 So we're here in Manchester, and this is the first ARM processor. Hello! Hi, I'm Steve Furber. I'm a professor of computer engineering here in Manchester, and back in the 1980s I was one of the principal designers of the first ARM microprocessors at Acorn Computers in Cambridge. So you designed ARM? Is that what you did? Yeah, I was the microarchitect. So Sophie Wilson designed the instruction set. I designed the microarchitect, which implements that on silicon, and we worked with a small team of VLSI designers who did the physical chip design of the first ARM. So you made the hardware and she made the software, kind of, or not really? Well, the instruction set architecture is the interface that the programmers, the software guys, use to program the machine. The microarchitect is the thing that implements the instruction set architecture and causes those instructions to have their appropriate effect on the values inside the processor, which are typically held in the register file. So here's the first ARM processor. So how do you make this? So what's going on here? What you're seeing there is the physical implementation, so the design of the silicon, and this is actually a photograph taken of one of the first ARM chips. What you're seeing here is this very regular area here, that's the register file. There's a data path that you can see flowing across the chip that way. It's 32 bits wide, bit zero is down here, bit 31 is up at the top. And here there are the ARM registers and there are 16 regularly visible registers, plus a few others used for handling interrupts and special purposes. And then if you look across the data path, the next big structure there, that's the barrel shifter. That's the thing that allows the 32B quantity to be shifted left or right, any number of bits, or rotated around. In this region here's the ALU, which does addition and logical operations and so on. At the left-hand end, this is where addresses come out of the processor and there's some hardware here which can increment the address to compute the next word address. You can see that it's missing the bottom two bits because addresses are, most of the time, are word addresses, so there are multiple of four, although the processor can address individual bytes as well. And you'll see the top six bits are missing at the top end because the very early ARMs did not have the full 32 bit addressing of the later ARMs. They used effectively 26 bits of byte address. And what's going on over here? At the top, the more random area of the chip is where instructions are decoded. So there's what's called a program logic array here, which takes various bits of instruction in and works out how to execute that using the data path. This big structure here is doing the register file decoding. So ARM has the ability to read two registers and write one register in the same cycle. So you need three layers of register decoder there. And then there are special purpose functions at the top. I'm not sure I can even remember what they are individually in order, but there are things that, for example, in a load multiple instruction, you have to search the 16 bottom bits of the instruction to find the ones and use those to access the appropriate registers. So did you come up with this design? Or how did you come up with this design? Was it a team? Like did you sit around the table and say we needed to have it over here and over there? Or was it standard? How do we make a processor? Well, so I took Sophie Wilson's instruction set architecture and looked at the instructions that were defined there, which were inspired by earlier risk work, particularly at Berkeley, on the Berkeley risk one. Risk is ARM's middle name, so it's very much inspired by risk, which stands for reduced instruction set computer. That means it's a computer where the instruction set has been simplified to make it more regular, so that you can use the silicon resource on the chip for things that do more for performance than complex instructions. And typically what you do is implement a pipeline and ARM one was a three stage pipeline machine. So I took this instruction set and I have drawings of how the data flows around in the machine and I literally photocopied these drawings and coloured them in. So for each cycle of each instruction I knew exactly what was happening on each part of the processor. And then I sketched a floor plan. The floor plan is if you like a hand drawing, which is similar to what you see on the final chip. It's not identical because the guys who did the chip design had to make some adjustments to get it to fit right. But I sketched this and then I built a reference model which was 800 and something lines of BBC basic that define exactly how the processor works. And you can run code on this basic program to test it all out. And then we wrote what were called block specs. So the various structures you can see here are called blocks. And there's a two or three page document that defines what the register file has to do that defines what the ALU has to do that defines what each of these blocks has to do. And the VLSI team then took these block specs and the floor plan and implemented these as physical layout and put the chip together. And the whole process from beginning to think about the instruction set architecture to having the full layout of the final chip took about 15 months. We started in about October 1983. We went to Munich to tape out the chip which was fabbed by VLSI technology in California, but we taped out in Munich in one of the coldest winters I can remember. All the taxes around Munich were stuck because their fuel their diesel fuel had frozen. And we taped out in January and the first silicon came back on April the 26th 1985. And by the afternoon it was running BBC basic. And we were cracking our bottle of champagne to celebrate working silicon. So right there in Munich, were you with Sophie Wilson or somebody else? No, Sophie was involved in the front end in the software and the instruction set architecture. I went to Munich with I think Robert Heaton from who led the VLSI design team, Robert and I and possibly I don't remember if other members of the VLSI team, Jamie Erkut was one member Dave Howard. And I'm not quite sure when Harry Oldham joined, but there were several people in the VLSI team that contributed to the layout. But I think it may have just been two or three of us that went to Munich. So do you think it was an important thing? Or were you thinking it was actually not going to work? Like how big is this chip? Is it like how big is the chip? The chip is about eight millimeters square. So it's about 64 square millimeters in total. But in the beginning, when you started to think you had to design a processor, do you think it would be impossible? It's like magic, no? Yeah, we thought that the microprocessor design was a black art when we started off on this and we'd been to visit various companies. For example, we went to Israel to visit National Semiconductors, a design team who were working on the 3216 and they were on RevG or RevH, you know, about the sixth or seventh version. There were still big bugs in the processor because it had a very complex instruction set. It was hard to debug. So we thought that designing microprocessors was a bit of a black art. But then we heard about this risk design that's been done by a postgraduate class at Berkeley and we thought, well, you know, if you can do this with a class of postgraduate students using this risk idea to make things simple, you know, maybe we can do it. And in October, 83, Sophie and I visited the Phoenix Arizona Design Center where they were designing the successor to the 6502, expecting to find the usual sort of glossy American building. But what we found was they were doing it in a bungalow in the suburbs using schoolchildren in the summer vac to design some of the cells for the cell library. And we came away from there thinking, well, if they can design a microprocessor, maybe we can. And we took that thought back to Cambridge and started on the first arm design in earnest. So you skipped, let me turn around over here. You skipped 16-bit. You thought you'd have to go directly to 32-bit. Is that a key also too? Yeah. So we started with the BBC micro, which is on the table here. And that used the 6502, which is an off-the-shelf 8-bit microprocessor. And that was a big success. And it was really the success of the BBC micro that underpinned our ambition to design the arm. But in developing the BBC micro, we started from a design that was designed as a dual processor. And the front end of the dual processor became the BBC micro, but it retained its capability of having a second processor attached. And we attached faster 6502s and ZHs to run CPM, but we also attached 302.016s and 68000s and other 16-bit microprocessors. And we formed the view that the performance of the microprocessor was much more defined by the amount of memory bandwidth it could access than it was defined by its details of its instruction set. And so we figured if we're designing our own microprocessor, if you build it with a 32-bit data bus, you immediately get twice the available memory bandwidth that you get if you built a 16-bit data bus. So why bother with 16? Why don't it go straight to 32? Get double the bandwidth and use that to deliver performance. And we figured that with the arm, we get something like 25 times the memory bandwidth available to the processor, so we get about 25 times the performance. And that turned out to be a pretty good estimate. So when you say we figured this out, so who were you, the team? Who's sitting there? Was a small table like this enough for the whole team? Well, the ACON team was led by Sophie Wilson and myself, with Sophie sort of more on the software and instruction set side and me more on the hardware side. But we had several people in the team. We had our chip design team. We had people like Tudor Brown and Mike Muller who went on to have very senior roles with Arm Ltd. Mike is still CTO at Arm. So he was on a team where you were defining? The Advanced R&D team at ACON was in total about 25 people and I guess sort of five to ten of those were involved in this kind of thinking. So in Cambridge, there was a room somewhere, it was a secret room, right? Like nobody knew you were working on a CPU? It was not a secret room, but the project itself was kept under fairly close wraps. I mean, I think most people in ACON knew it was going on, but the information was not allowed outside the company. And the core team actually working on it were maybe about a dozen people who were directly involved. How did you make sure nobody told anyone about it? Did somebody threaten their families or something? Just joking. No, no, it just made clear that it was a confidential development and people were I can't remember if people asked to sign a piece of paper or not. So Herman Hauser says that one of the reasons he's got successful is because he didn't give you a lot of money and he didn't hire many people. Herman's very good at retrospective analysis and I mean that's a good story and there's some truth in it, but of course that's not what he said at the time. This is thinking back about what it was. Why did ARM turn out to be so successful? Well, a key to its success was its simplicity, small size and low power. Why was it small size and low power? Well, because it was very simple. Why was it simple? Well, because it had to be because we had so few resources with which to build it and had we made it complicated it will probably never have worked. So there is a kind of rationale in that story, but it makes a good story and I think that's why Herman repeats it many times and often. How much did it cost to develop? Is there any way to estimate that? Like is it more than a million pounds? I would say it took about a dozen man years, a bit more than that. So if you think a man year is the order of a hundred thousand pounds with full overhead, I mean that's not what they paid us, but with the full overheads, that is probably the order of a million pounds, yes. So for a million pounds and a dozen man years and a small team, very awesome engineers. So you design something that is now more than 50 billion. So is this design very similar to what we see in 50 billion ARM chips right now? No, this is a full custom design. When you look at this you can see lots of structure and you can see that because humans have basically put everything down on that chip. Modern ARM processors look quite different because they're almost entirely synthesized by tools, so you don't see this nice regular data structure where you can see each bit is the same width all the way across. If you look at the physical layout of a modern ARM processor, you don't see much structure. You see a mess of stuff, but so physically you don't see anything that resembles the original ARM. But the instruction set architecture that defined the first ARM carried on through many generations of ARMs and even today's ARMs, the one we might use in the next generation of Spinnaker, which is a Cortex processor which runs what's called the Thumb2 instruction set, which is a completely redeveloped instruction set. You can still see lots of legacy of the original ARM instruction set in the way that's implemented. So how do you feel about that? Do you feel like every day you feel like awesome about it? Yeah, well I mean the success of ARM has been, as you say, awesome. And of course a little bit of that is to the start we made at Acorn, but a lot of it has been down to business innovation. So when Robin Saxby was brought in to lead the company when it was set up separately from Acorn, he created this idea of the licensing model, which has been the key to its success. And of course ARM has employed thousands of engineers over the years since then who've made the developments, who've understood customer needs, who've tuned the processor to meet customer needs more closely. So the success today almost certainly would not have happened had we screwed up in the 80s and not produced a working chip. So in that sense we sowed the seeds, but the success is down to many thousands of people's efforts over the quarter of a century since. But back then were you also involved and maybe also Sophie Wilson or some other people at Acorn involved in trying to figure out how to spin out ARM and figuring out what should be the business model? Like before Robin Saxby came in, what were you thinking, what was your ideas for the business model? Well the last couple of years I spent at Acorn before I came back to Manchester were some of that time was looking into business models to move the processor development out. I was still leading the processor development at Acorn and it was clear that Acorn's business was not growing fast enough to support processor development at the level that was required to keep it up to date. So we needed to find other sources of funding. The logical thing to do was to separate it from Acorn and sell the process there to other businesses in order to get more funds to keep the processor development moving. But in the business plans that I was involved in, which were led by the Acorn technical director, we could never get the numbers to add up because we were looking at a small processor where you might charge a modest royalty to include that processor on a chip and you'd have to sell millions of those processors for these royalties to add up to anything. And the other problem of course is that royalties are difficult for cash flow because they come a long time downstream and they start small. The magic of Robin Saxby's model was adding to the royalty stream the idea that people paid a license fee to join the club. So what you got was a significant chunk of money up front and of course that's great for cash flow because it's a big chunk of money and it's very early. And the royalties come later but you have to live off something while you wait for the royalties to start flowing. And this company right there, there's Apple came over and did the Newton and they kind of brought all the cash into the beginning of the spinout? Well the thing that made ARM happen as a company was just after I'd left Apple came knocking at Acorn's door saying they wanted to use the ARM in their Newton product. The Newton was in planning and design at the time. They'd been working for some time with AT&T and the Hobbit processor but they decided that the ARM was a better processor. But they didn't want the ARM owned by a competitor. They saw Acorn as a competitor in producing desktop machines so they approached Acorn about setting up ARM as a separate company as a joint venture. And that's what caused ARM finally to happen. The joint venture was based on the ARM, the Acorn team. So the people I led at Acorn became the dozen founders of ARM. Robin Saxby was brought in as chief executive to lead this. Apple put in a small amount of money and they became equal owners of the joint venture with Acorn. And VLSI technology put in the chip design tools and they became a third owner at a slightly smaller level than Acorn and Apple. And it's important to remember that the Apple Newton was vital to the early success of ARM because then as now Apple was a huge brand worldwide and the Apple brand name opened doors. By the time a few years later it was clear that Newton was not going to be a successful product. It didn't matter so much. But in the early days of ARM the Apple Association was vital to them making progress. Of course the real turning point was in the mid 90s when they got the Nokia business which allowed them to get their independence from their original owners of Apple and Acorn and get hooked into a very big rapid growing business of mobile phones. And this even saved Apple somehow because they were in money problems. They sold their armchairs. They could survive with that. And then all of their profits in the last 15 years has come from ARM part devices not from Intel part devices. Even though you have a Mac here over there but most of the profits is iPhone. Yeah I mean Apple has done very well out of ARM as you say. They got a chunk of cash because when they were in financial difficulties their arm shareholding was worth quite a lot of money and that helped them. And the Newton was followed 10 years later by the iPod and then the iPhone which turned into great successes. In some sense the Newton was about a decade too early. The technology wasn't ready. And when they came back and revisited the concept 10 years later they turned it into a huge success based very solidly on ARM technology. Alright so you're from Manchester and you went to Cambridge to study right? Can you show this? What is this? So this thing here is the first computer I built in about 1977. I was a PhD student. I was studying for a PhD in Aerodynamics and I heard about some people who were forming a club of computer hobbyists. It was called the Cambridge University Processor Group. And it was formed not by me although I went to the very first meeting. It was formed by people who built computers for fun. And I thought this sounded interesting. I was very interested in flying which led to my PhD in Aerodynamics. But I figured that simulation was probably a good way to go if you like flying. So I had this vision of learning about computer design and building some kind of flight simulator. And this machine was the first one I built. It was based around a Signetics 2650 microprocessor which almost nobody has heard of. Which somewhere in there? It's somewhere in here. I don't know if I can find it. This is built on boards. Yeah that's the 2650. It looks like most of the other microprocessors. So what's the performance of that compared to? Well this was something that would run at the order of a megahertz. Quite similar in era to the 6502 in the BBC Micro. This came a few years later so we run the 6502 in here at two megahertz. But the 2650 had things like an internal hardware stack for subroutine return. It had some nice features but got very little commercial use. Did you design the PCB? Well this is built on a standard Vera board. You can see the boards are all the same in here. And the components are put into sockets which are soldered into the board. And then they're wired together with these very fine Vera wire systems. So this was a wiring pen. You wrap the wire around one end. You wound it round through these combs to where you wanted it to go. You wrapped it around that end. Then you soldered it. And the wire had a very thin layer of insulation on and when you soldered it it melted the insulation and formed the connection. So it's all hand wired. And these boards were designed to take this kind of wiring system. Then you have a whole bunch of them and that kind of like is one system like this. Yeah so various bits of this were built at various times. It was expanded. None of it works anymore because bits have been plundered from all over to build successive bits. That's the actual one you built. Yeah this is I mean it's in a card frame but the card frame was hand built. I bought angle aluminium from you know a hardware supplier and soldered it and filed it and sanded it and built these are standard card rails which are just screwed into it. And the power supply is something that was built by hand and has an enormous capacitor in whose precise details I can't remember. And it started off in this form as you can see with a set of switches and LEDs and to load stuff into it you'd set up an 8-bit byte on those switches and then you'd write it into an address and it would auto increment the address and then you could set the next byte up and load it in byte at a time. A bit after this I gave it a simple operating system and an external keyboard and replaced the front panel by something that's much more boring that's just a plain panel with a single socket to connect to keyboard and it would connect to a TV display. So that was in Cambridge? Yeah. And was it like a nice big computer lab? Lots of what's called geeks? No I did this in my student lodging. I was living in 7th St John's Road which was a terraced house with three doctoral students living in it and it had an attic where nobody slept and that kind of became my informal workshop. And how do you buy all these components? You buy them mail order, the microprocessor I bought mail order from California using my credit card which is very scary as a student with very limited resources. The Vero board I think was from a UK company and the aluminium was just from a hardware store. You can see the ends of the box are just covered in sticky phablon that's a bit a bit blue peterish sticky back plastic. And the front panel was made just from sheet aluminium. I cut the holes and I anodised it in the kitchen sink and used a letter set to label it and then sealed the letter set with the kind of varnish of some sort. So it's all very much hobbyist home built stuff. So did Herman Hauser or somebody else see you do this and then you joined the team? How did it work? Yeah so when Herman and Chris Curry were thinking of starting Acorn they saw the Cambridge University processor group as a natural source of people who knew something about microprocessors and I'm not sure how they identified individuals but Herman approached me and invited me to a meeting in the Fort St George on Midsomer Common where I met Herman, Chris Curry and Chris Turner to talk about starting this company up and Herman said would I be interested in getting involved and my answer was well you know this is a hobby I'd have no professional qualifications in this game at all. I just built stuff by hand but if you think I can help I'm happy to get involved and that's where it started and for the next two or three years I was very much part-time I mean I was a PhD student and then a research fellow in the university just visiting Acorn you know one or two evenings a week out of interest and I built stuff like this for fun and where it was potentially the basis of a product they took it took the design and in exchange they gave me some bought bits to go build stuff with. Did they give you a salary? No no I was employed first on a PhD studentship which didn't allow me to undertake any additional employment and then I was a research fellow so in those positions I just took bits. The computers I built were useful for my research so not this one but the one I built after this which formed the basis ultimately of the BBC Micro it's built in a it's a double rack I used that for data logging on my university experiments I also used it for writing my PhD thesis so this was an area where most PhD students hand wrote their thesis and got a professional typist to type it up. I built a computer then I wrote a text editor and I wrote my thesis on this text editor and printed it on the a daisy wheel printer which were quite newfangled at the time these were printers where you had a sort of plastic daisy like wheel with characters at the end of each of the petals of this daisy and this wheel spun round and then a hammer just knocked the right character against the paper and because my PhD thesis was mathematical it contains greek so I had a twin track daisy wheel that had two of these heads on it for printing roman and greek and the thesis came out as about 100 feet of continuous paper which I then had to cut up to bind as a thesis. So it was a fun time to to tinker around in Cambridge? Yeah I thought building computers was great fun and it was perhaps an unusual time in the sense that the microprocessor had emerged so building a working computer was much simpler than it had been previously but the clock rates were still at levels where sort of hand wiring on the circuit board would produce a working system. I mean today with processors running with clocks in the gigahertz range you couldn't possibly build anything by hand you have to have a very professionally designed printed circuit board to make something that works at all and so the early 1980s was probably the one time in history where designing and building a computer was within easy reach of an average hobbyist. So that was a lot of the business that happened with the actually with the Sinclair and were you also involved a little bit with the Sinclair what they were doing before you acorn kind of started? Only only very peripherally and it wasn't directly with Sinclair it was with Science of Cambridge which was a company that he's that Clive Sinclair set up with Chris Curry when Clive was still a bit dubious about the prospects with the computer business but Chris persuaded him to have this little foray with Science of Cambridge and they bought effectively a design which was a development kit from National Semiconductor for the SCMP microprocessor and they modified this a bit to turn it into a hobbyist product and I hand-wired the first prototype of that using the wiring technology that I used in this machine so it was hand-wired and checked out and debugged before. That was actually a product that it was selling? No no this was a prototype before they before they designed the bespoke PCB to make the product. The products were always PCB based they were never based this hand wiring is very laborious and therefore expensive if you do it commercially. And that was very popular stuff and people they had soldering irons all over the UK and they were doing this kind of stuff? Yeah very popular as a relative okay. The first Acorn boxed product I mean we built the system one which was a Eurocard product which was a hexadecimal keypad this was Sophie Wilson's design and a seven segment display but the first product with the keyboard on was the Acorn atom and we did sell that as a kit and that was one where you needed to have a soldering iron and you could solder the components in there would probably be you know 30 40 components to solder so it's quite challenging and we started to get things returned that didn't work and one story I remember was one that came back where with a letter which said you know I know that components don't like heat and the instruction said solder but because they don't like heat I use glue instead and it still doesn't work. It was at this point we realized that the potential market for computers was probably bigger than the number of people who knew how to use a soldering iron and so towards the end of the atom product life we moved to selling a manufactured product and the BBC Micro was always manufactured we never sold the Bebe as a kit. But so you designed or you're part of the team but you designed the first arm but you also kind of designed the future arm employees they were all the soldering irons they were hacking on this stuff maybe later they joined arm right there's lots of UK engineers that came from that. Yeah I think the biggest impact on UK overall was the BBC Micro and there are many people who will say they learnt to program on the BBC Micro was the machine that introduced them to programming and that influenced their whole career. In particular quite a lot of the UK computer games industry bootstrapped off the BBC Micro and that became a very important industry in its own right. Now of course at the same time there were the Sinclair products and the spectrum was in some sense a cheaper games machine but I think the programmers preferred the BBC Micro with its sort of you know full moving keyboard. It was quite difficult to do real programming on a spectrum because the keyboard was very thin. This is a good keyboard right here and they sold over a million. We sold about one and a half million BBC Micro's in total. The keyboard is phenomenally good seriously over engineered I would say by today's standards. You could use it now. Yeah yeah each of these switches is an independently manufactured switch. Modern keyboards are a sheet of elastomerics. These are individual key switches. Each of these switches contains gold wires and when the keys raised a piece of plastic pushes the gold wires apart when you press the key the gold wires spring together and make the contact. It's very reliable. This kind of keyboard would survive at least the decade of primary school children hammering away at games and possibly even educational programs. It was a really robust keyboard. It was a very good reputation for most of the 1.5 million BBC Micro still work today. Oh I'm not sure you'll find many that work today. There are there are perhaps a surprising number. What breaks? What was the thing that broke? So the the first thing that breaks is the power supply. This side of the box well the cable is fairly messy but that's actually because it's this particular one that's spent most of its life in the expanded polystyrene packaging and you can see that the expanded polystyrene and the rubber don't get on very well. So that's particularly horrible but the power supply which the very first BBs were sold with a linear power supply but it really wasn't possible to get a linear power supply efficient enough in this profile. So very early on we switched to a switch mode power supply which was manufactured by Aztec in Hong Kong and the first thing that fails on most BBC Micro's are the capacitors in the power supply. After about 15-20 years the electrolyte in the capacitors begins to dry up and then the capacitors fail. I'm told by the retro computing community that if you have a BB that doesn't work if you replace those capacitors it'll probably work for another 20 years after that. So that seems to be the only component whose life is measured in decades. The rest of it seems good for a century. So as a microman is that a would you like this this show? I really enjoyed Microman. I mean it's not a highly accurate historical account it's a drama documentary and most aspects of the story are based on true anecdotes but they're all played up a bit. So for example the first prototype BBC Micro started working about three hours before the BBC arrived on the Friday morning not as shown in Microman five minutes after they arrived. So it wasn't quite as cliff edge as Microman purports and of course the actor who plays me Sam Phillips in Microman is shown as a chain smoker wearing glasses with a beard. I haven't touched a cigarette since I was 12 and I've never had a beard my my my hair doesn't grow right to grow a sensible beard anyway and I only started wearing glasses in my in my mid to late 40s. How about the other guys in this in the movie what do they think about their actors like Sophie Wilson, Hermann Hauser? I think most people enjoyed the movie I mean the the main thing is that they managed to find a good story okay so they told a good story so it was entertaining to watch and it was you know it was it was most of the things in there you would recognize if you were there you'd say well yeah that happened perhaps not quite like that but um but that sort of thing so the story about you know Hermann ringing Sophie and then ringing me and then ringing Sophie again I mean that that actually happened. Did they consult with you before they were writing the script and stuff or was it from books or oh yes the the the production team came around and you've bought me lunch so so they they they did their their homework I think they talked to everybody and they got so so all the accounts they got from individuals who were there they also told me that they got conflicting stories as well which is not surprising considering the amount of time that had passed um so is Clive Sinclair was an awesome guy like funny so I have never met Clive Sinclair in person never even after no no I've never been in the same room as Clive Sinclair to my knowledge so so I can't speak from personal opinion as to what Clive was like but Hermann Hauser and Chris Curry all these guys they were yeah the the actor that played Hermann I thought uh was was was was quite convincing I mean nobody's exactly the way they're portrayed but um and um the guy who played Clive who's whose name I should remember Martin um the actor that plays yeah that played Chris Curry I mean yeah um who also yeah he's in a lot of movies right now he's you know he was in the Hitchhiker's Guide he's been in the water stuff and and he's also played Dr Watson in uh in some is it accurate the way he played the Chris Curry I I thought he played Chris a little bit straight okay Chris Chris has a slightly more cunning side than the way he was portrayed he was portrayed portrayed as the kind of young innocent enthusiast and he's actually a he's a bit slier than that um so I I felt they slightly missed some of Chris Curry's more interesting personality features um but uh this is you know this is minor and the key thing is they they told a story and they they they they made it exciting which was great so the UK is is pretty awesome right and Cambridge is super awesome and this is a BBC micro but uh there was Apple and IBM and all did we thinking about trying to compete with those big guys why didn't you try to kill Apple or something like that was that something that was wrong a little bit at Acorn that you didn't focus on on itself fighting with uh uh maybe you should have joined up with Sinclair then the UK would have been the big boss today you know well um it is almost certainly true that we focus too much on competition with Sinclair who was local in Cambridge too and uh didn't think enough about the international position um Acorn the BBC micro grew very much on um on the UK education market we had some overseas markets uh Holland Germany Australia New Zealand were quite strong um but we did try uh to break into the American market but that was very difficult um partly of course because it's Apple's home territory but also because their their electronic product regulation regime was quite different and to get into the US market at all the BBC micro had to be modified to pass the FCC regulations and and the way the BBC micro is designed made that very difficult um because they had a lot of radio interference tests and you had to plug wires into all of these sockets the BBC micro had a lot of sockets um and and getting it through FCC required that you had screened earth cables and that you put the main circuit board inside a metal box and so on so the product became very heavy and rather clunky um and then we were taking on Apple on their home territory and so so that was always going to be a very hard game but I think there were some very important lessons learned from that experience um and those lessons fed through into ARM and and so from the outset ARM had the objective of becoming a global standard so ARM knew when they set off that their business was global if it was going to succeed uh and and that's a very important thing to understand at the outset acorn started off being UK and then being a few other places um and and really its global ambitions were very slow to develop and by the time they started emerging they were too late so how did you how did the BBC micro compare with whatever Apple had at the time were you like pretty good or were you slower or well I we were using the same microprocessor we were running it twice as fast um I would say we had significantly more sophisticated sort of operating system code in the machine um we had more sophisticated peripheral support so the BBC micro for instance um this was before there was common access to the internet and it had this tele software system whereby the BBC could broadcast programs um in the teletext lines of the old analog tv system and you could get an adapter for the BBC micro that would pick these up in the middle of the night and you get software downloaded so effectively there was a software broadcast mechanism we had a lot of that kind of thing uh that made the product very sophisticated um quite expensive uh quite difficult to support um but but very capable in in in terms of what it could do but expensive but apple was expensive also more expensive well the apple the apple 2 which was the product at the time was very expensive in the UK yes and even in the US it was expensive so you had there was an opportunity for you to go there and kill apple well we weren't I mean I don't remember the exact prices at the time but because of the FCC requirement we were pretty expensive in the States as well um so there were these meetings with the Hermann Hauser and you and others were you thinking should we try US should we try to compete with Apple or you actually never even thought about going there firstly I was principally on the technology side so I was I was an engineer responsible for the design of the product with with relatively little involvement on the marketing side um so we might hear about or or beyond the peripheral periphery of some discussion about some grand plan to conquer the world but that was not our concern our concern was to design the right product and and and keep that as competitive and as as performant as we could and and you know the BBC Micro was always about quality and performance rather than ultimate low cost we did from after the BBC Micro we did go on to design the electron which was really a kind of Chris Curry initiative to try and compete more directly with his old boss at Sinclair and and and that was not hugely successful because it wasn't really the acorn way to do things it was its functionality had to be cut back from the full BBC Micro and you know cutting back functionality was the wrong way to go but marketing was very difficult there was a lot of competition at the time the IBM PC had begun to gain traction early in the life of the BBC Micro so life was never going to be easy so uh but I mean arm came out of this so that was awesome right so so uh was it the Herman Hauser who got the like what what did Herman Hauser do every day uh at that point when when you were thinking about spinning out or was that the like the big priority well arm arm was designed in the early 80s the issue of spinning arm out came in the late 80s quite a long time later where did arm come from well it came from the fact that we we clearly needed to design a successor to the BBC Micro which was going upwards in performance there was you know some kind of capability roadmap and we spent quite a lot of time looking at the microprocessors we could buy the 68000 3016 Intel 286 and we didn't like them okay and we tested these things and they all had features we didn't like the two principal things we didn't like were firstly they all had pretty poor real-time response and this was because they were based on this concept of a complex instruction set and the the the sentiment at the time was closing the semantic gap between the instruction set architecture and the compiler so taking the instruction set architecture up towards the compiler by making it more complex giving it higher level functions but this compromised the real-time performance of these systems on the Bebe we were used to handling sort of every byte from a floppy disk with an interrupt you couldn't do that with the 16032 because it had for example a memory memory divide instruction which took 360 clock cycles to complete the clock was six megahertz so that was 60 microseconds while a single density floppy disk delivers a byte every 64 microseconds by this stage there were dual density disks that delivered a byte every 32 microseconds so these complex instruction sets were really hurting the ability of these processes to handle real-time IO we didn't like that and secondly and more fundamentally we adopted this thesis that the key to performance was accessing memory bandwidth and these microprocessors could not fully exploit the memory bandwidth that was available in standard commodity DRAM and we thought that was a big mistake because the memory is the most expensive part of the system it's the processor's duty to make the best use of that memory and these microprocessors failed in that duty so we were scratching our heads not sure what to do when I think it was Herman dropped some papers about the Berkeley RISC work on our desks and said you know here's some people designing a simple microprocessor and and and that's what stimulated the idea and and it started with with Sophie doodling instruction set architecture we then had this visit to phoenix to see the 65c812 development activity we came back from there thinking we could design a microprocessor I then started on the micro architecture all the time we thought well this RISC idea it's clearly a good idea okay it's so clearly a good idea that industry is bound to pick up on it and when industry picks up on this idea we'll just get trampled underfoot because we're tiny in comparison but at least in the course of attempting this design we'll learn something so when industry starts selling these microprocessors we'll know what to choose so that was really the pretext upon which we started sketching out the design of the arm the other factor which will sound quite strange is that acorn had decided it needed to have a silicon design capability Herman was saying in the future there'll only be two types of computer companies those that have learned to design on silicon and those that have gone bust and and he was determined that acorn would be in the first camp so we'd with Andy Hopper's guidance we'd trolled the world to identify tools to design chips we'd recruited chip designers but they didn't have any work to do okay we basically bought a capability without knowing what to do with it so as the sketch for the arm came off the drawing pad firstly from Sophie and then from me we just gave it to these chip designers and they were free so they just implemented the chip and so in 18 months we had working silicon now everybody was surprised when the first arm came back functional I mean we were surprised when I rang up a journalist with whom I had had quite a lot of contact and said we designed a microprocessor he simply refused to believe me because it was secret because we kept the project secret up to that point we decided I think in July 85 so three months after we got the microprocessor that it was time to tell the world and the world wasn't listening because they didn't believe we could have done this but you know we have the physical proof we built the second processor again exploiting the BBC micro's ability for you to add a second processor through the tube interface we could get a four megabyte arm development system up and running very quickly so the first arm was connecting to the tube yeah so if you look at the bottom the BBC micro it has this odd row of connectors along here and the one at this end which says tube that's where you could connect a ribbon cable it could be about six inches long and you put the second processor in a in a matching box right there and then the 6502 in here would sort of run the operating system and do the IO and disk access for you and your second processor could simply talk over the tube and that's how the first arm system was brought up it was a very easy way of testing out every microprocessor on the planet just about we built prototype boards for all of them and plug them in there and saw what they did so can we bring back the arm processor on the screen over there okay my computer's paying attention actually appears to be I have to type my password in all right so so why didn't Intel think of of what didn't they do risk or some who was the other big guys you were thinking we're going to make this well so so industry's reaction to risk was quite interesting firstly they were they were culturally in this path of building richer instruction sets closing the semantic gap and they were convinced this was the right idea so they were very resistant to the idea that it was completely the wrong idea they were resistant to the risk idea which is to go the other way simplify the instruction set give the compiler more work to do quite deliberately but just make a very good job of implementing the elemental components from which the compiler could build these more complex processes and and big industry was very resistant but in the second half of the eighties it began to realize what was happening so not particularly because of ARM but because of other companies who picks up on the risk idea and were building not microprocessors but but the computers made out of discrete components so who are they it's not the MIPS that was later I was worried you might ask me that question pyramid is the name that springs to mind but I'm not sure I've got that right was it a conflict with their business model the whole risk idea because they want to sell something expensive they don't want to sell something cheap well that's why they never that may have been a component but of course in the end they all adopted risk in some form so Motorola built the 88k in the in the in the later eighties Intel built a processor whose number I can't was it the i960 or something they built a risk processor that was not very successful national semi didn't go that way what they saw was was the MIPS company which was the Stanford emerged from Stanford with Stanford design which started off as a slightly odd risk concept but became if you like the pure risk processor very clean with some drawbacks as a result of being very clean and in particular when they were competing head on with ARM the issue was code density the very pure risk actually gave you significantly worse code density than arms slightly better balanced I think risk instruction set so how about those those initial maybe phd students or professors who were working on that risk idea in the beginning what are they thinking about you did they know about you I think they were probably very unaware of us I mean we were here in the UK you know most of the risk interest and activity was happening in Silicon Valley we kept fairly quiet we had our links with VLSI technology and they did actually license the ARM from acorn in 1987 and indeed one of its first uses was in a radius graphics accelerator for the apple 2 so you get a radius card to plug into your apple 2 and that used ARM for graphics applications but these things were not very visible I think so even apple maybe didn't really know there was an arm in their system well they probably knew but they would have been aware of it but the risk idea which was first expounded by Dave Patterson and Dave Ditzel at Berkeley I think Dave Ditzel went on to influence the AT&T activity and the Hobbit was another risk processor and that's what apple started off designing the Newton around and so there were there were new companies new startups that had picked on to the risk idea but the established industry was much more resistant how about those people in Phoenix the western the western design center what did they think about what they think oh those guys they came over now they like did this so I have no idea what they thought I mean we we did use the 65c816 in some later acorn products I think the when the 64k memory space for the bbc micro became a real headache we adopted the 816 to to give us a 24-bit addressable memory range and memory bandwidth is still the most important thing for all ARM processors right for all they need to access more and more memory bandwidth to have more performance it's actually more more it's more important than the ARM core itself yeah it's a but it's a universal truth of computing I think to a good approximation that the the amount of performance you can deliver is the amount of memory bandwidth the processor can access and exploit now there are architectural techniques that decouple the processor a bit so cache memory is a technique that we first implemented on the ARM 3 and and this makes the processor less directly reliant on the main memory bandwidth but the cache only papers over a certain number of cracks and if you look at at modern processors I mean the bandwidth that the main memory delivers to the processor is now formidable compared with what we had in the in the early to mid 1980s and of course what another thing the cache does is it it takes a lot of random accesses from the processor and joins them up so that you can use your memory in a much more efficient block transfer mode so you move bigger chunks out of main memory into cache and modern high-end processors have many levels of cache so you'll talk about L0, L1, L2 and L3 cache and and all of these levels have slightly different characteristics and they're all trying to cover up the fact that the processor is now using memory at a much higher rate than the main memory can deliver data so the idea of memory bandwidth optimizing for that jumping to 32 bit all these design ideas that was the first ARM right here and now we are at let's say 60 billion we're not sure the last number it was they certainly passed 75 billion 75 yeah 75 billion should we say more than half of those are very similar to like there's some modern arms and some not so modern arms or well the the growth of ARM shipments has been exponential okay so of the 75 billion something like 15 billion will have been shipped in the last year so that's 20 percent of all the arms ever shipped were shipped in the last year and previous year was probably 12 billion so the the great majority of those are relatively small arms they're they're sort of at the three or six stage pipeline level they're they're quite modest in performance and in cost they're in small microcontrollers and you know even if you think of of the Apple products the iPhone there are probably about 10 ARM processes in iPhone of which only a couple of the high end application processors which are quite complex and and and large in terms of silicon area the majority of them are doing very small and simple things such as the motion processor that continuously collects the sensor data even when the phone itself is basically off you've got these little processors managing the batteries managing the the radio communications the sensors the power management everything yeah yeah and and uh so maybe many of those like the eight remaining ones might be similar to all of them somehow but similar to what you were designing back then they are they have a heritage but but they've all been redesigned several times you know we're talking 30 years ago to the to the first arm um everything that's in use today has been redesigned and as I say they're all now synthesized and when you build a process of using synthesis you actually need to do some architectural changes you can't simply take the original manual design which exploited the ability of human designers to build very elegant regular structures that are quite efficient and when you synthesize the synthesizer has different skills it can do different optimizations but it can never produce this very tight coherent physical layout that a human designer can produce so there's a 25 000 transistors is that what you said it's it's 25k transistors yeah in the arm one yes so now they have billions well depends which one you talk about the the the the simplest arm processor the m0 I think is still around 8 000 gates so it's very similar in terms of complexity the 25 000 transistors on arm one correspond to about 8 000 logic gates the high-end ones will use many hundreds of thousands of transistors and if you include their cache memories I mean memory has a huge transistor count then you'll get up into the tens or hundreds of millions of transistors are the chip designers the smartest people in the world that's one thing I'm wondering is it more complex than rocket science and does it make any sense I think all all areas of human endeavor where people specialize they become very good at what they do chip design is is a very demanding discipline because the economics are such that getting a chip fabricated is very expensive making the mask set so that you can print a new arm processor on a new process technology you you spend several million dollars just to get the prototype made and then of course you can stamp them out very cheaply but the masks are very high precision pieces of equipment so the the the gamble in getting the thing right first time is huge and and the mentality of designers who can achieve you know a 100 million transistor chip and get it to come out of the fab working first time that's a very disciplined area of engineering it's a bit like you know writing a huge software program and getting it right first time without ever testing it and now it's more and more expensive right now the nanometers are smaller and smaller each each new generation is even more expensive than before maybe yes the costs of fabricating a chipper rising exponentially the cost of designing a chipper continuing to rise because they're more complex so you need more design resource to get the design complete so here at the Manchester University are you are you what's called the educating some of these guys there are in girls there are the chip designers of now in the future yes sure and and quite a lot of our graduates of now work at arm so um in fact the current CEO of arms simon sagas did a master's thesis with me in the mid 90s now he wasn't in Manchester he was working at arm at the time he he did it as an external industrial research masters but several of the people who who work for arm have graduated some have done phd's with me some have just done the undergraduate degree and gone on to arm but the the kind of thing we're doing now so you came back to to Manchester 1990 I came back in 1990 yes I was at acorn through the 1980s I was born and brought up in Manchester I came back in 1990 to this to the ICL chair of computer engineering and I've been here for the quarter of a century since teaching about chip design teaching yes we certainly teach chip design we don't just teach it we do it so the the spinnaker system that's sitting on my desk this is this is a machine this is a machine we've built for real-time brain modeling applications and each of these black packages with the attractive logo on contained a chip we've designed here in fact they contain two chips they contain a chip with 18 arm processors on and they contain a separate memory die which is an industry standard product and the the spinnaker chip we designed has a hundred million transistors on it that's awesome so there's all arm processors but very special ones well no they're standard arm processes they're in fact they're quite old arm processor designs but there are 864 arm processors on this board well all running in parallel and it's the way these processors connect which is the innovation in the spinnaker machine the processors themselves are very standard so can we walk around the university and you can show show more about this yes I can certainly do that so this is your office right here this is this is my office well I've been not all the time I've been in Manchester I started off in the other building and then I was head of department for a bit but I've been here quite a long time all right so here we see a spinnaker that's a spinnaker board that's under test at the moment I think well currently it's not actually powered on if it were powered on it will be under test so this is uh does it have to do with uh uh uh singularity the future of uh uh computers taking over humans or is nothing no no because you already designed the I'm a singularity skeptic I don't believe in the singularity because uh I don't think human intelligence is the kind of simple parameter that can be amplified in the way that the singularity assumption uh believes it can um because you already designed the uh uh most of the processing power in the world so maybe you also trying to no I'm just joking but me and me and a few thousand others yes so maybe you can introduce uh so this is Simon who's how long have you been here Simon seven years oh hello sitting here seven years you're being filmed yes I guess there's who's been in the department I feel I've been here since before records began I'm afraid what do you do what do I do I know I know I know a sister that's a firmer in the running of the group that I'm I did I did have an academic post here but I have retired but I've continued now three days a week just trying to help to run the group so it has tried retiring a couple of times was it was it it never quite worked so what are we looking at there is this uh that's um that's actually an earlier design that's actually an earlier design that the group did in the end of the 1990s that's amulet three and this is an arm nine class processor um but it differs from the commercial arm nine in that it's fully asynchronous in other words it doesn't have a clock that defines its throughput it just processes stuff as fast as the stuff will flow so there's no global synchronizing clock in any of this so the purpose of this is already originally what you led onto a spinnaker later it yes it precedes spinnaker um we we did a lot of work in the 90s on asynchronous design that was our major research focus and spinnaker uses a lot of that asynchronous design know how but but asynchronous design is not its headline feature um it's uh it's its application in building real-time neural networks that are its headline so uh you have uh let's go around so how many students do you have working in your uh cpu group oh in the group um there are about 80 people um upstairs working on various aspects of asynchronous design right um why are we heading now maybe let's go outside just to see the building a little bit you want to head outside okay so um Manchester University is has a lot of different things this the cpu group is just uh is it called the cpu group we call ourselves the advanced processes technologies group and and we cover a range of process related activities uh from um silicon design up to parallel computing uh this is not part of the group this is a clean room um this is where the work that led to the graphene Nobel prize was done uh so you'll see the the big poster advertising graphene we set up this nanotechnology lab jointly with with our physics department and the Nobel prize was awarded to Andrei Gallem and Kostya Novostvilov um back in 2010 nice and this is on the ground floor of the same building so there's interesting stuff going on so what are all these buildings around here so um this is the Alan Turing building which is uh where the math department is is home um Alan Turing department so he invented the computer right it's the Alan Turing building well Alan Turing did fundamental work in the 1930s on computer ability and he as a thought experiment um he created this concept of the universal computing machine and and he produced important theoretical results he was a mathematician and he produced important results about um what was and what is and isn't computable and uh these ideas led to the implementation of the programmable computer which again the first working programmable computer was built in Manchester and ran its first program on June the 21st in 1948 which is still before i was born but not by much that was that was that was a secret also that was a big secret no no they didn't really keep it uh the Manchester Bayview wasn't secret that was that was commercial no it's the wartime encryption work at Blessley was very secret yeah and because Manchester had the first machine that implemented his big idea from the 1930s Turing came to Manchester after that machine was running he didn't design it he came here to use it um and then he was in Manchester until he died in uh 1954 if he had lived a little bit more do you think the world would have been different do you think he would have launched uh acorn like 20 years earlier i don't think Turing was an entrepreneur and he wasn't an engineer maybe he had met somebody like Herman Hauser back then and started the european silicon valley before they started doing stuff these are imponderables that you speculate on one thing that Turing did when he was in Manchester i mean a lot of his research in Manchester was actually on biology he was very interested in morphogenesis which is the mathematics and symmetry breaking we all start as a single cell and that split into two identical cells and they split and they split at some point those cells have to differentiate and the mathematics of the instability that leads to the differentiation was was Turing's main research interest in Manchester but he also wrote a paper on computing machinery and intelligence where he starts with the words i propose to consider the question can machines think and he turns that around he says that's not a very well defined research question he turns it around to what into what he called the imitation game which of course is what hollywood called their movie on his life um but we simply know the imitation game as the Turing test for artificial intelligence and he speculated a lot only two years after the first programmable computer around its first program here in Manchester he was speculating about how far artificial intelligence might one day go are you trying to emulate artificial intelligence in the spinnaker project no spinnaker is specifically not about artificial intelligence it's about trying to understand natural intelligence my view is that artificial intelligence has proved much harder than Turing thought and many others have thought since because we've never really understood natural intelligence and and the the foundation of natural intelligence is the brain and we don't know how it works so spinnaker is a machine that's built specifically to help address the question of how does the brain handle information processing so it's research trying to understand the brain it's not a machine you want to mass produce no it's not a product a consumer consumer product it's a scientific research platform and in that role it's offered openly through the human brain project so we have in the Kilburn building we have a half a million core spinnaker machine occupying several large works yeah just just in that building can we go there we can go there you have the key uh no i'll have to collect the key on the way through okay um and anybody in the world can run jobs on spinnaker it's half a million armed processors it's freely available yeah half a million right so we're back in the it building yeah i'm calling i hope somebody's in in the lab otherwise we'll have to go back to my office to get a key to have to get the keys that's finished is it a lot of power consumption under the half a million arm processors um it's potentially reasonable it's uh the machine will consume up to 50 kilowatts if it's busy so it's not like cern it's not like a large hydrant collection no no it's not it's way way way down on even the typical supercomputer an arm is supporting uh helping youtube make this heavy they give us free free use of their cause for research use can i film in here we're always open we're lucky right so there's a you can see some spinnaker well all the all these students over there can we we check them out well i don't know if you can check them out what what do they do so this is this is the main lab area the spinnaker team is right that's a spinnaker board mounted on a robot designed by collaborators the technical university of munich so how about these guys what what are these are these are research students working on the spinnaker group working on some uh some programming no okay sorry sorry just jump in like this all right okay but you have to go on to sign a piece of paper if you want to right so they're they're your team right yeah they're the team that are here at the moment and uh if you want to see the machine then it's it's a fair walk over to the kilburn building okay so the machine is running on spinnaker one yes and that's uh 18 core arm arm nine is that what it is yes the the particular processor unit we use is the the arm nine six eight which is an arm nine tdmi core with thumb and signal processing support and uh tightly coupled memories the programs we can run are limited to 32 kilobytes um so it's a small code space people have to remember how to write small programs to run to program spinnaker um although the new neuroscience users don't see the machine at that level at all they they describe their networks in a high level language so this is this is the kilburn building kilburn building it has a nice garden out there which is about to be demolished to build more more office space no it's it's to allow allow light through into the floor below to build the spring pool no it was it used to be upon okay they decided keeping a large amount of water right on top of a data center full of expensive equipment was probably a bad idea so the it building what happens in the it building lots of different it research yes um it's home to my group um with its members of academic staff with various interests and it's the ground floor is home to the data storage group who who work on magnetic and optical storage systems and have the nanotechnology lab where they do graphene stuff all right so um they're all plugged in and people can remotely access and run their programs on it on on the half a million cores yes yes sir you have to apply to get the permission well run their apps yes the access is is under the auspices of the orbea million human brain projects so here we have half a million arm processors running on this side and behind here the red buttons all right so so what you see in here um to these cabinets 100 000 arm processors right here are wired together into a single toroidal surface so you can see the high speed serial cabling that connects them all together at present we have five of these cabinets um making the half million cores in total and our original plan was to go to a million cores and you can see the cabinets right here and we're beginning to populate these um in in the coming months so each s of c has 18 arm cores yes and plus what networking kind of stuff or plus plus a packet switch router so when the neuron spikes that spike becomes a tiny packet which we then route from chip to chip and around and so it can go from the processor that models it to many other processors around the machine it's a multicast system so it may go to 10 000 other processors elsewhere in the machine and the requirement to maintain real time is that we can deliver that packet in a small fraction of a millisecond anywhere in the machine in fact everywhere in the machine potentially well it's not a fraction of a millisecond anywhere on the machine yeah so there's a reliable networking but everything is doesn't have to be synchronized asynchronous it's it kind of works in a in a mainly asynchronous way yes so every processor has its own idea of real time and it generates spikes when when the model says it should and it has to cope with spikes when they arrive there's no sort of end-to-end handshake on the spikes they just arrive and land when they arrive and uh uh your your students up there and uh your team is working on the all the the platform the system for having this to be uh fault proof so if there's like one one cpu that doesn't wake up you go to the next one automatically and yeah so when we configure this um all the boards are tested and effectively each board has a little blacklist of components on that board that shouldn't be used and and they're either faulty or they're a bit unreliable um it's quite a small percentage but basically the software tools will will map to the reliable available subset of the machine now uh when you submit a job through the human brain project collaboratory yeah um it comes to the server which is in the sixth rack at the end there and the server then works out how much machine it needs to run uh allocates that part of the machine and uh isolates it so the serial leaks out of that part of the machine are turned off and so you basically get an isolated subset of the machine that's the right size for your problem so we can run quite a lot of small jobs at the same time um so how do you uh how do you know that the brain works in that kind of way like there's some research being done about how the brain is and you're trying to assimilate it we know the brain is made of neurons okay that's that's a well-established biological fact we have a fair idea how neurons work and what their characteristics are of course our models are approximations and one of the big debates is how much can you simplify a neuron before you've simplified something out that matters and that's not known that's an unknown um issue in computational modeling but we use pretty standard neuron models that are widely used in computational neuroscience and so we work with what the community expects and what they accept as as usable models and because the model is just a piece of software running on an arm core if somebody wants a slightly different model uh we can almost certainly support it and so we have flexibility in the neuron model we have flexibility in the synapse model and the learning rules all these things are just bits of software the thing that's implemented in hardware is the way we deliver spikes and the uh chillers are coming on again which makes the room a lot noisier nice let's go just over here uh so um potential customers for this um uh could be like psychologists yeah um neuroscience stuff i got sorry the neuroscience we work we work with psychologists and neuroscientists yes but particularly computational neuroscientists so these are people who are used to talking to the biological neuroscientists understanding what of interest to them and turning that into computer models and then we offer a platform um which will give unique hardware support for those computational models and and a software stack that makes the machine accessible so maybe this computer is gonna help us understand how we think and how we have like issues sometimes in the way we think like it could be useful for psychotherapists and all these depressive and all these issues people have it's potentially that i mean the computer of course will not answer those questions but it will enable the people who can build the models to if they can design the right model to test hypotheses and so it can help the users answer those questions the computer's not magic it's just a computer um but its ability to support very detailed biological models may allow its users to answer questions that are very hard to answer other ways so you have spinica one and spinica two how soon is uh is it is a work in progress spinica two is it's an early stage of conceptual design um we're taking our experience from spinica one the things that work well the things that work less well um and and we're transferring that knowledge to spinica two we're trying to keep the things that work well improve the things that work less well see what features people are asking for that would extend its capabilities and add some of those where they're economic um and generally you know it's a learning process you build a machine you see how well it works you learn from what users tell you when you build a better machine so uh there is a specific size of the chip right now the nanometers now they're going to be reduced and then spinica two it's got to be new arm cores spinica one uses fairly old technology the the processing layer in spinica one is 130 nanometers see what's technology um spinica two will be 28 nanometer even that's fairly standard now so that gives us 16 times the aerial density but if you look at leading edge chips they're built on 14 nanometer processes but that's too expensive for our application 14 nanometer only makes sense at the extremes of the business so we're we're still sticking to things that are economically attractive for the kind of work we're trying to do so do you have some phd's right now uh trying to suggest how they think you should use the 28 nanometer to die and what you should put on it and how many cores and how many what and doing what yeah that's all that's what you're working on we have we have weekly meetings to discuss how we want to architect the next machine we're also working with collaborators at the technical university of dresden um who are uh very competent at the back end design for the chip so spinica one we did all the chip design here for spinica two we've got a lot of very good support from dresden on designing the physical chip they have global foundries right yeah in the city are they partners or who's helping you to make these changes they dresden works very closely with global foundries and we are planning to implement spinica two on global foundries 28 nanometer technology they have all kinds of tools to help you do whatever you want on a chip right the design support for that stuff you can't design a modern chip without a formidable range of software tools to help you do the design it's uh kind of like how vlsi helped you in the beginning with arms precisely i mean the role that vlsi technology had went on was set up in 1990 um they had the state-of-the-art tools then that we still need state-of-the-art tools and that state-of-the-art has moved on a very long way with the process technology they should also have a big part of the credit for our arm right how how you got actually real how they got made yeah yeah i mean they fabricated it the chillers keep going on the log all right so uh so if you have one million in this room is that the plan to have one million in this room yes of uh and then how that's not a human brain no no a million on cause gets us to about one percent scale of the human brain um i prefer to think of it as 10 whole mouse brains uh the the mouse brain conveniently is almost exactly a thousandth the scale of the human brain um and and in many ways it's physiologically very similar to a human brain and so it's quite a good simpler model to start so then you will have all these million on cords you can load the software very near to each one of them yeah somewhere on the board is this your software you want to test the software's actually on the chip on the chip about a millimeter away from the processor core so how long does it take to load your software that you want to test on a million chips well we haven't done anything that big yet and software load is one of the weaknesses of the current design we're still working on the software load but for quite a lot of problems it takes you long there's a load the problem just to run it so you want to load pretty quickly that would be nice right and then you can have a lot of tests happening otherwise you have to wait for each test to load yeah but of course we can run many tests concurrently if it nearly everything that runs on the machine now is very small uses one or two or three boards and so we can run many copies and we can load each board independently um so we can do quite a lot of parallel loading so this is about parallel computing it's it's a lot about this well it's you know it's a million processors in parallel okay so yeah that's about as parallel as you get so it's the front the front uh the advanced most advanced parallel computing with arm kind of research you're doing i think this is probably the most parallel arm computer that's been built yes and the the latest chinese supercomputer the sunway machine has 10 million processors yeah you want to know that yeah do you want to turn on the light or no no no it's okay so that kind of thing is automatic you're just talking about the chinese is that an arm no no no it's the chinese sunway machine it's all chinese technology super computer and it's the fastest super computer in the world it's taken over the number one slot from the previous number one machine which was also in china yeah and that was tianhe two but tianhe two used intel technology um and was about 33 petaflops the sunway machine uses chinese technology and is about a 100 petaflops because i was trying to find out what architecture that we're using when i saw those news stories come out recently and it doesn't really show maybe they're trying to hide that it's arm could it be arm some of these new chinese super computers that no it's not arm the sunway sunway um there's been quite a lot of information about the design of sunway it uses basically it uses an approach which in some ways is a bit similar to spenica in that it uses a very large number of relatively thin cores rather than a smaller number of very fat cores and and this is advantageous because thin cores are more energy efficient than the fat cores maybe you can go right here so we can't go out okay yeah yeah so um uh so supercomputing research is is related to what what you're doing in some ways yes there are there are connections and um you know we very very carefully we very carefully use the most energy efficient cores rather than the fastest cores on spenica because we're using a lot of them uh if we used faster cores we wouldn't eat so many but they'd use more power because they'd be less efficient so arm has uh uh uh uh has a role in data centers and supercomputers and and research like you do yeah there's a lot of arm should be everywhere well arm is there are there are server variants of arm processors becoming available now and and server companies using arms um they have stiff competition from intel in the server and data center market um supercomputers are an interesting area because the next big challenge in supercomputing is exoscale and if you want to do exoscale economically which means you've got to deliver 10 to the 18 instructions per second for 20 megawatts that's 20 million dollars a year electricity bill that's what everybody thinks of as the limit you clearly got to do 50 gigaflops a watt i mean the numbers are easy and 50 gigaflops a watt is about 10 times more efficient than the best technology you'll find today so exoscale is a huge challenge in energy efficiency terms and it may be that very large numbers of arms are the right way to do that arm has always been the best at performance per watt that's the most important thing right but as you push that performance up to compete at the server end then that advantage reduces because to go very fast you have to add lots of accelerator stuff and accelerator stuff uses power and cost you efficiency so it's more efficient you know them the most efficient arm core is not the biggest it's their smallest all right so this has been a long video maybe i need to ask you one last thing you know there's a soft bank happening in japan hopefully all this awesome stuff can continue in the uk and for example what you're doing is very important with brexit maybe it's influencing with the eu and everything so what's your hope for the future of uk innovation well my hope for the future is that firstly brexit never happens although i think that's a full-on hope now because politically it has to happen um but i i only see that as as as damaging for uk science i see our pan european network our engagement in european programs has been has been crucial now of course we could continue that um but um you know as as as as the swiss and norwegians know the quid pro quo for engaging in european programs is you have to accept the mobility of labor and that's one of the fundamental political issues behind the brexit vote i think um now was the biggest consequence of brexit was arm and the japanese deal and i don't know if the japanese deal for buying arm is related to brexit in any way i just don't know um my view is that arm is is is much more valuable to the world as an independent company its independence underpins its ability to to sell you know process the designs to all and sundry now softbank has said that will continue um but whether that can always continue when you know they become part of softbank and inherit part of softbank's enormous debt um i don't know um i have no issue with softbank or with with japanese ownership but i think arm was was much more valuable to the world as an independent company hopefully maybe it helps with the what's called uh there's all these investor quarterly reports that need to be done when you have investors and now if you don't have them anymore then maybe arm can do whatever they want a little bit more maybe invest more without asking anyone only need to ask the japanese owner it's it's it's possible but the japanese owner is in huge debt uh according to the reports i've read and so how they can possibly afford to let any of the companies they own do other than generate the maximum possible profit to try and fund that debt i don't know but this is not an area i'm an expert in okay i'm a technologist not a not a not a an investment banker so all right hopefully Theresa may is watching this video and uh can influence her opinions in the future maybe okay thanks a lot okay