Currently, Google is using AI to design chips. This is much faster than a human engineer would do.

In just 6 hours, the model was able to generate a design that optimized the placement of the various components on the chip.

Koko / Getty Images

A team of Google researchers has announced a new AI model that can enable complex chip designs in hours. This is a tedious and complex task that a human engineer usually takes months to complete.

Researchers provided a machine learning model using a dataset of 10,000 chip layouts and trained it in reinforcement learning. It turns out that this model can generate a design that optimizes the placement of various components on the chip in just 6 hours, creating the final layout that meets operational requirements such as processing speed and power efficiency. ..

The success of this method is that Google has already used this model to design a next-generation tensor processing unit (TPU) that runs in its data center to improve the performance of various AI applications.

“Our RL (Reinforcement Learning) agents generate chip layouts in just a few hours, but human experts can take months,” said Anna Goldie, a research scientist at Google Brain who participated in the study. Tweeted. “These superhuman AI-generated layouts were used in Google’s latest AI accelerator (TPU-v5)!”

Modern chips have billions of different components placed and connected to fingernail-sized silicon. For example, a processor typically contains tens of millions of logic gates, also known as standard cells, and thousands of memory blocks, called macroblocks. These need to be wired to each other.

Placing standard cells and macroblocks on the chip is important in determining how fast the signal can be transmitted on the chip and therefore how efficient the end device is.

This is why much of the engineer’s work focuses on optimizing chip layout. It starts by placing a larger macroblock. This is a process called “floor planning” that consists of finding the optimal configuration for a component, keeping in mind that standard cells and wiring need to be placed in the remaining space.

The number of possible layouts for macroblocks is enormous. According to Google researchers, there are 2,500 different configurations that can be tested to the 10th power, or 1 followed by 2,500 zeros.

In addition, after the engineer has come up with a layout and the addition of standard cells and wiring, it may be necessary to fine-tune and tweak the design thereafter. Each iteration can take up to several weeks.

Given the painstaking complexity of floor plans, the whole process seems to be clearly in line with automation. But for decades, researchers couldn’t come up with a technology that could take the burden off the engineer’s floor plan.

Chip designers can use computer software to assist in their tasks, but it still takes months to find the best way to assemble components on a device.

And the challenges are getting harder and harder. Moore’s Law, often quoted, predicts that the number of transistors on a chip will double each year. In short, engineers are faced with an equation that grows exponentially over time, while having to deal with tight schedules.

That’s why Google’s apparently successful attempt to automate floor plans could change the game. Facebook Chief AI Scientist Yann LeCun tweeted that “Google’s incredible work on deep RL-based optimization of chip layouts” overcame the “40-year” attempt to solve this challenge. Congratulations to the team.

Google’s new AI model rarely lands any longer. The semiconductor industry is currently swayed by a global shortage of chips that is hitting many sectors, from consumer electronics to automobiles.

Reducing the time it takes to invent next-generation chips can be a welcome remedy for the entire supply chain, although it is lacking due to lack of capacity at the manufacturing level rather than semiconductor design. ..

The scientific journal Nature welcomed a new method as one of them. “Google researchers have succeeded in significantly reducing the time required to design a microchip,” they said. “This is an important achievement and will greatly help speed up the supply chain.”

While machine learning models can impact the industry as a whole, it’s also worth paying attention to the use of Google’s own technology.

The search giant has long demonstrated that its ambition is to create custom processors in-house, especially in the form of system-on-chip (SoC).

Currently, Google is using AI to design chips. This is much faster than a human engineer would do.

Source link Currently, Google is using AI to design chips. This is much faster than a human engineer would do.

Back to top button