AI demands a different procesor unit, not the present CPU, GPU and FPGA turbo
Moore's law is no more that valid in terms of costs, and the present mass-used CPU architectures hardly handles streaming real-time high-bandwidth data without latency. They also consume lot of power. The long lasted Von Neumann architecture is sliding towards obsolescence.
GPUs and FPGA accelerators could provide bypass to the processing bottlenecks for some data types. But that's a temporary relief to improve efficiency of present computer architecture. Like addition of turbo to diesel internal combustion engine. Future human brain type computing processors need a new processing engine similar to the electric motor replacing gasoline engine in automotives. And in fact autonomous driving demands modern architecture computer to deliver fast processing, consuming 10s of watts instead of 100s of watts, and decide with zero error while driving a car on a highway with speeding traffic.
In series of articles, we will bring to you processors and other related products emerging in this area:
1. Intelligence Processing Unit (IPU) from Graphcore Systems:
Graphcore designed a new intelligent processor unit designed for machine learning workloads. The IPU employs massively parallel neural network like, low-precision floating-point processing with graph software tools and libraries. Graphcore is also offering its Poplar software framework supporting current and future machine intelligence applications to bring down the cost of Artificial Intelligence (AI) in the cloud and datacenter.
Applications written for Tensorflow will work with less of an effort on an IPU with Poplar software, which is a C++ framework that provides a seamless interface to standard machine learning frameworks, like Tensorflow and MXNet.
Poplar has a full set of drivers, application libraries and debugging and analysis tools to help tune performance and a C++ and Python interface for application development to speed up the development.
IPU features large amounts of on-chip SRAM supporting training and inference across a large range of machine learning algorithms.
Is the Graphcore' IPU performance significant?
Graphcore claims its devices improves performance by 10x to 100x compared with other AI accelerators. Is that good enough? Not really if we look at the computing needs required in a robot/drone powered by a low capacity battery.
Investors in Graphcore include Bosch Venture Capital, Foundation Capital, and Samsung Catalyst Fund.