data centers. Recent reports indicate that the company will be using its new proprietary processors in data centers by the 2025 timeframe.
This is not the first time that Google has succeeded in the R&D process of a dedicated chip, having already produced an ASIC-based design for servers and a system-on-a-chip (SoC) for smartphones. The Internet search giant had started the development process of Tensor Processing Unit (TPU) since 2015. TPU was built on the basis of ASIC and the purpose of its design was to increase the speed of machine learning, artificial intelligence and neural networks. This processor was also used in custom SSDs, network power supplies and NICs. TPU was placed on the TensorFlow platform to perform artificial intelligence processing, but Google used CPUs and graphics made by other companies for a number of processes and processing tasks. TPU has now reached its fourth generation, and now Google plans to use its own processors in the server space.
Google has also developed a more complete system-on-a-chip called Tensor G2 for smartphones, which uses ARM’s Cortex cores, Mali G710 graphics with a custom TPU, image signal processor (ISP), security core, and cache memory. 5 nm has been produced. The development of TPU and this chip in recent years has probably had a huge impact on the better development of Google’s new server chip development plans.
Apparently, Google is trying hard to start a more serious competition with Amazon in the business cloud server segment. In 2018, the online retail giant unveiled the ARM-based AWS Gravitation processor and now uses its third generation in its data centers. Compared to similar examples made by other companies, this processor has brought very good performance and efficiency to Amazon servers.
to report TomsHardwareGoogle’s server chip research and development teams are currently working on the construction of two 5nm chips based on ARM architecture. The mentioned system-on-chip codenamed CyPress is one of Google’s internal designs. Meanwhile, there is another scheme codenamed Maple, which is referred to as the second scheme and is apparently the main priority with CyPress.
Considering that the mass production of dedicated Google server chips will probably start from 2024, it is possible that the company will use them in its data centers from 2025, which is not good news for companies like AMD and Intel.