Chipsets will drive Artificial Intelligence
Chipsets will continue to drive computing tasks in the foreseeable future. Processing power doubles every two years, so if we’re at four right now, the next exponential step will take us to eight, then to sixteen and so on. By 2025 computing power looks set to match that of the human brain. That’s awesome. Lot’s of processing power is needed to handle the massive amount of data that IoT generates. It’s needed because instructions run much slower in software than they do in dedicated hardware and this limitation is being exposed when support for AI is needed. Brute processing force hits a wall and using a technology that works sequentially isn’t fit for purpose.
Chipsets that employ multi-core processors can handle more than one task simultaneously. Different types of processors, each having a specialised function, are used to boost performance or save energy, but the idea isn’t new. “AI chips” advance this concept by embedding dedicated analytic software that executes machine-learning tasks. This enables AI systems to uncover hidden patterns in the data, predict actionable insights and go on to perform actions based on specific situations.
The market for these chips is at an early stage, but the ability to enhance AI’s capacity to solve real-world problems is not in question, as evidenced by the fact that more than 60 companies have announced some sort of deep learning chipset. Moreover prominent names in the technology industry have indicated the need for hardware acceleration of AI algorithms and the semiconductor industry has responded by releasing products for validation by the market.
Taking it to the next level
The big names in the semiconductor industry carry legacy baggage and they need to protect their manufacturing assets and market position. Therefore start-ups like Graphcore, a UK fabless chip company founded in 2016, can start with a clean sheet of paper and if things work out come up with something new and genuinely innovative.
In this case it’s a brand-new intelligent processor employing parallel computational resources that’s said to be the first that is specifically designed for machine intelligence workloads. Together with corresponding software, Graphcore states that processing speeds are ten to one hundred times faster than with conventional hardware. This should allow researchers to explore machine intelligence across a much broader front than current solutions. Moreover it would enable deep learning to evolve towards general artificial intelligence, which is AI’s ultimate goal, i.e. the creation of systems that learn about a problem and then solve it, which is what people do all the time.
Bringing it to the market
Realising a technology breakthrough is one thing; bringing it to the global marketplace and competing is another. Significant financial resources are needed as well as powerful partners. Graphcore has raised more than $300 million in fresh capital, of which $200 million came from BMW and Bosch in December 2018. In addition Microsoft is said to have participated as well as Dell and Samsung Electronics. Right now the competition is coming from ARM, Advanced Micro Devices, Huawei, IBM, Intel and Qualcomm.
Technology Editor, Beecham Research.