The race towards artificial intelligence chips is mounting

Making things smarter by connecting, integrating sensors and building artificial intelligence software within them for most organizations is the biggest challenge in the next five years. Everything will become smarter in theory, but the limitations of current computer chips slow down the process, today’s technologies are simply not at the task level, which is recognized by Bristol-based Graphcore, which is responsible for developing a new chip to help accelerate the adoption of artificial intelligence. Some neuronal networks work well enough using cloud computing and large data sets. However, the most powerful artificial intelligence systems are struggling to develop rapidly complex complex calculations when using current computer processing units that run sequentially in other words, reduce time Response.

Nigel Toon, co-founder of Graphcore, said: “70 years ago, we programmed computers to work according to step-by-step instructions. Artificial intelligence, however, involves learning computers and adapting to the data they process. The talk is simple enough to understand, With the existing technology at the moment, but understanding the whole language and the context in which the words are more difficult, and requires systems to store data and deep in memory to understand the background of conversations, and the things required to learn from the data significantly on the traditional process, Completely out of the workload. “

Temporary solutions, including CPU status in the cloud to share the amount of work to be done and the use of graphics processing units, are not fast enough for the rapidly evolving artificial intelligence world, and many companies like Google, Amazon and Apple are working on devices to solve the problem, Resulting in an unprecedented influx of capital into emerging technology companies.

In 2012, Nigel Tun launched the Icera semiconductor company , in collaboration with co-founder Simon Knowles, which was sold in 2011 to the NVIDIA chip maker for $ 435 million. This project helped him to think about hardware restrictions Which faces artificial intelligence.

In 2016, Nigel Tun and Simon Knowles will continue to work with researchers to identify their problems and future plans. They have decided to work from the first principles, think less about software and focus more on the same computer. Their new solution requires building a whole new type of processor , And think about the workloads borne by the computer in a different way.

CPUs usually solve problems by assembling data blocks and then running algorithms or logical processes on that information sequentially. The quad-core chips have four parallel processors, and the graphics processors for the games have parallel processors that can perform multiple tasks Same time.

Computers with artificial intelligence systems need to pull large amounts of data in parallel from different locations and then process them quickly. This process is called graphical computing, which focuses on nodes and networks rather than instructions. The new Graphcore chip, IPU, on graphical computing in parallel with low-resolution floating-point computing.

“The hardware structure is simple and straightforward, and you simply can not access the hardware stage and then try to figure out how to write the program,” says Nigel Tun. The difference in how individual processors on the chip communicate with each other and external memory, which is run by Graphware’s Poplar program .

Poplar moves the data through the chip more efficiently, which means less wasted processing capacity, as it does in a timely manner, using all the processors sequentially, so performance improvements are very important. The Graphcore chip can handle advanced artificial intelligence algorithms Up to ten times as fast as today’s largest processors.

Graphcore claims that its architecture for handling and processing data will be 100 times more powerful than the most powerful graphics processing unit, opening up new opportunities and applications for people.

Sequoia Capital, the investment arm of Sequoia Capital, which focuses on investing in the technology industry and has previously invested in companies like Google and Apple, has provided $ 50 million to the project to help it grow and reach the next generation of Silicon technology and Moore’s law persistence, we can get more transistors in a smaller space, making us see new breakthroughs thanks to artificial intelligence.