Artificial intelligence simulates microprocessor performance in real time

This approach is detailed in a paper presented at MICRO-54: The 54th IEEE/ACM International Symposium on Microarchitecture. Micro-54 is one of the top conferences in the field of computer architecture and was selected as the conference’s top publication.
“This is a problem that needs to be studied in depth and traditionally relies on additional circuitry to solve it,” said Zhiyao Xie, lead author of the paper and a doctoral student in Yiran Chen’s lab. Professor of Electrical and Computer Engineering at Duke. “But our approach runs directly on microprocessors in the background, which opens up a lot of new opportunities. I think that’s why people are excited about it.”

In modern computer processors, the calculation cycle is 3,000 billion times per second. Tracking the power consumed for such a fast conversion is important to maintain the performance and efficiency of the entire chip. If a processor draws too much power, it can overheat and cause damage. Sudden fluctuations in power demand can lead to internal electromagnetic complications that slow down the entire processor.
By implementing software that can predict and prevent these unwanted extremes, IT engineers can protect their hardware and improve its performance. But such a plan would have a cost. Keeping pace with modern microprocessors often requires valuable additional hardware and computing power.
“APOLLO is close to an ideal power estimation algorithm that is both accurate and fast and can easily be integrated into a processing core at low power cost,” Xie said. “Since it can be used in any type of processing unit, it could become a common component in future chip designs.”
The secret to Apollo’s power is artificial intelligence. The algorithm developed by Xie and Chen uses artificial intelligence to identify and select the 100 signals most closely related to power consumption from millions of processor signals. The company then built a power model from those 100 signals and monitored them to predict the performance of the entire chip in real time.
Because this learning process is self-contained and data-driven, it can be implemented on most computer processor architectures, even those yet to be invented. Although it doesn’t need any human designer expertise to do its job, the algorithm can help human designers do theirs.
“Once the AI ​​has picked out 100 signals, you can look at the algorithm and see what they are,” Xie said. “Many choices make intuitive sense, but even if they don’t, they can provide information to designers about processes relevant to power consumption and performance.”
This work is part of a collaboration with Arm Research, a computer engineering research organization that aims to analyze disruptions affecting the industry and create advanced solutions that can be deployed years in advance. APOLLO has been validated on some of today’s best processors with the help of Arm Research. But the algorithm needs to be thoroughly tested and evaluated on more platforms before it can be adopted by commercial computer makers, the researchers said.
“Arm Research has partnered with and secured funding from some of the best-known companies in the industry, such as Intel and IBM, and power consumption forecasting is one of their top priorities,” Chen said. “Programs like this provide our students with an opportunity to work with these industry leaders, and these results inspire them to continue working with and hiring Duke graduates.”
This study was conducted by the Arm Research High-Performance AClass CPU research program and was partially supported by the National Science Foundation (NSF-2106828, NSF-2112562) and Semiconductor Research Corporation (SRC).

Comments are closed.