It's amazing to think how far computing has progressed over recent decades, from mainframe computers that literally filled a large room and needed an army of people to operate, to what we have today. The smartphone in your pocket has many times the processing power that took astronauts safely to the moon and back in the ‘60s and ‘70s. It's almost impossible to exaggerate the progress made in increasing the processing power, whilst at the same time shrinking the size of modern-day CPUs that power computers of all shapes and sizes. What has been accomplished is truly amazing, yet believe it or not CPUs (Central Processing Units) and traditional compute remain incorrigibly dumb! By which I mean, they only do as they are told, they cannot think for themselves. They do what the programmers have asked them to do using logic and algorithms etc, they do it very accurately and very fast but that is all they do. The ‘intelligence’ to solve a particular problem is basically ‘in the programmers’ not in the CPU. So CPUs have done a great job and will continue to do so for the majority of compute, however we are now in the early stages of intelligent machines or AI as it is more commonly known, learning from the data they are given.
UK-based Graphcore is shaking up the traditional CPU market with a chip and PCIe card designed exclusively for AI processing.
The birth of Artificial Intelligence
Way back in the 1950s Alan Turing foresaw that machines one day will themselves learn: “It seems probable that once the machine thinking method had started, it would not take long to outstrip our feeble powers…” AI has been gaining traction over the past several years with so called machine learning and deep learning. These are topics beyond the scope of this article but the hardware used to drive these solutions has principally been the GPU (Graphics Processing Unit). As the name suggests, GPUs original purpose and design point was only to provide the capability for a computer to display graphics on a screen. It turns out (by serendipity?) that the features of a GPU to display pixels on a screen, i.e. do a simple task but repeat it very quickly and many times over is also very useful for a number of AI workloads. This is basically why GPUs have been so successful in AI applications in the early years of intelligence computing.
"CPUs and traditional compute remain incorrigibly dumb... they only do as they are told, they cannot think for themselves."
The race to AI supremacy
But what of the future? Both CPUs and GPUs are deterministic, in other words, they work in a very precise way, either with scalar constructs in the case of CPUs or vector in the case of GPUs. Whilst this has worked well in deterministic environments, what is called for with future AI workloads is rather different! For 'next gen' AI workloads what is needed is not just lots of processing units but also a very flexible and capable interconnect. For example, if you were to examine any part of the brain, it is about 25% ‘processing’ (neuron) and 75% ‘interconnect’ (synapse). It is not so much the raw intelligence of the synapse but the flexibility of the interconnects that gives the brain its full capability. One of the leading new technologies aiming to power this new era is the Intelligence Processing Unit or IPU, as pioneered by UK company Graphcore. This is designed from the ground up for next generation AI workloads, unlike current solutions that were originally designed for other workloads and later adapted to suit AI. It's a sea change in the architecture of a 'processing unit', and elsewhere other large enterprises are pushing the boundaries in subtly different ways, such as Alibaba's Neural Processing Unit (NPU). Of course any new system requires both new hardware and a new software ecosystem to drive it as well as a thriving developer community. This is no small task but performance improvements of 100x have been cited with the IPU, so we wait with anticipation the development of these AI-focused chips.
"This has only happened three times in the history of computers. CPUs in the 1970s, GPUs in the 1990s. Graphcore is the third. Their chip is one of the great new architectures of the world."
Investor Hermann Hauser, co-founder of Arm Holdings, is betting that IPUs will unleash the next wave of computing.
At TBT Marketing we're on the leading edge of technological developments and are passionate about communicating the next generation of enterprise computing. Need ideas and strategies to take your products forward? Then talk to us or call on +44 (0)1373 469270.