NVIDIA is well aware of modern power efficiency woes in relation and designing circuits is the start of remedying this growing concern. With NVIDIA being one of the foremost companies that progress AI forward, its no surprise the company is now planning to use AI to develop future GPUs.
NVIDIA is utilizing the PrefixRL model that uses AI deep reinforcement learning to better outperform EDA design tools from traditional vendors. EDA vendors have their own AI solutions for their own processes but NVIDIA’s PrefixRL method has proved to be very impressive for NVIDIA.
In their company blog post, NVIDIA highlights the benefits of this compute-intensive workload with NVIDIA tasking this process to a cluster of 256 CPUs which took 32,000 GPU hours to train for a 64-bit adder circuit. NVIDIA has developed Raptor which is a deep reinforcement learning platform which utilizes NVIDIA hardware to further improve on this process.
The results are quite impressive but the underlying system is complex with the hardware supporting it equally demanding but with the resulting circuits becoming smaller and more efficient, so far the company says its worth.