Co-designer of the RISC chip which is now found in over a billion mobile devices worldwide.
Dr Steve Furber is the ICL Professor of Computer Engineering in the Department of Computer Science at the University of Manchester. He received his BA degree in Mathematics in 1974 and his PhD in Aerodynamics in 1980 from the University of Cambridge, England.
During his PhD Steve heard about the formation of the Cambridge University Processor Group and would later work with Sophie Wilson, Dr Hermann Hauser and Chris Curry.
From 1981 to 1990 he worked in the hardware development group within the R&D department at Acorn Computers Ltd with Hauser and Curry, and was a principal designer of the BBC Microcomputer and the ARM 32-bit RISC microprocessor with Wilson, both of which earned Acorn Computers a Queen’s Award for Technology.
He moved to the University of Manchester in 1990 and established the Amulet research group which has interests in asynchronous logic design and power-efficient computing, and which merged with the Parallel Architectures and Languages group in 2000 to form the Advanced Processor Technologies group.
Interviewed by Richard Sharpe on the 15 June 2018 at BCS London Office.
Steve Furber was born in Manchester in 1953. His father was a mechanical engineer, who spent his early career at UMIST in Manchester as a PhD student and lecturer, but in the mid-fifties moved to The Nuclear Power Group and he spent the rest of his career in the nuclear industry. After bringing him up, his mother, who had trained as a physiotherapist, took up the work of a maths teacher.
Steve was educated at the Manchester Grammar School, and it was clear from an early age that mathematics was his forte. He went to Manchester Grammar at the age of ten and didn’t take the Eleven Plus until toward the end of his first year there. MGS put a lot of its mathematicians in for the National Maths Contest, which was used to select members of the British team to go to the International Mathematical Olympiad, and he was fortunate to be chosen to go to Hungary, in 1970.
In the sixth form he had three mathematics teachers; Mr Wilkinson taught pure maths, Mr Schofield taught applied maths, and Mr Copley, taught whatever he felt like (or so it seemed to the pupils). He (Mr Copley) had all sorts of mathematical stories to tell and was involved in the development of the resonant cavity magnetron in the Second World War, as the first system that could generate radio waves that were sufficiently short that they could be used to detect submarines. After being used for hunting submarines in the Second World War, they then turned up in everybody’s kitchen, in their microwave oven, which Steve found an interesting story.
Steve finished at MGS with A Levels in maths A, further maths A, and physics A and with S Levels in maths and further maths. MGS was keen to send as many people as possible to Cambridge. It was partly reputational, but partly because, for mathematicians, it felt it was the right place to go. On an UCCA form you could fill six universities in, but on the school’s advice Steve just put Cambridge on the first line and left the rest blank. He was sixteen when he took his A Levels, and the school’s advice was, if he didn’t get in that year, to try again the next year. But he did get in that year with a Baylis scholarship to St John’s College.
Steve finished school eighteen months before starting at Cambridge. The first nine months he spent working with his father’s employer, on a temporary job at The Nuclear Power Group, and then for the second nine months he went to North America. He spent a term at McGill University in Montreal, attending any courses he felt like. Then he worked on a summer camp in Meadville, Pennsylvania, over the summer.
Steve got a First in his BA in Maths from Cambridge and went on to take Maths Part III which at the time didn’t result in any formal qualification. However, 35 years later the university decided that Part III was worth a master’s, so he got his MMath from Cambridge. He took the course in 1974/5 and got the degree in 2010.
In previous years Professor Ffowcs-Williams from the Engineering Department pad offered a Maths Part III course on aeroacoustics, but this course was not offered the year Steve took the course, which was disappointing as he had always had an amateur interest in aeroplanes. He contacted Ffowcs-Williams to ask him about PhD opportunities, and he was happy to take him. So, he changed from maths to the Engineering Department for his PhD,which he was awarded in 1980.
Cambridge University Processor Group - CUPG
During his PhD Steve heard about the formation of the Cambridge University Processor Group, CUPG. He was not a founder but went to the formation meeting. Because he was very interested in aeroplanes, he had turned this interest round to the idea that maybe it would be fun to build a flight simulator. What do you need to build a flight simulator? Well obviously, you need to start with some kind of computer. So, he thought he’d go along to this society and see what was going on. It was founded by people who build computers for fun. The real men used discrete logic, TTL chips as they were then, to build their computers, and the wimps like him bought these new-fangled microprocessor things. They were doing very scary things, ordering chips by mail order from California, using credit cards, which was all very new-fangled in those days. They then hand-assembled them into machines, and then went around to each other’s houses to be impressed by what each other could build and finding bugs and so on. It was in the Processor Group that Steve first met Sophie Wilson, who took great pleasure in finding bugs in Steve’s computer. At this time Hermann Hauser was a postdoc at the Cavendish Lab. He had started talking to Chris Curry about forming a company in the microprocessor area, because microprocessors were clearly going to be important. As the idea formulated, they clearly thought that the Processor Group was a good place to go and look for potential staff, people who could do the technical work. So Hermann found Steve through the Processor Group, although he wasn’t himself an active participant.
Chris Curry and Hermann Hauser decided to set up a company and looked for people who could do some of the technical work. Steve was involved from the outset and remembers that the first meeting, where they talked about the company, was in the Fort St George on Midsummer Common. He thinks it was Chris Curry, Hermann, Chris Turner, and him. Steve moonlighted with them, although he was initially a student funded from SERC, and then a Research Fellow, both circumstances where you can’t arbitrarily undertake additional employment. However, he found what they were doing interesting, so they had a kind of deal whereby he would design bits and pieces, and hand the designs to them, and then they’d give him more bits and pieces to play with. So, he was funded in kind rather than employed.
Acorn started as CPU Limited, which did a bit of consultancy work. When Sophie Wilson designed what became the Acorn System 1, Hermann and Chris’s thoughts turned to selling computer products rather than running a consultancy. The Acorn System 1 was marketed as a kit, and sold in quite reasonable numbers, and the whole business moved over to that side.
That model was followed by the Acorn Atom, which was their first single-box machine. At least, it had the computer and the keyboard in; you still needed an external screen and an audio cassette or something to save your programs and data on.
The Atom was solid, but you could make it less robust by adding things to it, and people added a lot of things to it. It was the first machine that could have an Econet module attached and do some basic local area networking for instance, which in ’79 was pretty advanced. But you could also put a ROM extender card in, so you could have more different bits of software. And, if you put enough things in, the power was regulated by a little regulator, and the printed circuit board was kind of upside-down, so the keyboard was on the back, and all the main components were underneath. This little 7805 regulator would get so hot supporting all these added components that it would literally melt the solder and drop out and rattle around in the box.
When the BBC were looking for someone other than Newbury Labs to build a computer to accompany their computer literacy project, Chris Curry was able to attract their attention and get them to come and look at Acorn. This resulted in the famous weekend when Hermann played Sophie and Steve off against each other saying, ‘The BBC are coming on Friday. Can we show them a prototype?’ They hadn’t even got a design at this point, let alone started to build anything. They had a sketch for a thing that they called the Proton, a dual processor successor to the Atom. Steve had got as far as sketching a preliminary circuit diagram, but Hermann’s chicanery persuaded them to go for a prototype and, by the Friday, just, it was working. In parallel, Hermann had got Allen Boothroyd, who had done some of the industrial design for the Atom, to come and mock up a case for the BBC Micro. So five days later when the BBC came, over here they had a prototype running code, and over there a case model. This is what it’ll work like, and that is what it’ll look like. Clearly the fact they had done that in a week was a factor in persuading the BBC to back the Acorn horse.
At their initial discussion with Acorn, the BBC were expressing confidence that 12,000 machines would be sold on the back of their programmes. In the end they sold about one and a half million. Nobody anticipated the public appetite for computing equipment at the time that the discussions were taking place. The BBC Steve thinks, quite rightly, felt their duty to offer education to the wider public, but they had no sense of the extent to which the public would lap this up, and spend real money to back up their interest.
They took the show on the road; Chris Turner, Sophie and Steve initially went to all of these. The first one was set up at the IET (it was then the IET) in Savoy Place. They have a lecture theatre which seats seven or eight hundred in the middle of the building. The first talk they set up there was oversubscribed, he thinks three times the number of people they could allow in turned up, and there were people who booked coaches from Birmingham. The degree of public interest was just phenomenal, and it was a turning point in the development of computers. It was also very unexpected as a lot of them found their way into homes; this was puzzling to many professionals who couldn’t see why anyone would want a computer at home. Initially, it was just satisfying curiosity. People just wanted to familiarise themselves and see what these things would do. Very rapidly of course, other factors came into play, and one of the big ones was computer games.
The BBC Micro was really Steve’s entry into the business. Acorn started off dealing with the BBC but then it became clear that the Government was going to give significant backing for its computer literacy project, and so the machines were going to end up in schools in quite a big way. And then they were going into homes. The whole thing basically just developed its own momentum and there was no sense in which they were pushing. They were more standing back and watching and trying to anticipate what the next move would be, so that they would be ready for that.
The obvious next step from the BBC Micro, which used an 8-bit microprocessor, the 6502, was 16-bit. This was driven by what was becoming known as Moore’s Law. The number of transistors you could put on a chip was doubling every eighteen months to two years so the amount of functionality you could put on there was growing at that rate and going from eight to sixteen bits was a natural progression. They looked at all the 16-bit processors that could be bought off the shelf, and they didn’t like any of them. All of them had microarchitectures which were to some degree based on the very successful minicomputers of the 1970s. The minicomputers had shown how you scale a machine up from quite small, to significant computing power, and the 16-bit microprocessors were trying to emulate that direction of progress. But this had, in Acorn’s eyes, two significant drawbacks, the first one of which was that their real-time performance was very poor. The second, that was equally important, was, they could not make full use of the available memory bandwidth. Acorn had done several tests on a few microprocessors that drove them to the conclusion that the thing that determined the performance of a computer more than anything else was the computer’s ability to access memory bandwidth.
They were scratching their heads about this problem when Hermann dropped some papers particularly from Berkeley, but also from Stanford, on their desks, where Dave Ditzel and Dave Patterson were expounding the virtues of what they called the reduced instruction set computer (RISC). The thesis they were expounding was that the minicomputer style architecture was the wrong answer if the goal was to put the whole processor on a chip. The number of transistors that it took to implement these complex instruction sets did not leave enough transistors to do some more important things. So, in 1983 Acorn decided they would try to design their own microprocessor around the RISC concept. Fed with these published papers, Sophie started playing with the instruction set design. They were encouraged by Hermann who saw this as an opportunity and, as he knew them quite well, if they thought it was an interesting way to go, he trusted and backed them. Sophie did the instruction set architecture design; Steve did the microarchitecture. There were other factors, such as they had employed some experienced chip designers but didn’t really have anything for them to do. So, when Steve sketched a microarchitecture, they were there waiting, and they took the microarchitecture specs and did the silicon implementation. Within eighteen months, on April the 26th 1985, they had the first working ARM chip.
Concurrently with the development of the ARM, Olivetti came on the scene. Acorn had overextended itself in attempting to break into the US market, and the Electron timing didn’t work quite right. Christmas was very important for Electron-like products, and the first Christmas they couldn’t make enough because of problems with the yield of the ULA. For the second Christmas, when they could make huge numbers, suddenly the market had gone. Both of those things cost Acorn a lot of money, and the company was then effectively bust. In the negotiations with Olivetti, they were not allowed to talk to them about the RISC microprocessor, it was kept secret. When Olivetti finally agreed to buy Acorn, they were told about it, and had no idea what it meant, what its significance was. They were apparently quite happy to allow Acorn to continue doing this bizarre thing with their own processor and put it in their own products, and Acorn sold ARM base products, starting with the Archimedes, then developments of that, moving on to the RISC PC in the Nineties.
Acorn’s business had grown exponentially in the first half of the Eighties, but then it went very flat because it was pretty much the case that each year the Government declared its budget for computers in schools, and that was Acorn’s business. They didn’t get all of that, but it set a cap on Acorn’s business. So Acorn stopped growing and it was becoming increasingly hard to maintain the processor development, which was getting more expensive to keep it competitive.
Sometime very soon after that, Apple came knocking on the door at Acorn, and they said they were interested in using the ARM in their Newton product. Apple was a very big name then, as now. But they were not comfortable with ARM being controlled by a competitor, albeit a rather puny one, in the form of Acorn. So, they approached Acorn with the proposition to set up a joint venture. They were all pushing on an open door; in fact, in Steve’s last two years at Acorn, he had spent quite a lot of that time trying to find ways to set up the processor activity as a separate business. So, lots of the groundwork had been done, but they couldn’t find a model that made business sense. Then Apple came and wanted the joint venture and by November that year ARM Limited was set up in an old barn in Swaffham Bulbeck, and the processor was now in independent hands. Robin Saxby was brought in and he introduced this business model that made it all work. He found the trick that Steve had failed to find, and the trick was, the standard licensing business is a royalty business, but royalties come very slowly, and downstream. Saxby introduced the ‘join the club’ model, where you pay your licence fee upfront, quite a big licence fee. And of course, that’s brilliant for cash flow. Royalties are terrible for cash flow, until they’ve really built up. The upfront licence fee is perfect for cash flow. It’s a big chunk of money, very early in the engagement.
In 1990, just before Apple approached Acorn, Steve decided to leave Acorn and moved to take up the ICL chair in Manchester on the 1st of August 1990.
There are two things that he has really contributed a lot to since his move to Manchester. The first was the Amulet microprocessor series. Amulet was a series of research projects, principally funded from the EC ICT programme. It originally was the Open Microprocessor systems Initiative, which Acorn and Inmos were instrumental in encouraging the EC to set up. The Amulet processors were ARM compatible microprocessors, with the fundamental difference from the commercial ARM processors being that they were clock-less. They were designed on a principle where they timed themselves, they did things as fast as they could, and all parts of the circuit interacted. There was no global synchronisation. Asynchronous technology is very interesting. In some sense the role of a clock in a chip is to slow the fast bits down so that the slow bits can keep up, everything must run at the same rate. In asynchronous design, you move away from that straightjacket so you can potentially do things much more energy-efficiently, because you only use activity where you need it, rather than having a clock buzzing away everywhere all the time. Particularly significant is that they have very good radio interference advantages. A clock generates as much interference as possible; by desynchronising you get something which has much lower interference and it is very broad spectrum. There is a global asynchronous design community, Manchester weren’t the only people by any means in this game. However, by the end of the Nineties, it was clear that processors had moved from being manually designed hard cores to being synthesised. The state of tools for synthesising asynchronous circuits was much less developed than the state of tools for clocked circuits. So at that point, they shifted the emphasis a bit from designing the cores by hand to designing tools. Then it proved very hard to compete with industry, because there were huge resources going into conventional tool flows. The Amulet processors occupied them nicely for the Nineties, and they did some nice work, but nobody really managed to make a commercial success out of asynchronous design.
The second was SpiNNaker. The motivation for SpiNNaker is that, although Steve has been designing conventional processors for several decades and they’ve become a thousand times faster, they still struggle to do things which we humans find easy, even when we’re babies. He got a grant from the EPSRC to look at associative memories, which he’s always liked in digital systems but found he was just reinventing neural networks. So, he thought about what can we do as computer engineers to contribute to the world progress in neuroscience, and in computational neuroscience? That led to thinking about what kind of machine could be built that will make a difference? Brains are big, they’re very complex and we each have 100 billion neurons inside our heads with ten to the fifteen connections. Clearly, the machine must be scalable, so how do we make sure we make it scalable? Let’s set a big target. And the target was, what can we do with a million ARM processors? Real-time brain modelling in a single machine. A million was an arbitrary number. It was clear from the outset that with a million you barely get to one per cent of the scale of the human brain, although you had passed a mouse brain by then. The mouse brain is, very conveniently, quite like the human brain but a thousand times smaller.
So, the motivation was, what can we do with a million processors? And, “how do you do the interconnect”, is the big question, because that’s the question that people have wrestled with in the past when they’ve tried to build this kind of machine. They came up with what Steve still thinks is the best answer available to that question and he thinks all the competing neuromorphic platforms that are out there, even from IBM and Intel, are still lacking on the connectivity question. They found a very effective answer, which is the heart of SpiNNaker. They put eighteen processors on a chip, 48 chips on a board, that’s 864 processors on a board, and put 1200 boards in a machine, and that’s a million cores. That’s how you do it, but the innovation in SpiNNaker is how these processors talk to each other, and that’s the interconnect network.
Manchester vs Cambridge
Having worked in both Cambridge and Manchester, two hubs of computing innovation, Steve was asked to compare the two. He has allegiances to both. He was born and brought up in Manchester and has worked there for more than quarter of a century. But equally he spent 20 years in Cambridge, at the university and then at Acorn. He does sense an historic tension between the two places going back to the very early history of computing. Some of that boils down to who really built the first stored-program computer? The historic answer is that the first operational stored-program computer was in Manchester, but it was rather small and prototype-y and not very useful and needed expanding a bit. The first usable stored-program computer was the one Maurice Wilkes built, the EDSAC, in Cambridge. Of course, Williams and Kilburn were not actually trying to build the first stored-program computer at all. They had this idea for memory based on cathode ray tube storage, which they had used during the war, in analogue form, for radar, and their question was, can we now turn this to digital purposes? They built a very simple computer around it, as their idea of the easiest way to test this concept. But Manchester can certainly claim the first operational stored-program computer, and Cambridge can clearly claim the first stored-program computer that could support a sensible user service. Both of those are important.
Since then of course, Manchester has a continuous tradition of building big machines that isn’t evident in Cambridge. Cambridge’s contributions were very diverse; like Andy Hopper’s work with high speed networking, which was highly influential. Manchester kept building machines with Ferranti’s, and then ICL. Cambridge’s influence you can feel in all sorts of other dimensions. Manchester really owns the big machine story.
Steve is still excited by the work that he does and the possibilities that there are. He thinks that computer technology doesn’t just advance, it changes nature. Roughly every decade you get a qualitative change in the role of computers in the world. Although the technology is hitting some fundamental limits – Moore’s Law is now slowing down; Gordon Moore in ’65 reckoned it was good for ten years; it’s lasted 50 – that’s not bad, but it’s not going to last another 50 – there are major innovations happening in technology. To continue delivering more functionality we’ve got to find different answers from simply making the processor go faster, because that doesn’t work anymore.
Steve thinks his grandest mistake is the BBC Micro disk error 14. The story there is, when he was doing ‘my little bit of moonlighting’ he built Acorn’s first floppy disk controller, using the Intel 8271 floppy disk controller. He got this going at home and read the datasheet very carefully, and set up all the parameters, and got it to work. He took it into Acorn, and said, ‘Look, I’ve got a floppy disk controller here.’ They took it over and it found its way into Acorn products, and then into the BBC Micro. Five years later, with a million and a half BBC Micros in the field, there were strange reports of disk error 14 happening rather more than they’d like. So, it was investigated, and it turned out to be his fault, because when he had read the datasheet and set up the parameters, he got one of them wrong. In the five years and one and a half million BBC Micros later, nobody had checked these numbers. So, he doesn’t feel that guilty about it. He was just knocking something up in his back room to prototype.
Interviewed by: Richard Sharpe on the 15th June 2018 at BCS London Office
Transcribed by: Susan Hutton
Abstracted by: Helen Carter