Deep learning is a form of artificial intelligence that excels at integrating the human brain, resulting in improved decision-making capabilities. The integration of AI chips is one such application that has gotten a lot of publicity.
Around a year ago, Jeff Dean, an American computer scientist who also serves as Google’s brain chief, stated that Google will use artificial intelligence to advance its internal creation of custom chips. This will pave the way for the company’s applications to be faster in the long run. Dean discussed how machine learning could change the way we think about things and be used to make some low-level design decisions.
Dean discussed how machine learning could change the way we look at things and be used for “place and route” design decisions at a low level. Chip designers will use software to decide the layout of the circuits that shape the chip’s operations in this case. This is eerily close to designing a building’s floor plan.
Dean discussed how machine learning could change the way we look at things and be used for “place and route” design decisions at a low level. Chip designers will use software to decide the layout of the circuits that shape the chip’s operations in this case. This is eerily close to designing a building’s floor plan.
Google recently announced Apollo as one of its research projects, which is a fascinating creation. This goes beyond what Jeff Dean said a year ago. However, what can be seen in Apollo is that the software is exploring architecture. This is in direct contrast to what Dean said at the time. Apollo doesn’t mention a floor plan in particular. Apollo’s goal is to test various methods systematically and determine which ones work best.
This degree of “architecture discovery” is far higher than place-and-route, and it’s also where there’s more room for performance improvement. It operates on a variety of chips, and there is a dedicated team dedicated to developing AI accelerator chips.
The real challenge now is to create AI chips. Apollo’s architecture would be based on neural networks, so it would be similar. And here’s why: it’s all about linear algebra, which consists of basic mathematical units that perform matrix multiplications and add the results together. Well, this isn’t going to be a simple mission. The quest isn’t restricted to only a few criteria. The quest isn’t restricted to only a few criteria. To name a few, areas such as how many of the math units, known as processor components, will be used, as well as how much parameter memory and activation memory would be ideal for a given model, must be addressed.
Follow and connect with us on Facebook, Linkedin & Twitter
