Posted: Peter McGuinness 05.21.18
The story of the last twenty years has been that everything is software. As long as standard CPUs and DSPs were able to step up and deliver the performance, programmable solutions ruled. Now, with the arrival of machine learning and—especially—deep neural networks, this dominance has evaporated, no longer is everything “just another workload.”
Under the twin pressures to deliver extreme computational performance at the same time as reining in power consumption, the search for new hardware architectures is well under way. Established players are working to leverage their existing strengths in DSP and GPU to gain a foothold and for the first time in two decades, significant amounts of venture capital are flowing into fabless chip startups revisiting old data-driven architectures and developing completely new ones to suit the rapidly emerging market that demands intelligent devices everywhere.
Where there is hardware, there will be APIs. Every major operating system now has a machine learning API available to it, and every major hardware vendor is working to establish a deployment environment that will drive OEMs to use their hardware. Where Nvidia has established a dominant position with CUDA for GPU acceleration in the cloud, a range of upstarts are working to disrupt that by making it possible for developers to go direct from Tensorflow to their specialized hardware. Arm, late to the game, has realized its opportunity to dominate the embedded world is slipping away and is scrambling to remedy the situation.
While the automotive industry has been aware for some time of the potential of machine learning, the vastly larger IoT world is starting to realize that this is a foundational enabling technology for them. This is especially true of Chinese semiconductor companies and system integrators, who are acting on a government sponsored strategy to bring smarts to everything that touches the consumer. Given their strategic focus and investment, could this be the technology that gives China its long awaited breakthrough into the first rank of innovators?
It’s not just China: Sweden, Russia, France, Italy, Germany, The UK, Greece, Ireland, Hungary and Israel are all significant, credible players in this game. It’s a changing world: the last major disruption based around new architectures for computational workloads was the development of accelerators for graphics, largely based in the US and Japan. Machine learning is a worldwide phenomenon, broader in scope and likely to have a much greater impact on our daily lives and on the semiconductor industry.
Hold onto your hats, it’s going to be a wild ride.