Special FOSS Ep: Raspberry Pi With Eben Upton

Yes, we had. I mean, obviously this isn't something that we'd ever imagined was going to sell a lot of units. It's a thing we were building to teach, to encourage kids to come to Cambridge. And so our idea of success was, you know, our idea of success was a thousand units, and our idea of wild success was 10,000 units. I think the point where I realized that we were in trouble was in the autumn of 2011. So we had been working on it since 2008, and we'd sort of told people in May of 2011 that we were doing it. And we did this by going to see a guy called Rory Cellan-Jones, who's a BBC technology correspondent, and he took a little video of one of us holding up a very early Raspberry Pi prototype and talking about what we were trying to accomplish. And he got this vast number of YouTube views for his video, he got 600,000 YouTube views in a couple of days, which was kind of cool. But it's very easy to apply a very large discount factor to that number to measure interest, so we thought we were fine. And then we spent a lot of time in 2011 getting from the sort of point it was the kick that we needed, it was the kick in the pants that we needed to actually go from noodling away at this thing to actually try to turn it into a real project. So we were spending 2011 going through that process of productionizing Raspberry Pi. But we didn't imagine that we had 600,000 people who were interested enough to buy a Raspberry Pi. But the point where we realized that we were in trouble and that we would have to scale the organization to support that was when we put an operating system image online in about November of 2011. And it got something like, and this was a Fedora-based operating system, and it had something like 50,000 downloads. And you could only run it in QEMU, right? There was no hardware out there. So you could run it in a modified QEMU. We had 50,000 downloads of a very badly broken operating system for a computer that didn't exist yet. And that was the bit where we realized that somehow this had tapped into, it had found in that kind of big parameter space of the things people are interested in, we had accidentally dropped a product into a completely empty space where a lot of people were standing. It was empty of products but very full of people.

Yeah, I think it was certainly when we started, I guess probably 2009, 2010, when we were first working on what would become the Raspberry Pi platform. ARM Linux itself, I mean the kernel, was an unusual thing, right? I can't remember because it existed for a long time outside the tree. I guess it may have been merged by then, but ARM Linux was outside the tree for a long time. It probably had been merged but probably fairly recently. There were various distros available, but it was all a bit everything was a bit special. And how you built things, I remember trying to find a compiler, trying to find a pre-built GCC to bootstrap yourself, and to build your kernel wasn't trivial. And yeah, there's just kind of a moment, if you draw that line between that point kind of 2010 and probably, you could say maybe the launch of the M1 Max was probably the point where ARM as a client architecture kind of emerges completely into the mainstream. There is that point somewhere in the middle where large numbers of applications were starting to provide binaries, generally Raspberry Pi binaries, generally binaries for this slightly unusual Raspbian world. So that's ARMv7hf, so hard float, ARM Debian but built to remove the V7-specific instructions so it could run on the ARM1176 that we had in Raspberry Pi 1. Those sorts of things started to appear in kind of 2013, 2014, 2015.

I mean, of course, the BeagleBoard predates us, probably the first broadly available. I guess it didn't have the continuous, one of the interesting things about Raspberry Pi, and of course we've been challenged with this over the last couple of years, but for most of its life, one of the interesting things that's characterized Raspberry Pi is kind of continuous, high-volume availability. And that was probably a thing that the BeagleBoard didn't have. It was something which was sporadically available. They would build some and then sell them out and then build some more. But it did, and it was probably three or four times the price, or five times the price of Raspberry Pi. But it did fit that same niche. But yeah, certainly I think we popularized it as a platform. Going on to that point about the M1 machines, of course we were super happy. We have this object called Raspberry Pi 400, right, which is a Raspberry Pi 4 built into a compact keyboard. We launched that about a week before the M1 Max. Now, I grew up with, not owning but lusting after, the Acorn Archimedes, the kind of the original ARM computer, the original ARM desktop. And it was kind of fun that we actually were first back in market by a week. We were, after a 20-year period in the wilderness where there were no kind of consumer-oriented PCs, we beat them by a week. We were first back in market.

Oh yeah, it's interesting, right? It's all about how you tell stories, which bit you want to emphasize. I think for me it was a coming together of those. It was a perfect coming together of those two strands. There wasn't really a chicken and egg. I was building little computers, actually quite Arduino-like little computers. So there's this thing that if you search for Raspberry Pi 2006 Edition, there's this rather cobbled together looking piece of veroboard that has an ATmega, same chip architecture as the Arduino Uno, it's an ATmega644, it's got 512K of RAM, and it will drive a standard definition video signal. And it kind of, it bashes, actually uses the microcontroller both to run your code but also during the display period to bash out video addresses to get this SRAM to drop bytes onto the bus in the order required to draw something on the screen. So I was building those things. And I was also a director of studies in computer science at the University of Cambridge, and so I was also experiencing the collapse in the number of young people who are interested in computer science. We saw this terrible decline between the late 1990s and the mid part of the first decade of this century. And so I was kind of exposed to both things. It was a sort of natural coming together. And really, Raspberry Pi is testing the hypothesis that we used to have loads of kids who were interested in computers because there were loads of programmable computers in kids' lives. And that when you took those computers away, because you largely replaced them with games consoles in a lot of children's lives, the ultimate piece of very powerful non-programmable hardware, non-end-user-programmable hardware, the disappearance of general purpose computing made the kids go away. And if you bring back general purpose computing, the kids will come back. That's the hypothesis we're testing. And it was a kind of hypothesis that was very easy to stumble upon if you happen to be both building little computers and confronting the disappearance of computing among young people. So there's not really a chicken and egg. It's a perfect coming together of these two strands of my life.

I think if you imagine there's this scalar value that we care about, which is number of applicants to the University of Cambridge. The deadline is the 15th of October something like that, so on the 16th of October every year you're presented with a scalar number. That value went from about 600 in 1999 to about 200 to 250 in 2008. And last year that scalar was about 1,450. So we have a little over twice as many applicants to computer science at the University of Cambridge as we had at the height of the dot-com boom when everyone thought that computer science was a meal ticket. So if you just look at that number, yes, there has been an absolute sea change over the last 10 or 12 years. And I think Raspberry Pi has had a part in that. I think the kids who rock up at least have the good sense to pretend that Raspberry Pi is the reason for this. Fascinating for me, as somebody who grew up writing computer games on my BBC Micro and on my Commodore 64, and had imagined that the primary use case for Raspberry Pi in education would be software. The vast majority of these people say Raspberry Pi and robotics. So the vast majority of these people are not having a purely software-mediated interaction with Raspberry Pi. It's about physical computing. Which means it's very, very lucky. I mean, that's obviously somewhere Raspberry Pi kind of stands out. It's not something that I was super excited about. Those general-purpose IO pins on the Raspberry Pi which allow you to do all those kind of cool physical computing things, that was Pete Lomas who designed the first generation Raspberry Pi hardware. He was very keen to expose some of the interfacing capabilities of the chip we were using, and I didn't care enough to argue with him. And of course, how lucky we were, both of course in the education space, but also that was the capability that unlocked the industrial use of Raspberry Pi, which really now accounts for the majority of our sales.

The nice thing with the core platform, so the SBC, the single-board computer platform that's kind of classic Raspberry Pi, it's still the dominant volume runner for us. Out of, you know, seven million obviously last year was a rather volume-constrained year, but if you look back to 2021 when we sold 7 million Raspberry Pi computers of all sorts, the SBC product, the single-board computer product with its four USB ports and an Ethernet connector, that was between 5.5 and 6 million units. And compute module was on the order of a million units. So the SBC is the volume runner. We're lucky that we've never had to really make any decisions to prioritize say educational users of that product or industrial use of that product. The decisions you make to make it a great educational product also make it a great industrial product, because particularly the kind of robustness and reliability are very important. You know, it's always said, what's the worst environment? What's the more challenging environment? An oil rig or a kid's bedroom? They're both pretty tough. I mean, I've got two kids and I've never been on an oil rig, but I've been to some power stations. And the kid's bedroom is a pretty tough environment. So we've kind of always been able to have a happy medium. You never have to choose one consumer over the other. We have now compute module which is purely targeted to the embedded and industrial space, and of course we have Pi 400 which is purely targeted at the consumer space. So we have put our toes in the water in both directions. The nice thing is actually both of those are kind of most of the investments we make are at the platform level, so deriving those specific platforms from the core platform, from the SBC platform, is actually quite cheap from a cost and engineering perspective. So we have tried to put this product in the market in the middle of the market, and then we just kind of tried to feel out what people don't like. And the main thing that people don't like as they go to scale with the SBC is they don't like being enslaved by our form factor choices. And you do see people, you do see industrial users, I have a keen eye for the appearance of that boat tail, the two pairs of USB connectors and the Ethernet jack. If you look carefully at lots of industrial products, you see that boat tail a lot. It's the sign that someone's built a big industrial object and there's been enough space inside it, and so their solution for how do they embed a Raspberry Pi is they put it inside, they have a wiring loom or a HAT that sits on top of it and mounts their circuitry. The challenge that people have is when they try to address more consumer-oriented use cases, they don't want to have to follow our choices about onboard peripherals, they don't want to have to follow our choices about connector layout. And that's really where the compute module products have been kind of a slow burn. We've actually been doing compute module since 2014, and it's really only the last couple of years, and I think with the introduction of Compute Module 4, that started to take off like a rocket. So it's still a small part of our volume, but by far the largest growing segment of our business.

Yeah, and we've had two distinct form factors for the compute module. So up to Compute Module 3 and Compute Module 3+, we had an SODIMM form factor. You put a JEDEC SODIMM socket down on your board and then kind of slide it in and clip it in like a memory module. And then Compute Module 4, we threw that away. We had backwards and forwards compatibility in that space, and then Compute Module 4 we replaced that with a pair of high-density, managed impedance connectors that you clip down onto the board. That's definitely been a good choice. It's a more space-efficient choice, it's a more cost-effective choice. I guess the other decisions I think we made with Compute Module 4 which were informed by experience with earlier modules and have really helped us: we put Wi-Fi on the board as an option. And there's a platform, I'm not sure if it's defunct now, called Android Things, which is effectively an IoT platform built around the Android kernel and bits of the Android userland. And I had a very revealing conversation with a member of the Google Android Things team in, I think, back in about 2018. And he said, 'Look, you really need to put Wi-Fi on your module because what it means is as soon as you want to go to scale, you either have to do a bunch of Wi-Fi engineering yourself or you kind of have to abandon the Raspberry Pi platform and the compute module platform for another platform.' So we put Wi-Fi on as an option. That obviously eases a lot of people's integration, eases a lot of people's migration from the SBC platform which has integrated Wi-Fi into the modular space. So that was one thing. The other thing we did was we massively simplified the power chain. So earlier compute modules, you have to give them a bunch of rails. You have to give them a 3.3V rail and at least a 3.3V and a 1.8V rail, and I think another rail as well. You have to put circuitry on your baseboard that delivers those rails and brings those rails up and takes them down in order in the proper order so that nothing latches up. Compute Module 4, you give it 5 volts and it makes its own rails. It has a power management IC on there, it makes its own rails. In fact, it gives you the rails back. So it'll give you a 3.3V rail for you to use as long as you don't draw too much power from it. That simplified things hugely. I struggle sometimes when we're spec'ing and designing these to understand what the pain points are going to be. So we put these products in the market and sometimes we find that we've done something that people find painful and it limits the success of a product. That was definitely a big factor in the success of CM4.

So if you think about the evolution of the Raspberry Pi software environment from 2011, 2012 through to the present day, I think what you're seeing is we have this thing that I occasionally call start.elf, and that is sometimes disparagingly called the blob, which is a large closed-source software binary that runs on the closed CPUs in the platform. So it runs on the VPU, effectively a DSP, it's actually fairly powerful DSP, but really these days we only use the scalar core there as a kind of system management processor. Early days, everything ran on there. Everything multimedia related ran on there: camera processing, video decode, video encode, display scan out, control of the HDMI subsystem, 3D graphics. All of that stuff ran inside the blob. And then what you've seen over the years is a kind of carving away. So where you had an open, famously where you had an OpenGL API on the ARM side, what that was was effectively a serialization shim that would take the calls, serialize them, pass them over a FIFO down into this blob which would then do what it was told. Actually, that's kind of great for OpenGL because OpenGL is designed for that exact implementation style. But it wasn't very pleasing to people because people couldn't see what was going on inside. And we were very lucky that in about 2014 we got permission to release the architecture documentation for the 3D graphics core, the VideoCore graphics core in the chip. And we released this, and this is pretty unusual actually to have good public documentation. I was involved in designing the core, and it's pretty unusual to have good public documentation that still isn't really the case, say, for the Mali cores that you see in a lot of other SoCs. And Eric Anholt joined my team at Broadcom basically to build a Mesa driver stack around this using this documentation. So from probably 2014, 2015 onwards, we did have the option to turn off the 3D graphics element of the blob and use Mesa instead. And then Eric went on to develop a driver for VideoCore 6, which is the core that we use in the modern Raspberry Pis, and that's then being further improved by Igalia. So on the 3D graphics side, you saw us turning off a bit of the blob. And then in the camera processing side, you saw us turn off the camera processing and move across to libcamera. And in display scan out, you saw us turning off that element and moving across to KMS. So really all that's left actually is video encode. H.264 video encode still runs inside the blob. And that's really, I mean, it's great because obviously we can leverage the efforts of these enormous teams that have put together these very powerful middleware systems. We can leverage some of those efforts, we can contribute back to those efforts. If you look at our education mission, it means that there's no point where you have to say, 'Here be dragons, don't look inside this black box that does OpenGL.' You really want someone who's sufficiently inquisitive to be able to go all the way down and see when I ask 'please draw this triangle,' what actually happened deep down at the bottom of the stack and what do we think is happening inside the hardware. And I suspect that video encode is still there in the blob. I suspect probably in the next or next one generation of Raspberry Pi, we will get down to a point where there's effectively only clock, power, and reset. And that's fairly traditional. I mean, Intel chips have the Intel Management Engine in there. Most systems have some little microcontroller which is handling bringing the big cores out of reset effectively. And we'll probably get to that point. That'll be really desirable. And hopefully we may even, when we get to that point, even though there will still be a binary and probably a signed binary down there that does those things, hopefully we'll even be able to publish the source for that binary so people can even understand what that's doing.

We have an interesting philosophy about how we decide what to put into a Raspberry Pi hardware platform. And it's broadly you can summarize it as: we take the cost of the feature and we divide it by the fraction of people who will find it useful. So a feature, a $1 feature that everyone uses, or a 10-cent feature that's used by a tenth of people, get the same priority in our decision-making. This is actually a really hard barrier for new features to overcome. The only really, if you look at a modern Raspberry Pi, the only qualitatively new feature that's in a Raspberry Pi today that wasn't in the 2012 product is wireless, Wi-Fi and Bluetooth. And that's the only thing that is a $1 to $2 feature and it's useful to almost everybody. Pretty much everything else, there's no ADC, there's no analog-to-digital converter on the Raspberry Pi because it would cost 10 to 20 cents and we don't think that the vast majority of our users need one. So what that philosophy tends to do is it tends to push lesser-used features off onto external hardware. And that's always been good to us as a decision-making heuristic. So if you want to do some machine learning on a Raspberry Pi, actually for 1.8 GHz A72 is quite a lot of machine learning throughput. So a lot of people's machine learning requirements will actually fit onto the CPUs. If you want to do more machine learning than that, you can buy a Coral, say Google Coral accelerator. They're about $60, they've got about 2 TOPS of inference on there. And that applies to everything from analog-to-digital converters all the way up to machine learning accelerators. It's been good to us. It's actually created space for other people in our ecosystem to innovate. The fact that we don't try to slurp all of the functionality into the platform means there is a business, you could start a business selling really nice digital-to-analog converters that you connect to a Raspberry Pi, and we won't come and drive over you. We won't come and add those features into our platform.

For us, because we always kind of knew that you can do better than PWM audio. Even if the hardware you have is effectively a filtered pin, you can do a lot better than just piping PCM audio into the duty cycle register for a PWM block. You could do all sorts of clever sigma-delta noise shaping stuff, one-bit DAC effectively. What if it was a hi-fi? Would be marketed, if it was a hi-fi in my childhood at least, would have been marketed as a, you know, Panasonic would market these things as one-bit DACs. And what we've ended up with, if you go back and you run a modern Raspberry Pi operating system on the Raspberry Pi 1, and we're quite proud of that, that's kind of a core bit of our ethos that you can run today's operating system on the 2012 hardware and it performs pretty well. Actually, you'd find the audio quality is pretty good, and that's because it's picked up a lot of noise shaping, a lot of fairly alarming noise shaping code over the years.

It was a real wrench for us having to, and we limited ourselves to two. So I mean, there's another obvious choice which would be three. So you could have an ARMv6hf Raspbian effectively image that ran on Raspberry Pi 1 and Raspberry Pi Zero. You could have an ARMv7hf, armhf in Debian speak, that runs on Raspberry Pi 2. And then you could have the ARM64 version that runs on Raspberry Pi 3 and later. So we decided not to do that. So there's no special Pi 2 Cortex-A7 build. Yeah, it was a bit of a wrench, I think, because of the way...

That the Debian World Works actually not too much of a wrench. You know, an enormous amount of effort goes into, you know, if you sort of pick what couldn't we have done Raspberry Pi without obviously Debian and the way that Debian thinks about the world, and so the fact that a lot of effort goes into making the system fairly architecture neutral allows us to support fairly radically different architectures, of course ARM32 and ARM64, with relatively little engineering effort.

Oh, we have sporadic contact. So I think you can sort of divide operating system support for Raspberry Pi into four buckets. So you've got Linux, various Linux distros. You have Windows, and of course there are sporadic efforts from Microsoft and from the community to make various incarnations of Windows work on Raspberry Pi, some pretty successful, some fairly rough. You have the BSDs, and we do have some light touch interaction with the various BSD ecosystems, generally trying to answer questions and also trying to provide, where we can, software that interfaces with the various subsystems on the device which isn't GPL licensed, because of course they have a challenge that there are various subsystems where they can't really take bits of Linux kernel where you're close to the Linux kernel drivers being the reference software reference architecture, but of course they're inappropriately licensed, and so we occasionally do work to try to help people with that problem. And then you have other operating systems, that's everything from various research operating systems to RISC OS, of course. We've always liked the fact that you can run the original ARM operating system on Raspberry Pi, and we do have a very friendly relationship with RISC OS who are just up the road from us.

I think if I had my time again, yeah, I would, what wouldn't I give today for a million, you know, for an extra million well of any of the Raspberry Pi ships actually, BCM2711 which is what we used for Pi 4s, but what would I give for another million of any of those chips. And of course it should have been obvious, right? The core dynamic was a pandemic happened, everyone assumed there would be a massive recession so stopped ordering. Businesses assumed there would be a massive recession so they went into destocking mode. At the same time, everyone was at home needing electronic equipment to continue to butt society, and also in most certainly European and North America having their income protected by various government schemes and also unable to spend money on services. So you have the same income but you can no longer go on holiday or go to a restaurant, so what are you going to do? You're going to buy consumer durables. And so it was that pair of transients, the negative going supply transient and the positive going demand transient, were, if you sat someone down in a business school class and you described the situation to them in those terms, of course they would tell you what was going to happen. But at the time when we were in the thick of it, and of course a lot of us were focusing on not getting COVID, nobody called this one right. The big car companies didn't call this one right.

So yeah, if I had buy time again, obviously I would, what would I like? I'd maybe have liked somebody to talk me through it in business school terms because then I could have spotted what I needed to do. So towards the end of 2020, you see things start to come apart. So you see first blowouts in component lead times. Components that you would expect to be able to get on 20 week lead times going out to 52 week lead times. We were pretty quick in getting orders on in that environment, but obviously if lead times instantaneously blow out by say three months in one day, it's inevitably going to create a three-month void in your supply chain. There's no way you can get orders on fast enough to fill that 3 months. It's just an instantaneous hole. And it's not like there are parts, it's not like you have warehouses full of parts in China sitting waiting to go. Stuff comes in and it gets stuck on boards and the boards go out. And you have some inventory, I mean you have some component inventory on any given day, you have some component inventory, and you have some finished goods inventory, and your downstream channel has some inventory, and industrial customers of that downstream channel have some inbound inventory which is your finished product. So there's lots of inventory slushing around in the system, but no, it's certainly not the case that you have millions of units of every component that you need to build a Raspberry Pi. So you have this kind of blowout.

Then once you get back into a situation where your orders that you've placed that were now 52 week lead time arrive, you end up in an allocation situation. So you end up in a situation where even though you have orders on, you don't get supplied at one to one ratio because the total aggregate demand exceeds the total aggregate production capacity, and so the only thing you can really do is to give everyone a haircut. So that was the kind of inbound situation for us, which has persisted really up to the present moment. It's a situation which is now improving quite rapidly, and I think what we said when we gave an update in December was broadly the first calendar quarter of this year is pretty tough. It has some bright spots, and I think if you go on our locator you'll see what some of those bright spots are right now, things like Raspberry Pi 3A plus. Later in this quarter we'll see Raspberry Pi Zero, we'll see a significant improvement in Raspberry Pi Zero availability, which is of course kind of cool because it's still my favorite Raspberry Pi product, Zero W. So we'll see some improvement over the course of this quarter, but it's still a constrained quarter. Next quarter is broadly speaking a pre-pandemic quarter.

So you can see next quarter as being one and a half to two million units of inbound inventory across the various components that we use, but because you have backlogs you need to do better than that. So that's enough to keep up but it's not enough to start to get back. So I think what you'll see in the second quarter, much better availability, but a continuation of things like single unit limits for consumers, and what we call active management for our industrial customers, where a customer orders something, and rather than just taking their order, we go back to them and we actively say to them, do you really need that many units? It's anti-sales. We have sales people here doing anti-sales. Oh, so can I have 10,000 Raspberry Pis? Well, would a thousand Raspberry Pis each month for 10 months be enough? You know, because you're trying to detect inventory building, what we call kind of toilet roll hoarding in the context of the pandemic.

Because that will destroy that, you know, even if you have a sufficient supply, that kind of beggar-thy-neighbor behavior which is totally understandable and totally defensible, and I would absolutely be doing it if I was buying these units, you have to find ways to help people. It's not about stopping people doing it, it's about making your industrial customers comfortable that if they don't do it, they'll be fine, that we will look after them, that they can disclose their true demand to us, and that we will use that in a good way to help make sure that they don't go lines down, rather than in a bad way to give them a haircut. So you'll see that those behaviors will persist through Q2, but it will be a good environment but it will be a heavily managed environment. And then from Q3 onwards this year, we effectively have an unlimited supply of pretty much everything we need to make Raspberry Pis. That will allow us to catch up on our backlogs and then get back into the situation we want to be in, which is where we don't actually need to touch our customers so much, we can just put units in channel. The dream with Raspberry Pi, and where we were for nine out of 11 years in the market, we've been in a situation where there are hundreds of thousands to millions of units of Raspberry Pi in channel, and if you wake up one morning as a hobbyist and you want 10, you can have them. If you wake up as an industrial customer and you want 100,000, you can have them. And that kind of law of large numbers averaging effect means that actually not many people wake up on a given day and want 100,000 units, and therefore you can pull from the channel and it's as if there's an infinite supply of Raspberry Pi.

I just it's, you can think of Raspberry Pi as having explored another use of Moore's Law. So the classic PC industry use of Moore's Law is to pick a price point and every 18 months to two years, fill that price point up with twice as much compute, and so you get an exponentially increasing amount of compute at a given price. The other thing you can do with Moore's Law is you can pick an amount of compute and you can ride an exponentially decreasing price curve. And you can sort of see Raspberry Pi as a piece of hardware having said in 2010, 2011, 2012, hey, what would a 2000-era PC cost now? Actually, because of Moore's Law, it's $35, and that was kind of great. And then we launched Pi 2 in 2015, and what did we do? We picked $35 and we filled it with an exponentially increasing amount of compute. We'd actually just gone back to doing what at a different price point, we'd gone back to doing what everybody else does with Moore's Law. And really, Raspberry Pi Zero I love because it's an example of us, a few months later, realizing that we'd conformed, and then asking ourselves, well, what if we didn't conform? What if we said, let's take the 2000-era PC, what does it cost now? Oh, it costs $5 or $10. And so I like it for that. And the projects you get with it are very embedded. It shows up in some of my favorite hobbyist projects like the high altitude ballooning because it's very light. It shows up in some of those projects. So I do love it, and we've sold a lot of them. I don't know what the total numbers are, but it's probably three or four million out of the 50 million roughly Raspberry Pis we sold. Probably three or four million of them have been these tiny little computers. It lives in a fun spot between a microcontroller and a full-blown PC. And of course we've explored with Pico which came out a couple of years ago, which has our own silicon on it, it's the first product that has Raspberry Pi designed silicon on it. We kind of then have explored what does the world below Zero look like. We have this $4 Pico product and $6 Pico W product, just sort of trying to understand how low can we take it. And then the SoC, which you can also buy, you can buy the microcontroller from us, and it's a $0.50 microcontroller. So, you know, how low can we take that general purpose computing mission?

It suddenly helped that we had access to two process nodes. We had core silicon across two process nodes, 40 nm for the Pi 3 earlier, 28 nm. So being exposed to more process nodes is good. If I had my way I would have been exposed to three or four process nodes, but two was certainly better than one. We saw people migrating in both directions. We saw people driven by month-to-month availability behavior. So we saw people who were using Pi 4 era hardware looking to see whether applications could be optimized to run on earlier hardware. We saw people who were using Compute Module 3, we have a sort of what's currently a pseudo product called CM4SCI as a way to give people a migration path up to the modern silicon platform without having to retool their baseboards. But it actually ended up doing duty as a way for, in the periods when we had better 28 nm availability than 40 nm availability, it did duty as a way for people to get continuity of supply. It's a product that is on our website, it's a real and official Raspberry Pi product, but it has never really been launched. It's only been used as a tactical aid to some of our industrial customers. And even today, you're going to see us come back into availability in broadly the order in which our products were launched. So you're going to see the old 3A Plus come in, see Zero W come in, and then probably Pi 4 will be towards the end of the consumer unrestricted volume availability. Raspberry Pi 4 will probably be towards the end of the transition.

It's very hard for me not to be all in on ARM, as somebody who spent his childhood drooling over Acorn computers and as a former colleague of Sophie Wilson, who was responsible for the design of that instruction set architecture. It's very hard for me not to be all in on ARM. I think RISC-V is an interesting development. People, it's important not to understate the challenges that a new instruction set architecture faces in terms of adoption. There's an enormous amount of accumulated invested engineering effort in the ARM architecture and the Linux kernel and the various userlands that run on top of it. Your optimizations, go look at the FFmpeg source code, a huge number of NEON optimizations for both 32 and 64-bit. Actually, that accumulated base of engineering investment pales in comparison to the even more enormous level of investment in the Intel architecture. If you go into some of these places, you find fast paths for everything from MMX to SSSE3 to SSE2 onwards, a whole museum of Intel multimedia instructions. And stuff that needs to be replicated in the RISC-V space. It's not to say it won't be replicated there, but it does need to be replicated, and that would be a necessary precursor to us being able to do any sort of RISC-V based Raspberry Pi product. So I'm very satisfied with the ARM architecture. The ARM ecosystem is a quite good place to hang out, and it's notable that quite a lot of other computer companies have chosen to go hang out there recently. And just as the Linux kernel provides a venue for cooperative collaborative innovation, so the ARM instruction set and the various cores that you can license that execute that instruction set provide a similar environment and certainly a similar source of scale advantage. So it's going to be interesting to see what happens with RISC-V. I think in the microcontroller space, those advantages are much less pronounced, and I think where you do see penetration, and I think you'll continue to see penetration of the RISC-V instruction set architecture in microcontrollers, deep embedded.

The funding is interesting. Foundations are top charities. The thing I run is Raspberry Pi Limited, so I run the engineering subsidiary of the foundation. We do the engineering. If you buy a Raspberry Pi computer, it was designed by us, and the software was written by us, the Raspberry Pi specific software, and the brand is maintained by us. So we do that bit. And then we make money, and then we give that money, well we reinvest some of the money in making new Raspberry Pi stuff, and then we give the surplus money to the foundation that does educational things with it. We've returned on the order of £35 million I think over the years to the foundation to fund its educational work. So that's the kind of structure. Where did the money come from in the first place? A handful of us put money in. I think we put in about £100,000. The founding trustees of the foundation lent the foundation about £100,000 between us, actually for a period of very few months, and then those loans were then repaid. So that's the initial capital. And actually until 2021, until the autumn of 2021, that was the only money that had gone into Raspberry Pi. So the whole of Raspberry Pi was bootstrapped off that money and probably about 30 or 40,000 of donations from high net worth individuals in Cambridge who thought it was a good idea. So that's kind of fun. And then we did a $45 million funding round in the autumn of 2021, a couple of absolutely fantastic supporters of ours put in $45 million into the trading business effectively, bought new shares in the trading business. The trading business then ceases to be still a very much majority owned subsidiary of the foundation, but no longer a wholly owned subsidiary. But that's the only external money we've actually brought in.

In terms of how my job has evolved, I guess it's evolved from kind of overseeing every aspect of Raspberry Pi early on, through to very much a focus on the commercial and engineering side of it, which is natural for me because I'm a software engineer originally, I'm a chip engineer subsequently. So having the opportunity to focus in on that aspect of things has been great. We found ourselves a fantastic chief executive for the foundation, Philip Colligan, who we poached from Nesta, the National Endowment for Science, Technology and the Arts in the UK. He was deputy chief executive at Nesta, and he joined us in 2015. I think he was about for the early years of the Raspberry Pi, the engineering function was big and the charitable function was small, so Philip was probably only at that point about the fourth employee of the foundation, and then grew that very quickly through both organic growth and acquisition to about 150 people. So that's the foundation in terms of headcount, now rather bigger than us. And he's stayed with us, he will have been with us eight years this year, and he's transformed the foundation into something that we didn't envisage when we started. The existence of the Raspberry Pi computer was supposed to be the contribution, right? Now we have a foundation which, you know, obviously we continue to ensure that Raspberry Pi computers exist, but the foundation does much more conventional charitable things. They train teachers, they run clubs, they create curriculum and things like that. So it's kind of wonderful to have both that original anarchic strand of just making sure general purpose computing is everywhere, and also to take some of the money you've made from that and go do more traditional interventions in the space. But for the most part, the day-to-day of both the charity and the trading company are funded by sales of hardware, right? Yes, that's right. Which is cool. We've sold 50 million computers, and we don't make very much money when we sell it, we don't make very much profit when we sell a Raspberry Pi, but if you sell 50 million of them, then you do make a bit of money.

For the longest time, we were a licensing company, taking a leaf actually out of ARM's book. We designed the Raspberry Pi and we licensed that design and the use of the Raspberry Pi brand and obviously the community behind it to two partners who then made Raspberry Pis and paid us a royalty. And the nice thing about that business is as you grow, a traditional manufacturing business sucks capital in because you need to fund inventory. Your working capital loop expands, you need to fund all of that inventory, component inventory, finished goods inventory, receivables, waiting for your customers to pay. You need money as your business gets bigger, so you suck capital in. As a licensing company, the unit economics are a little bit less good because you're only getting a royalty rather than the full margin on the product, but it does mean you don't need to suck money in. And that's why the foundation was able to retain 100% ownership of Raspberry Pi Limited while it grew, and thus we were able to devote the entirety of the surplus profit to funding the foundation rather than having to share it between the foundation and people who had come in over time to contribute capital to the organization. So then the reason 2021 was to transition off of the licensing, it was certainly to transition to more of a blended model. We are a much more blended organization now. Certainly last year, a substantial majority of Raspberry Pi computers were actually made by us, or we commissioned the manufacturing. Nobody makes anything, right? So we paid some people to make it for us, but it was a traditional company with us inside. That's obviously capital intensive. Also a certain amount of that money has gone to pay for advanced R&D stuff, because of course we continue to try to make cool new Raspberry Pi things, and that's not cheap.

It's a wonderful, and most of us who are involved in Raspberry Pi have had some sort of business experience before, a more traditional business experience. I think for all of us, it's kind of been a wonderful toy, I guess. This experience that no one's built an organization like this before, and it's interesting to see how far we can take it. We've sold 50 million computers, we fab our own chips, again, nobody fabs our own chips, we pay people to fab our chips, we design our own chips. You design your own chips, TSMC, just like everybody else. And it's kind of interesting to see the breadth of what you can do within this interestingly constrained environment.

I love our own Astro Pi program where we've put Pis on the station for kids to run code on. A lot of European children have done that, it's a European Space Agency collaboration, so that's a fun one. I love that actually. Although there are two Astro Pis on the space station, there are a bunch more, on any given day a huge number more Raspberry Pis on the ISS because they get used as payload controllers because they're very robust. And actually we have no idea how many Raspberry Pis there are on the ISS, and we only tend to find out about these uses when they get downmassed, so that's kind of fun. What else? I just like all the people who've built businesses around it, whether they're resellers or people who make DAC accessories, or people who've built it into some product that they've dreamt up. Just this idea of enabling entrepreneurship, enabling distribution. I think entrepreneurship, capitalism, business, they're wonderfully powerful tools for good change in the world, and it's wonderful to have had an opportunity to kind of unlock that kind of innovation which was getting hard to do. You know, you could walk down to an electronic store and get a 6502 out of a bin for about the same price that the biggest company in the world would have paid for a 6502. It was a very level playing field in the 1970s and 1980s, and the playing field is not level at all anymore in terms of access to advanced technology. It's very hard, it's not that advanced technology is expensive for some of these companies, it's just unavailable. And so taking really advanced technology and making it available at a very flat price, the thing that characterizes Raspberry Pi is not just the thin margins but the very flat pricing structure. So people come to us and say, 'Oh, so Raspberry Pi is $35, how much will a thousand cost?' And the answer is $35,000. You know, a thousand might be $34,000, but it's not $25,000 or $15,000. And that's a commitment we've stayed very true to all the way through. I think it's obviously important from the educational side because it means that the people at the bottom end of the volume curve are kids, so you don't want to be rinsing the kids. But it's also the thing that's powered this kind of community, this ecosystem of innovation around the platform.

Remember most big Raspberry Pi products use merchant silicon, so we don't develop the chips that are in the big Raspberry Pis, that's Broadcom. So we're not involved in that. But just doing the board bit of that work is many millions of dollars now. Raspberry Pi 4 was probably a $2 or $3 million program. Raspberry Pi Pico was probably a $7 or $8 million program because that did have a chip development in it. That has the RP2040 microcontroller wrapped up in it. So it's probably RP2040 is about a $5 million program just on its own, and then you have the board development and the software development on top of that. So it's only getting more expensive. I would imagine, we've kind of taken a pause on new big Raspberry Pis because we focus on just making the existing Raspberry Pis available. But I would imagine that any subsequent Raspberry Pi will be a more than $10 million effort even without a chip development in it.

We have this idea of reluctant hardware companies. Actually, the world was full of reluctant hardware companies when we started. Digital signage companies, thin client companies, people building widgets to go into factories to log data. All these people who were really software companies, they all wanted to be software companies. Their differentiating value was in whatever software stack it was that they provided, but in order to get that software stack, you can't sell people software, it needs something to run on. And the PC was too expensive, and there weren't readily available embedded PCs at a respectable price point. Therefore, people were building a little thin client, doing all the work of going out and sourcing cores, looking at memory, bringing up a board, writing a board support package, getting an RS-232 port, just so they could get to the point where they could then run their application. And that's just so much wasted effort. That's really what I'm proudest of with Raspberry Pi's contribution, I think. We freed a whole heap of companies. So I talked about the boat tail, you know, now you'll see a thin client on the desktop and you look at it, it's got a boat tail on the end because it's actually a Raspberry Pi in a fancy box running somebody's differentiating software application. And that, in terms of contribution to the public good, not having thousands of talented engineers wasted doing basically undifferentiated grunt hardware design, it's a good contribution. I think it ranks actually probably alongside the educational impact in terms of good done by Raspberry Pi.

Terror. Okay, sadness. Sadness in that the progress towards Raspberry Pi 5 has been derailed by all of the challenges pandemic and post-pandemic that have confronted us. I mean, obviously one day there will be a Raspberry Pi 5. We've told people it's not this year, and it isn't this year. I think once we get back into a stock position, into our traditional super robust stock position with Raspberry Pi 4, we'll be in a position to think about it. But it is some distance away. These are not cheap programs, and the fact they're not cheap implies that they're not quick, because it takes time to spend that much money. But we'll get there. And we're lucky that this happened to us when Raspberry Pi 4 was the flagship platform, not Raspberry Pi 5, because of course Raspberry Pi 4 is enormously more performant and has a lot more headroom in terms of how much more performance we could eke out of it by clever software optimization.

Yes, absolutely. It's the first one that delivered on our dream really of no compromises. You can sit down, and when people, lots of kids did, we had a lot of charitable work getting these out to people in the pandemic. It's the first one that you can sit down in front of and use as a no-compromises entry-level PC. You're not going to run a big CAD package on it, but you can run a web browser on it. So yeah, it was good that that came along. It's good that it was available. And obviously Pi 4 and the P400 during the pandemic, it was one of the things that certainly sustained me when I was locked in my house was the pictures of kids getting their Pi 4 and P400 kits. These kids, we had so many, it was such a crime, the sending all the kids in the country home to work from home and forgetting that huge numbers, I think 700,000 households in the UK had kids but no general purpose computer. It didn't create the inequality, it reminded us all the inequality was there. So to the extent that we were able to get units out to a few tens of thousands, we had some great philanthropic support, we were able to get a few tens of thousands of units out to kids mostly in the UK, just the looks on their faces when they opened the box and it was a computer inside was amazing.

So raspberrypi.com for Raspberry Pi Limited stuff, engineering and products. Raspberrypi.org for the Raspberry Pi Foundation and all of their good educational work.