“I’m sure Nvidia will be successful as well,” he ventures. “They’re already being very successful in this market… And being a viable competitor and standing alongside them, I think that would be a worthy aim for ourselves.”
Toon also flags what he couches an “interesting absence” in the competitive landscape vis-a-vis other major players “that you’d expect to be there” — e.g. Intel. (Though clearly Intel is spending to plug the gap.)
A recent report by analyst Gartner suggests AI technologies will be in almost every software product by 2020. The race for more powerful hardware engines to underpin the machine-learning software tsunami is, very clearly, on.
“We started on this journey rather earlier than many other companies,” says Toon. “We’re probably two years ahead, so we’ve definitely got an opportunity to be one of the first people out with a solution that is really designed for this application. And because we’re ahead we’ve been able to get the excitement and interest from some of these key innovators who are giving us the right feedback.”
Bristol, UK based Graphcore has just closed a $30 million Series B round, led by Atomico, fast-following a $32M Series A in October 2016. It’s building dedicated processing hardware plus a software framework for machine learning developers to accelerate building their own AI applications — with the stated aim of becoming the leader in the market for “machine intelligence processors”.
In a supporting statement, Atomico Partner Siraj Khaliq, who is joining the Graphcore board, talks up its potential as being to “accelerate the pace of innovation itself”. “Graphcore’s first IPU delivers one to two orders of magnitude more performance over the latest industry offerings, making it possible to develop new models with far less time waiting around for algorithms to finish running,” he adds.
Toon says the company saw a lot of investor interest after uncloaking at the time of its Series A last October — hence it decided to do an “earlier than planned” Series B. “That would allow us to scale the company more quickly, support more customers, and just grow more quickly,” he tells TechCrunch. “And it still gives us the option to raise more money next year to then really accelerate that ramp after we’ve got our product out.”
The new funding brings on board some new high profile angel investors — including DeepMind co-founder Demis Hassabis and Uber chief scientist Zoubin Ghahramani. So you can hazard a pretty educated guess as to which tech giants Graphcore might be working closely with during the development phase of its AI processing system (albeit Toon is quick to emphasize that angels such as Hassabis are investing in a personal capacity).
“We can’t really make any statements about what Google might be doing,” he adds. “We haven’t announced any customers yet but we’re obviously working with a number of leading players here — and we’ve got the support from these individuals which you can infer there’s quite a lot of interest in what we’re doing.”
Other angels joining the Series B include OpenAI‘s Greg Brockman, Ilya Sutskever, Pieter Abbeel and Scott Gray. While existing Graphcore investors Amadeus Capital Partners, Robert Bosch Venture Capital, C4 Ventures, Dell Technologies Capital, Draper Esprit, Foundation Capital, Pitango and Samsung Catalyst Fund also participated in the round.
Commenting in a statement, Uber’s Ghahramani argues that current processing hardware is holding back the development of alternative machine learning approaches that he suggests could contribute to “radical leaps forward in machine intelligence”.
“Deep neural networks have allowed us to make massive progress over the last few years, but there are also many other machine learning approaches,” he says. “A new type of hardware that can support and combine alternative techniques, together with deep neural networks, will have a massive impact.”
Graphcore has raised around $60M to date — with Toon saying its now 60-strong team has been working “in earnest” on the business for a full three years, though the company origins stretch back as far as 2013.
In 2011 the co-founders sold their previous company, Icera — which did baseband processing for 2G, 3G and 4G cellular technology for mobile comms — to Nvidia. “After selling that company we started thinking about this problem and this opportunity. We started talking to some of the leading innovators in the space and started to put a team together around about 2013,” he explains.
Graphcore is building what it calls an IPU — aka an “intelligence processing unit” — offering dedicated processing hardware designed for machine learning tasks vs the serendipity of repurposed GPUs which have been helping to drive the AI boom thus far. Or indeed the vast clusters of CPUs otherwise needed (but not well suited) for such intensive processing.
It’s also building graph-framework software for interfacing with the hardware, called Poplar, designed to mesh with different machine learning frameworks to enable developers to easily tap into a system that it claims will increase the performance of both machine learning training and inference by 10x to 100x vs the “fastest systems today”.
Toon says it’s hoping to get the IPU in the hands of “early access customers” by the end of the year. “That will be in a system form,” he adds.
“Although at the heart of what we’re doing is we’re building a processor, we’re building our own chip — leading edge process, 16 nanometer — we’re actually going to deliver that as a system solution, so we’ll deliver PCI express cards and we’ll actually put that into a chassis so that you can put clusters of these IPUs all working together to make it easy for people to use.
“Through next year we’ll be rolling out to a broader number of customers. And hoping to get our technology into some of the larger cloud environments as well so it’s available to a broad number of developers.”
Discussing the difference between the design of its IPU vs GPUs that are also being used to power machine learning, he sums it up thus: “GPUs are kind of rigid, locked together, everything doing the same thing… all at the same time, whereas we have thousands of processors all doing separate things, all working together across the machine learning task.
“The challenge that [processing via IPUs] throws up… is to actually get those processors to work together, to be able to share the information that they need to share between them, to schedule the exchange of information between the processors and also to create a software environment that’s easy for people to program that’s really where the complexity lies and that’s really what we have set out to solve.”
“I think we’ve got some fairly elegant solutions to those problems,” he adds. “And that’s really what’s causing the interest around what we’re doing.”
He says Graphcore’s team is aiming for a “completely seamless” interface between its hardware — via its graph-framework — and widely used high level machine learning frameworks including Tensorflow, Caffe2, MxNet and PyTorch.
“You use the same environments, you write exactly the same model, and you feed it… through what we call Poplar [a C++ framework],” he notes. “In most cases that will be completely seamless.”
Although he confirms that developers working more outside the current AI mainstream — say by trying to create new neural network structures, or working with other machine learning techniques such as decision trees or Markov field — may need to make some manual modifications to make use of its IPUs.
“In those cases there might be some primitives or some library elements that they need to modify,” he notes. “The libraries we provide are all open so they can just modify something, change it for their own purposes.”
The apparently insatiable demand for machine learning within the tech industry is being driven — at least in part — by a major shift in the type of data that needs to be understood from text to pictures and video, says Toon. Which means there are increasing numbers of companies that “really need machine learning”. “It’s the only way they can get their head around and understand what this sort of unstructured data is that’s sitting on their website,” he argues.
Beyond that, he points to various emerging technologies and complex scientific challenges it’s hoped could also benefit from accelerated development of AI — from autonomous cars to drug discovery with better medical outcomes.
“A lot of cancer drugs are very invasive and have terrible side effects, so there’s all kinds of areas where this technology can have a real impact,” he suggests. “People look at this and think it’s going to take 20 years [for AI-powered technologies to work] but if you’ve got the right hardware available [development could be sped up].
“Look at how quickly Google Translate has got better using machine learning and that same acceleration I think can apply to some of these very interesting and important areas as well.”
In a supporting statement, DeepMind’s Hassabis also goes to far as to suggest that dedicated AI processing hardware might offer a leg up to the sci-fi holy grail goal of developing artificial general intelligence (vs the more narrow AIs that comprise the current cutting edge).
“Building systems capable of general artificial intelligence means developing algorithms that can learn from raw data and generalize this learning across a wide range of tasks. This requires a lot of processing power, and the innovative architecture underpinning Graphcore‘s processors holds a huge amount of promise,” he adds.