Bristol-based Graphcore’s “intelligence processing unit” aims to do for AI what the graphics processing unit did for computing.

In September 2015, hardware veterans Nigel Toon and Simon Knowles were doing the rounds of venture capital offices in Silicon Valley and London, touting their latest startup. The pair had a dazzling track record – among other achievements they’d sold their previous semiconductor company Icera to NVIDIA for $435 million (£346 million) four years earlier. And their vision for Graphcore – a new Bristol-based venture – was bold: they were building a new generation of microchips known as intelligence processing units (IPUs), designed for the rapidly approaching artificial intelligence age.

Yet early reactions to their pitch for series A financing were distinctly muted. “In many cases, we were laughed out of court,” recalls Toon, Graphcore’s CEO.

Typically, Toon says, they’d find a partner in a VC firm who was excited by what they were doing. “But then they’d go to their partner meeting, where the first question would be: ‘What’s AI?’ It’s stunning to think that was a conversation that was happening [as recently as] 2015.” From there, it was an uphill struggle. “Even if they got the fact that AI might be interesting, they’d then say: ‘Your business model is to build a chip for this AI thing? Well, nobody’s made money from chip investments in the last 10 years.’”

Toon, who is 55 and has the mellifluous voice of an old-school BBC continuity announcer, says that chip development, in the eyes of most investors at the time, was considered highly capital intensive, with returns failing to justify the upfront financing required. “It’s not more capital intensive than software,” says Knowles, Graphcore’s co-founder, and CTO. “But software has this joyful property that you can try it out on small scale first, whereas with a chip you’re all in. If it doesn’t work, you’ve spent all your money.

That was 2015. Fast forward to today and, of course, AI hardware is a white-hot category for investors, with VC funding for US AI companies jumping by 72 percent in 2018 to a record $9.3 billion (£7.4 billion), a fifth straight year of growth, according to a report by CB Insights and PwC.

What changed over those three years? Toon points to two things. First, in 2016 traditional chip giant Intel acquired an AI software and hardware startup called Nervana for $350 million (£280 million), raising eyebrows all over the Valley. Second, Google announced it was going to build its own chips – evidence, Toon says, that existing chips weren’t up to the task.

Knowles describes the impact of Google’s decision as “seismic”. The fact that Google thought AI was going to be a sufficiently big deal to justify the pain and expense of building its own chip team helped make the Graphcore founders’ case for them. He and Toon had been arguing that it was worth digging deep financially to develop new processor hardware because existing graphics processing units (GPUs) – used, for example, in mobile phones, games consoles, and personal computers – weren’t designed for AI workloads such as machine learning and deep learning.

By then, their startup was already ahead of the pack in developing a new processor architecture. Soon top-tier investors – including Atomico, one of Europe’s best-known VCs – were beating a path to their door. Atomico, which went on to lead Graphcore’s $30 million (£24 million) Series B round in July 2017, was followed six months later by one of the Valley’s biggest guns, Sequoia Capital. At the time, Graphcore, having recently closed its Series B, didn’t need investment – but the west coast investor wasn’t taking “No thanks” for an answer. “They came to see us here in Bristol and said, ‘No, you don’t understand, we want to invest in your business,’” laughs Toon. “So we work out terms and they invest $50m into the company. And that’s one of the very few investments they’ve made in the UK because they’ve got so much opportunity on their doorstep.”

Sequoia partner Matt Miller, who now sits on Graphcore’s board, admits he was somewhat bemused to find himself chasing down a company based in Bristol. “We knew there was an opportunity for a new architecture that would be designed from the ground up that could massively accelerate our entry into this AI age, and we were trying to landscape all of these companies in China, the US, and Europe,” he says. “But our references were all pointing to this one company in Bristol, whom we hadn’t met yet.”

A roar of laughter distorts the line from the Valley. “Lemme tell you, if you’d asked me a month prior if I’d ever [sit on] a board in Bristol I’d have said ‘No way!’ It’s not your typical destination on your tour of Europe. But to be honest, it’s been surprising for us in the Bay Area because the quality of talent in the UK, and particularly in Bristol in the semiconductor space, is very strong. The team they’ve been able to build there is on a par with the best in the world.”

Following a $200 million (£160 million) Series D round in December 2018, Graphcore was most recently valued at $1.7 billion (£1.36 billion), with investors, innovators and large corporates now seemingly convinced it will be the company to power the AI era in much the same way as Cambridge-born chip giant ARM dominated mobile devices, shipping over 130 billion chips and reaching 70 percent of the global population. The opportunity at stake is nothing less than the future of AI, with applications ranging from medical advances to autonomous vehicles, space exploration and just about everything in between.

Graphcore engineer Joanna Taylor belongs to “a semiconductor team on a par with the world’s best”

Nick Rochowski

In fact, Bristol has a strong history as a hub for hardware engineering, which can be traced back to 1978 and £50m of seed investment (another £150m would later follow) made by the UK government in Inmos, a microprocessor startup with fabrication facilities in Newport, South Wales. “We often forget the importance of government investment,” says Hermann Hauser, the Austrian-born entrepreneur, and investor is best known for spinning out ARM from Acorn Computers – and Graphcore’s first backer. “It was the £200 million that the Callaghan and, later, Thatcher governments originally spent on Inmos that created the infrastructure and ecosystem around Bristol that really understood semiconductors. It created brilliant people like [leading computer scientist] David May, Simon, and Nigel, who would not have been there had it not been for the government initiative at the time.”

Knowles first came to Bristol in 1989 to work for Inmos. “Historically, Bristol has been the center of chip design [in the UK], and in many ways ARM and CSR [formerly Cambridge Silicon Radio] were anomalies,” he says. “I mean, they’re very successful, large anomalies, and now everyone associates Cambridge with chips. But in terms of numbers of chip startups, and how many years back it goes, Bristol is the dominant place in the UK.”

Graphcore emerged from a tangled family tree of semiconductor companies. Toon and Knowles were introduced to each other by Stan Boland, former CEO of Acorn Group and now CEO of autonomous vehicle startup FiveAI, who had worked with Knowles at chip company Element 14. When this was acquired by Broadcom for $640 million (£512 million) in 2000, the pair went on to found Icera in 2002 with Toon, who was previously with electrical equipment manufacturer Altera Europe. When Icera was sold to NVIDIA, it meant that Knowles had already exited two chip design startups at a total value of over $1bn. But he and Toon were far from finished. What motivated them to start all over again with Graphcore?

Sitting across the table from one another in a fifth-floor meeting room at their Bristol HQ, the founders exchange a fleeting glance. After a while in their company, it’s clear that this long-established business double act has acquired some of the hallmarks of marriage: they have an easy rapport, finish each other’s sentences, and occasionally talk over and correct each other.

“Simon maybe has a different view,” says Toon, “but my sense of it is that this is what we get up in the morning for. The fact that the opportunity in front of us is so enormous, I feel like I’ve been waiting my whole life for this.” He adds that it comes down to purpose: “You might get some satisfaction from connecting people together in a social network, for example, or delivering food to them through an internet app. What we’re doing is potentially changing the future of computing – we’re potentially allowing lots of people to create major breakthroughs; maybe someone will come up with a cure for cancer using the tech we’re creating.”

“We’re building the motors of AI, really,” says Knowles. “And what people will build-out of those motors is far greater than our motors. We want to be the Rolls-Royce jet engines of AI machinery.”

In essence, the problem Graphcore is solving is that previous generations of microprocessors – central processing and graphics processing units – weren’t designed for machine intelligence, which requires a new way of processing data.

Knowles holds up a Graphcore chip. The size of a small cracker with a dark grey, metallic center, contains 23.6 billion transistor devices all connected by several miles of wiring. As transistors were progressively shrunk over the decades so that more of them could fit on to each chip, the chips themselves grew correspondingly hotter as energy demands increased. “We’re almost at the end of that gravy train now,” says Knowles. “The objective of chip design always used to be to go as fast as possible; now it’s to make the most use of the energy available.”

To make them as efficient as possible,” clarifies Toon.

“Exactly,” says Knowles. “And actually, you design things in a completely different way if you’re most interested in energy and less interested in speed per se. So why do we want more computing performance? We’ve just started to work out how to mechanize intelligence. And what do we mean by intelligence? A machine that can learn by its experience, or by being given examples, or by itself, discovering things. In no sense, historically, has a computer solved a problem – it was always the person who wrote the program. AI flips that on its head.”

Suddenly, there’s a surge in demand for more processing power due to the AI workload, at precisely the moment when traditional silicon shrinking won’t offer it. “Explaining to a computer how to learn is quite different from explaining to it how to do traditional supercomputer maths for example,” says Knowles. “So we’ve set about trying to solve those two problems – intelligence is a different workload, and focusing on efficiency and not speed – with our IPU.”

Whereas other AI hardware companies have focused on neural networks – a type of knowledge model for capturing the sort of intelligence in the human cortex, which is essentially designed to recognize numerical patterns – Graphcore has built an architecture that is more flexible. It can run current machine-learning approaches, as well as new and emerging approaches that simply don’t work efficiently on today’s hardware. “What most of the [rival] startups are doing is building a machine to do fast neural networks, and that’s what you do if your ambition for your company is to sell it for a couple of hundred million in a year or two,” says Knowles. “What we’ve tried to do – because our ambition for the company is to be permanent, and broad enough to encompass engines for AI as opposed to just chips for perception – is build a much more general-purpose machine. Nigel and I were very clear about our ambition for this company: we’ve grown and sold companies before, but this one is our magnum opus.”

Toon chips in: “This is a once-in-a-generation opportunity. If we get this right, the IPU will define the future of machine intelligence, powering world-changing innovations for decades to come.”

Graphcore’s Colossus GC2 IPU is a new generation of microprocessors built for the artificial intelligence age. Image: Nick Rochowski.

VCs are rarely sparing in their use of hyperbole. But when a big-hitting Valley investor like Sequoia’s Miller says “We think [Graphcore] can be a company with a market cap in the tens of billions of dollars”, and flies halfway around the world to make an investment in a startup that wasn’t raising money in the first place – with the likes of BMW, Microsoft, Bosch, Dell and Samsung also queuing up to invest – there tends to be a pretty good reason.

The answer lies in the almost limitless fields Graphcore’s IPU can be applied to – anywhere, in fact, that machine intelligence can enhance human activity. “There are still some things humans are going to be better at, typically creative things,” says Atomico partner Siraj Khaliq, a computer scientist, and former entrepreneur. “But when it comes to looking at patterns and making predictions – for example looking at a radiology scan and deciding if there’s cancer there or not; looking at someone’s viewing habits and deciding what they should watch next; even looking at the attributes of a person, what they do and what they like, and recommending who they should marry via dating apps – all of these things machines will now do because they’re just better at it. So I don’t think I’d be doing it justice by saying ‘Here are one or two things that Graphcore’s IPU will be used for’, because it is really pretty much everything.”

Back in Bristol, Knowles cites medicine and law as two areas on the brink of AI-driven transformation. “What is the definition of a good doctor or a good lawyer?” he says. “It’s someone with a lot of wisdom acquired by experience, someone who’s seen a lot of cases, read and digested a lot of research material and comes up with good answers. They can’t always be correct, but given the knowledge that exists they come up with the best reasonable answer based on their experience.”

The most exciting opportunity for machine intelligence is being able to do that with all of the human knowledge, he says. “Take a medical oracle which can read all of the medical research that’s ever been published and can resolve and identify discrepancies. It can read all of the patient records that have ever been recorded. And it can come up with the best answer based on all of the human knowledge. It’s not perfect, because all of the human knowledge isn’t all knowledge, but it’s the best we can possibly do and the opportunity there for totally solving a whole load of human conditions must be enormous.”

(To be continued…)

Source: Wired

Related posts: