News on June 19, according to media reports in Taiwan, China, in order to develop the latest server-oriented AI chip, Google has approached MediaTek to cooperate, and plans to hand it over to TSMC’s 5nm process OEM, and plans to start mass production next year. Produce.
According to reports, sources revealed that this cooperation between Google and MediaTek will provide MediaTek with serializer and deserializer (SerDes) solutions and assist in integrating Google’s self-developed tensor processor (TPU) to help Google Create the latest server AI chip, which will be more powerful than CPU or GPU architecture.
The industry points out that many of Google’s current services are related to AI. It has invested in deep learning technology many years ago and found that using GPUs to perform AI calculations is very expensive. Therefore, Google decided to develop its own processing dedicated to deep learning. processor chip TPU, and launched the first generation TPU processor in 2015. Compared with the CPU and GPU at that time, its TPU can provide 30 to 100 times higher floating point computing performance per second, and the overall computing performance is It was improved by as much as 15 to 30 times, and even in terms of performance and power consumption, it was improved by nearly 30 to 80 times. The chip was then deployed in Google's own data center. Now Google's TPU chip has developed to the fourth generation.
This cooperation between Google and MediaTek is mainly based on MediaTek’s strength in the high-speed Serdes field. Specifically, high-speed SerDes is two independent devices, the transmitter (serializer) and the receiver (deserializer), which are important gates for cloud data traffic and analysis. As the scale of data processing in data centers continues to expand, various forms of networks are required, from servers to the top of rack routers, from rack to rack, and for long-distance transportation from data centers around the world or around the world. All require the use of high-speed SerDes chips. For example, data interconnection and transmission between servers usually uses optical fiber. The optical module needs to integrate a high-speed SerDes chip. To increase the bandwidth of the optical module, it is necessary to increase the number of SerDes channels or increase the transmission rate of a single SerDes channel. Therefore, high-speed SerDes chips have also become key components in data centers.
MediaTek announced in November 2019 that it would expand its ASIC services to cover 112G Long Range (LR) SerDes IP chips. MediaTek's 112G remote SerDes use silicon-proven 7nm FinFET process technology to enable data centers to quickly and efficiently process large amounts of specific types of data, thereby increasing computing speeds.
△MediaTek’s 7nm 56G PAM4 SerDes IP prototype chip
Sources revealed that MediaTek’s self-developed Serdes has already had a number of products in mass production and the technology is mature. The AI server chip that is being assisted by Google this time will be produced using TSMC’s 5nm process and is expected to be delivered to TSMC Tape Out before the end of this year. We are striving for mass production early next year.
Previously, MediaTek General Manager Chen Guanzhou has publicly stated that whether it is AI on the terminal or in the cloud, and how to meet the development trend of large language models (LLM), these are topics that MediaTek is thinking about. MediaTek has invested a lot in the field of AI computing technology. A few years ago, it believed that future computing will not only be CPU and GPU, but also the trend of AI-based APU will rise, so it also introduced the APU architecture into its Dimensity 5G chip. MediaTek will launch the third-generation Dimensity series chip, Dimensity 9300, at the end of October, which will support AI computing technology using LLM.
Editor: Xinzhixun-Linzi
The above is the detailed content of MediaTek is rumored to have won a large order from Google for server AI chips and will supply high-speed Serdes chips. For more information, please follow other related articles on the PHP Chinese website!