Google claims its AI is faster in chip design – Desde Linux
Desde Linux » News
4 minutes
Google claims to have developed a software of artificial intelligence capable of designing computer chips faster than humans. In an article published a few days ago, Google claims that a chip that would take humans months to design can be imagined by its new AI in less than six hours.
Artificial intelligence has already been used to develop the latest iteration of the chips Tensioner Processing Unit (TPU) by Google, which are used to perform artificial intelligence-related tasks, Google said. Google engineers said the advance could have “major implications” for the semiconductor industry.
Essentially, it’s about figuring out where components like the CPU and GPU cores and memory are placed against each other on the chip. Their location on these small boards is important because it affects the power consumption and processing speed of the chip; the wiring and signal routing required to connect everything is of great importance.
Google engineers Azalia Mirhoseini and Anna Goldie, along with their colleagues, describe in their publication a deep reinforcement learning system capable of creating “basic patterns” in less than six hours, while sometimes it takes months.
In other words, Google is using artificial intelligence to design chips that can be used to create even more sophisticated artificial intelligence systems.
Similar systems can also beat humans in complex games like go and chess. In these scenarios, the algorithms are trained to move pieces that increase your chances of winning the game, but in the tile scenario, the AI is trained to find the best combination of components to be as efficient as possible in the game.
The neural network also uses some techniques that were once considered by the semiconductor industry, but abandoned as dead ends. According to the article, the artificial intelligence system received 10.000 blueprints for chips to “learn” what works and what doesn’t.
“Our approach has been used to design the next generation of Google’s AI accelerators and has the potential to save thousands of hours of human effort for each new generation,” the engineers wrote. “Ultimately, we believe that more powerful AI-designed hardware will drive the advancement of AI, creating a symbiotic relationship between the two fields.”
According to the article, when designing a microprocessor or a workload accelerator, it is usually necessary to define how its subsystems work in a high-level language, such as VHDL, SystemVerilog, or maybe even Chisel.
This code will eventually translate into what is called a netlist, which describes how a set of macroblocks and standard cells must be connected by wires to perform the functions of the chip.
Standard cells contain basic elements such as NAND and NOR logic gates, whereas macroblocks contain a set of standard cells or other electronic components intended to perform a special function, such as providing on-chip memory or a processor core. Therefore, the macroblocks are much larger than standard cells.
Then you have to choose how to organize this list of cells and macroblocks on the chip. According to Google employees, it can take human engineers weeks or even months to work with specialized chip design tools and iterate many times to get an optimized plan based on needs for power consumption, timing, speed, etc.
What usually happens in this process is that the location of the large macroblocks must be changed as the design develops. And then you have to let the automated tools, which use unintelligent algorithms, drop in the multitude of smaller standard cells, and then clean and repeat until you’re done, the doc says.
To accelerate this step of chip schematic design, Google artificial intelligence specialists created a convolutional neural network system that performs macro-block placement on its own in a few hours to achieve an optimal design.
Standard cells are automatically placed in empty spaces by other software, according to the article. This machine learning system should be able to produce an ideal diagram much faster and better than the human engineers method using traditional automated tools in the industry, Google employees explained in their article.
Source: https://www.theregister.com/
Be the first to comment
Your email address will not be published. Required fields are marked with *
*
*
*
