Intel develops custom AI tool for SoC thermal sensor placement

Intel develops custom AI tool for SoC thermal sensor placement

Technology News |
By Nick Flaherty

An augmented intelligence tool developed in-house by Intel engineers is providing data on sensor placement in system-on-chip (SoC) designs in just minutes.

For decades, determining exactly where to place heat-sensitive sensors on Intel processors required equal parts science and art.

Circuit designers would be guided by historical data when deciding where to place thermal sensors on the central processing units (CPUs). Experience provided data on exactly where hotspots tend to happen, and this process could take up to six weeks of testing, running simulated workloads, optimizing sensor placement, and then repeating the process all over again.

The tool was developed by the Augmented Intelligence team led by Dr. Olena Zhu, senior principal engineer and AI solution architect in Intel’s Client Computing Group (CCG).

“This tool has completely revolutionized the way we do thermals today. It’s so much more efficient and gives us so much more visibility to thermal risks before we turn the SoC on. We’ve been feeling our way in the dark, but with augmented intelligence, we’ve been given a flashlight to guide the way forward,” said Mark Gallina, CCG principal engineer and senior system thermal and mechanical architect.

The augmented intelligence tool allows engineers to key in the boundary conditions and the tool provides design suggestions in minutes. It as used on the Intel Core Ultra Meteor Lake mobile processor family and is being used on the coming Lunar Lake and successor chips.

“Client products like laptops rely heavily on turbo and peak frequencies. You want the SoC to burst to higher frequencies, which in turn generates thermal heat,” says Gallina.

Engineers must precisely analyze complex, concurrent workloads that activate the CPU core, input/output (I/O) and other system functions to accurately determine the location of thermal hotspots. Complicating the process is determining where engineers place the thermal sensors.  

“That process takes a few weeks, and we are limited to looking into one or two workloads at a time,” he said.

The team trains AI models based on simulations or measurements of a small number of workloads. These AI models then predict other workloads that are not being simulated or measured by Intel. “Augmented intelligence leads to a new breed of tools that allows us to manipulate data much more efficiently than ever before,” said Olena. “When we combine AI with our existing engineering bench strength, we can find the needle in the haystack much more efficiently.”

Augmented AI for chip design 

Augmented AI tools are being used in other areas at Intel. A fast and accurate signal integrity analysis tool for high-speed I/O reduces design time from months to an hour, while an AI-based automatic failure analysis tool for high-speed I/O design, deployed since 2020, led to 60% efficiency gains.

An augmented intelligence tool called “AI-Assist” uses an AI model to automatically determine customized overclocking values for different platforms. This reduces overclocking time from days to just one minute.

An AI-based automatic silicon floor plan optimizer is incorporated into Intel’s SoC design flow while a smart-sampling tool to help power and performance engineers crunch smart design experiments reduced the number of testing cases by as much as 40%.

“There’s a fast growing movement in the industry to infuse AI into similar engineering usages, and Intel is definitely taking advantage of it and embracing it,” says Olena.

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News


Linked Articles