Google and AMD Collaborate on TPU: Next-Generation Hybrid Design on the Way
It is claimed that Google is collaborating with AMD on its 10th generation TPU projects. Details of the reinforcement learning-focused hybrid design are in our news.
New information has emerged regarding Google’s collaboration with AMD on the development of its 10th-generation TPU. A report shared by SemiAnalysis suggests that Google may be interested in AMD’s CPU cores for processor-intensive workloads.
This potential collaboration could be AMD’s first major undertaking in a dedicated AI ASIC project. It is claimed that Google is working on a new type of TPU that combines its own accelerator technology with general-purpose processor cores.
Hybrid Design and Processor Needs
Google has been developing its own AI accelerators for nine generations, working with Broadcom in the process. However, the CPU IP, programmable logic, and advanced packaging technologies required for the new generation TPUs may have led Google to work with a processor manufacturer like AMD.
Reinforcement learning and reasoning models, in particular, require significant general-purpose processing power in addition to accelerator operations. Google had already begun increasing processor resources in its TPU 8i systems by using one Google Axion CPU for each TPU.
The CPU to accelerator ratio approaching 1:1 on the server side indicates that future AI systems will be processor-centric. Integrating CPU cores directly into the TPU package can improve performance and reduce power consumption by reducing the gap between general-purpose processing power and tensor computing.
AMD’s Experience and Future Expectations
AMD has valuable experience in combining x86 and accelerator chips in a single package with its Instinct MI300A design. It is seen as a logical step for Google to adopt this design approach and combine its TPU chips with AMD’s CPU and HBM technologies.
Although Intel is also considered a candidate due to strategic collaborations, its lack of expertise in hybrid data center designs leaves it in the background. While details of AMD’s participation in this project are not yet clear, it appears that Google is focusing on a new family of TPUs with reinforcement learning.
What are your thoughts on Google’s intentions regarding its increased focus on CPU power in AI hardware?
