AI data centers require powerful network chips. These chips smoothly link custom hardware, graphics processors, and core units. Consequently, the optical fiber industry is seeing a rapid surge in new orders.
Furthermore, linking multiple servers demands superior connectivity. Therefore, developing high-performance network silicon has become a vital goal for tech leaders. For instance, Google plans to build custom chips instead of purchasing off-the-shelf vendor products.
The Marvell Alliance and Intel’s Foundry
Google originally built its Tensor Processing Unit as a custom chip. This hardware explicitly speeds up machine learning and AI tasks. Specifically, the TPU focuses on low-level matrix math within neural networks.
Currently, Google Cloud hosts the firm’s own Gemini models. These advanced systems rely heavily on TPU hardware for operational speed.
Strategic Silicon Design
Recently, industry sources revealed a major infrastructure shift. Google commissioned Marvell Technology to design custom network chips for the TPU ecosystem. This project aims to improve TPU efficiency and enhance data center connections.
Presently, advanced factory capacity at TSMC remains tightly limited. Therefore, Intel will likely manufacture this custom network silicon. The foundry will utilize its modern Intel 18A or 18AP processes.
According to internal reports, mass production will begin by late 2027. Consequently, these custom processors will likely debut inside the future TPUv8e model. Google will subsequently deploy these units at scale in new data centers by 2028.
Deep Engineering with MediaTek
Concurrently, industry reports highlight an active partnership between Google and MediaTek. The two firms are co-developing the next-generation TPUv8e framework.
Division of Labor
Within this project, Google retains primary control over the main chip logic. Conversely, MediaTek manages the input/output blocks and backend engineering. Upon successful check, Intel will manufacture the silicon and apply its advanced EMIB packaging.
Presently, Google runs active TPU deployments across eleven global data centers. This footprint explicitly includes two dedicated AI zones. Ultimately, this massive network underscores the urgent need for efficient chips to link global clusters.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
