AI chips are designed to perform complex AI tasks efficiently. They consume much less power and provide much more data security and latency. But developing an AI chip has its challenges. What if we could program AI to design computer chips?
Google chip implementation and infrastructure teams and scientists at Google research want to take a Learning-based approach to chip design. Based on past experience and improvements over time, AI can equip itself to design an AI chip in under six hours on an average.This could be achieved by mainly automating the placement of on-chip transistors.
If the public can gain access to this research, a lot of cash-strapped startups could develop their own chips and do much more. Furthermore, this reduction in the chip design cycle will enable hardware to adapt better to rapidly evolving research.
Jeff Dean, Google AI lead, aims to place a “netlist” graph of memory and logic gates onto a chip canvas. So that, while adhering to constraints in placement density and routing congestion, the design optimizes power, performance, and area (PPA.) The researchers plan to devise a reinforcement learning training agent. After a lot of training over a lot of chips, they were able to speed up the training process and generate high-quality results faster.
Google’s pre-print paper co-authored by Jeff Deans describes that their technique:
- Leverages knowledge gained from placing prior chips to become better over time.
- Enables direct optimization of target metrics.
- Easy to incorporate new cost functions
- Weigh relative importance of cost functions according to the needs of a given block chip.
#AIMonks #AI #ArtificialIntelligence #Google #GoogleAI #ComputerChips #AIChips #Netlist #Infrastructure #Research #PPA
0 Comments