Pretty exciting news here: we launched a podcast! We had the opportunity to record a one-on-one chat with Alexandros Marinos, balena Founder and CEO, and Eben Upton, Founder of the Raspberry Pi Foundation for our very first episode.
The two talked shop about how hardware manufacturers manage the global component shortage, the future of RISC-V, and how manufacturers aren’t paranoid enough (about quality)! Check out the full recording above, and please subscribe to our new YouTube channel.
There’s also a full episode transcript toward the end of this article.
Prefer to listen to the audio version?
We talk to people with remarkable stories about delivering the fire of physical computing to the masses. Now that the power is in everyone’s hands, we ask our guests: what’s next?
Full transcript of Episode 1
As the viewers will probably know this is the first of hopefully more conversations like this, and we’d like to talk to people who have had remarkable stories in the industry, and trying to get the fire of physical computing to the masses. And who better to start with than Eben Upton.
So, Eben, without, without further ado, um, we’ve got, there’s so much we could talk about, but the, the, the one that, uh, always I always wonder about is, you know, the, the Raspberry Pi Foundation basically started with a very strong educational mission and wanting to make computing accessible to children everywhere. So, this is a bit of a two-pronged question: So, first of all, how do you think you’re doing in relation to that, but also, what are other of the biggest impacts you’ve seen the Raspberry Pi make outside of that core mission and maybe the most surprising ones, or maybe the biggest ones, or, you know, however you want to approach that.
So I think the core mission is interesting. We, I think if you kind of think about, I mean, to put it in context, we saw this enormous decline in the number of applicants to Cambridge. So we have, you know, the Raspberry Pi Foundation still has a very simple scale of metric for how things are going, right. Which is how many in December of each year, how many people picked you up and try and get into the University of Cambridge to study computer science. And we’ve seen this decline from, you know, 600 ish, at the height of the dot com boom, down to about 200 to 250 in 2008. So that’s kind of still though the aspirations of the foundation have grown enormously beyond that. Um, that’s still the number I look at if I want a single metric to show me how things are going.
On that basis, things are going pretty well. Right? So last year, we have on the order of 1400, um, applicants, um, to the university. So we’re kind of getting on for 2.5x the number of applicants that we had at the height of the dot com boom, when people fought that, um, competing was a meal ticket, you know, when people thought it could be. So that was really good, right. I think that it’s actually kind of the upside and it took a little while to get there. And I think we took a non-obvious road to that destination. So if you think who was buying Raspberry Pis in 2012, it wasn’t teachers and it wasn’t really children. It was adult technologist, hobbyists, geeks, people like me. At the time there, I think there probably was a little bit of a sense within the organization of disappointment that we weren’t kind of doing this kind of steepest descent straight to every child on the Raspberry Pi.
But what we didn’t realize and what we, and came to realize later is that those people are your way in, right? So those people, they are parents, they are teachers, they are volunteers of after-school clubs, they’re the people who are going to take your product into the original target market. And so what we saw off to two, three, four years, was some of that enthusiasm bleeding down from the kind of this kind of super geek world down into kind of regular adult humans and then regular little humans. So that was really, really good, that was a really positive development for us. And so we kind of go up to this point via an indirect route where at least our scale up metric looks pretty good.
Well, I think that’s all the, all the good stuff I would say. So what, what are the bad things? Um, I think we’ve come to realize that the challenges are much larger. You know, if you sort of thing that we’re an organization that started off trying to recreate the 1980s, we were trying to build up pocket version of the, um, the, the circumstances that provided in the 1980s and the led to that being a large number of applicants to be a science in the 90s.
Actually, when you look at the 1980s, it’s not a fantastic place in a lot of ways. Many of the demographic challenges that flipped our industry now, you know, most of us are male, most of us are white. Most came from middle-class backgrounds. The unrepresentative nature of the employee base and the technology industry is a consequence of the unrepresentative nature of the hobbyist community in the 1980s and nineties.
There’s definitely a strong standard trying to do better. And hey, you know, we have some very encouraging leading metrics. So a very known 50% of the participants in code club is our network of after-school clubs for nine to 13 year olds, all girls. And that’s really powerful, particularly at the upper end of that age range, which is the kind of point at which something in society starts to discourage girls from being involved in STEM subjects. So the fact that we can acquire and retain people’s interests through that is kind of a pretty good sign that we’re all gonna get to a better place.
The other statistic I love is that if you look at schools in the UK, if you use the proportion of children who received free school meals, as a proxy for deprivation, if you’re in the top quartile for deprivation, you are slightly more likely as a school to have a code club than if you’re in the bottom quartile.
And that’s really encouraging. So I think we’re finding even within the UK, we’re finding kind of new challenges, we’re finding new things to do to make a difference. And of course, you know, the UK only has about 1% of the world’s population. We have a lot of work to do in the developed world, there’s an enormous amount of work to do with the developing world. There’s probably actually a pretty good product market fit between the Raspberry Pi and some of the needs of educational computing. We kind of increasingly draw this distinction between computing education, which is our original wheel house, that’s kind of where we come from, teaching people about computers, and computing for education, providing computing devices that people can be educated in all sorts of subjects.
And there’s quite a good alignment between some of the modern Raspberry Pi platforms, like the Raspberry Pi 4, and the needs of general-purpose computing for education. And we’ve seen that in the context of the pandemic, you know, we see a lot of people using Raspberry Pi as a general-purpose computer for home learning. We’ve done some philanthropic work ourselves in this area with the support of a number of a number of donors through the Bloomfield trust, primarily, and that’s been done has been very successful for us, and we think there’s probably a pretty good product market fit in the developing world for people to want to provide their children with the computing experience. So those are all kind of, that’s the kind of core. And I mentioned that we got to the core via this hobbyist thing.
The other thing we didn’t appreciate about the hobbyists, of course, is that many of them are professional design engineers. And that’s probably led them to the other strand of probably more surprising stuff with Raspberry Pi, which is the people take Raspberry Pi with them into that work. And you know, somebody will buy a Raspberry Pi at home to hack on, and the next time they’re asked by their boss to do, uh, a piece of work, it will end up in their “tool bag”– it will end up in the, the collection of things that they might choose to use. And so that’s good. And then now, you know, presently, we sold 7 million Raspberry Pis last year, and we just got past 50 million units, total mark half last year, over half of those were going into some industrial occupation.
NOTE: corrected below to 40 million
So of all the Raspberry Pis that have been sold, almost half of them were sold last year. I mean, up to the end of last year.
So of all the Raspberry Pis that have been sold, so 40 million. Sorry. I should enunciate clearly. Um, yes, 40 million. So, roughly one in six. Basically in a decade, roughly one in six was sold in the last year.
ALEX: Still pretty staggering. Yes. So, that’s incredible. So, sorry, go ahead. You were talking about the broader impact.
So I was saying is that, you know, what, you actually, what you find that is the interesting thing for us, is we haven’t had to make any compromises during that process. So the same things that make Raspberry Pi an attractive computer for education and make an attractive computer for industry, you know, the robustness is actually a little bit, it’s an interesting question. You know, what is the, what is the more demanding environment? An oil rig or a child’s bedroom? Yeah, well, it’s not completely obvious to me,
(laughter) No, absolutely not.
It’s a combination of robustness, low cost, low power consumption, high performance programmability of being a general-purpose computer. Those are all incredibly attractive things to people in the industrial applications. And so there may come a day where we have to make some sort of choice, but at least in the core single-board computer, the core SBC product, we’ve always been able to do the same thing and have a single unified product.
You see kind of some divergence now where certainly last year you saw us launch Compute Module 4. So, the Compute Module line, which I know you guys are very familiar with. CM4 has been a really surprising early success for us, you know, our model. So, so for those people who don’t know, Compute Module products pack the core logics, generally the SOC, uh, its memory, its power supply, circuitry, and more recently, some of it’s networking peripherals. So a wireless and wired networking, um, into, uh, into a module format rather than a computer format integrated into a product. Now our model for how modules work, for how the sales modules work as well, is that we, um, we sell a few in the first year, uh, you know, thousands, maybe 10,000 of an investment, then those go into a design process and then you get that design.
And that’s really what happened with Compute Module 1, Compute Module 3, Compute Module 3+, very, very slow sales intentions, or the sales. And then it’s kind of a rapid growth later on.
Strangely with Compute Module 4, maybe because I think we’ve kind of got the feature set right from the point of view, simplifying integration into end-to-end user products. All of a sudden, particularly the power supply design is simpler. You just give it five [volts] where the previous ones there were some power sequencing responsibilities that the main board had, and with Compute Module 4 you just give it five volts and you’re done. We’ve integrated wireless networking onto the motherboard as well. So it’s kind of quite easy to get started with. So we’d sell tens of thousands of CM4s a month, it’s only been on the market for six months, but we’re selling tens of thousands of these a month, which is kind of scary for us actually having to play with our mix, our manufacturing mix, in order to keep CM4 in an environment where all of our products are selling pretty well as well.
So you can see how the whole Compute Module experience is kind of doubling down on the industrial side. Then you see probably 400, which we launched in November of last year . There’s a Pi inside the keyboard as being our first really completely explicitly consumer-focused product. So we do have a little bit of specialization there in terms of all of those Raspberry Pi 4, CM4, Pi400, they’re all built on the same basic core technology platform. We haven’t had to really customize that platform where you’ve got some form factor customization, but no real functional customization as in the different markets. And that’s great long may it continue, of course, to see it helps us do a lot with a relatively small engineering team.
So, yeah. Awesome. And you open up so many, so many avenues to go on, I’m having a hard time picking. I’ll pick one that’s definitely been near and dear to our hearts here at balena, which is the Compute Module– Compute Module 4. So famously the PI foundation has been very, very careful at breaking interfaces, right? And you guys take that really, really seriously, and in my view, it’s been key to your success, right? Because the ecosystem cannot accumulate unless they can expect some sort of continuity. But you know, at some, you know, the risk, uh, that the sort of the return on investment sort of starts to tip, right? And you kind of make the big, uh, leap. So, how was that decision for you guys with, you know, with CM1, CM3, [CM]3+, you kept all of that sort of under the same sort of interface.
We did grow a millimeter in wide. (laughter)
ALEX: (laughter) Okay. You consider that an issue and we didn’t even notice.
Yeah, you’re right. It it’s, it’s, it’s one of those ones where, you know, we, do like continuity, um, and you see that in the single-board computer product, um, where really the form factor was completely stable from [Raspberry Pi]1+ through to 3B+. So you, you get little changes. So, um, the LEDs moved in 3, and the LEDs moved because we needed room for an aerial. Um, and so they moved down there, where they actually belong down to the power supply. The 3+ group grew a Power-over-Ethernet header. That’s just, just a function up. Um, and then Pi4 is the first one that has significant churn in the form factor. So we flipped the, Ethernet around by simply routing on the board. Yeah, one of the ways that we keep the cost of the product down is we use fairly, um, limited level pieces of new technology.
Um, so we use the [inaudible] PCB stack up, which doesn’t give us an enormous amount of reach and capability. And so it would have been extremely challenging to bring the, um, the, the Ethernet down to the bottom right-hand corner, and end up in the top right-hand corner. So, then obviously we had went from one HDMI to two micro HDMIs from right down on the bottom edge, and we went from micro B to C for power supply. So it was quite a collection of individually, relatively minor changes that add up to something just feels substantially different.
Um, Compute Modules is obviously much more advanced and much more aggressive case of that, that we’ve, yeah, we’ve stuck with the JEDEC SO-DIMM, the L2 SO-DIMM form factor, um, for several generations, uh, we got to the point, the, the interfaces that were exposed by the BCM2711s stock that’s on the Pi 4 are so different than the ones on the 2708 or 2710 designs we felt that we had to, we were just throwing away a lot of interfaces. So I think we would have at least thrown away the second HDMI, we probably would have thrown away PCI express. I think we could just squeeze PCI express onto an alligator pin. Um, but we were, we would have lost at least those two and we would have lost the doors into, um, uh, we would have ended up with the older power sequencing approach, or we probably wouldn’t be able to fit Wi-Fi on there in any meaningful way.
So I think we, um, and, and people were not a hundred percent happy about SO-DIMM. SO-DIMM is enormously popular as a form factor for SOMs. Um, but actually that perpendicular connector format that we’ve gone for now, um, is a, uh, is I think more popular, partly because it enables a smaller footprint and meatier carrier board design. Um, so it’s all of those things, like I say, they all stack up together and we did a rare thing, which is compatibility to work.
Yeah. I mean from our point of view as balena, we build the [balena]Fin, which is relying on the Compute Module and, you know, while on the one hand, you know, would it have been great if there was something compatible to CM3, I guess. Sure. And to be completely honest, we were trying to read the tea leaves. Like when the Pi 4 came up, we’re like, okay, how is the PCI? Is there, do they even have the theoretical capability to do this? Or should we just, you know, but honestly, like seeing the result, I, you know, we’re fully, perfectly happy because I mean, one of the things we say within the team is always short term pain for long-term gain. Right? So at some point you just gotta draw a line and do the new thing and, you know…
We changed every generation.
ALEX: Yeah, exactly. Yeah.
EBEN: If we made changes, which were kind of just fancy changes, you know, pointless changes that would be different. But I think, you know, we, people can at least trust us to make sensible changes that we can articulate reasons for something or obviously somebody will be unhappy, but you know…
With, uh, you know, what, we’ve, what we’ve found, you know, once you reach the scale that you guys are at, um, there’s always going to be somebody there’s nothing you can do that will make everybody happy.
Yeah. You know, I mean, it is. So it’s been a challenge for us that was once, as you said, once you reached a certain scale, everything happens. Large numbers, right? Um, you know, so you get every kind of technical problem. Um, you get every kind of customer interaction, um, you get every kind, all, you got every kind of opportunity, you know. You get every kind of every kind of surprise application. Someone discovers it and does it, um, so in, in positive and negative ways, that scale has a lot of consequences.
As you say, it comes with the territory and it also comes with a good, so we shouldn’t, we shouldn’t grovel too much. Uh, so the other thing you touched on actually, which was, I know a lot of people are fighting through right now is component shortage, and like manufacturing and having to make the hard decisions, you know, which child do you feed? I mean, how do you do that? Yeah. Uh, so how do you think about that? I mean, again, you could, uh, there’s the specific, what you’re dealing with, which I’m sure like you’re dealing with every day, and also the, the broad, like what advice you’d give to someone who’s dealing with that in other circumstances.
We’re lucky because obviously we are very scaled. Um, and we have a great team, um, outside of being scaled, we get great support from our suppliers and I mean, it’s, it’s tough for everyone. Um, but we do at least get great support from all suppliers to the extent that they’re able to support us. And you know, it’s not easy.
You gotta remember, it’s not easy for your suppliers. It’s easy to be angry with your suppliers. “Why aren’t you sending me stuff?” But you know that they’ve got suppliers as well, that aren’t sending them stuff. Um, so, so, you know, we’re getting great support from pretty much everyone.
Um, we’ve had to, what are we prioritizing? Um, how we’ve ended up prioritizing the, the, uh, we’re lucky we went in with very, very strong entries, all of, um, of the three products, the, the 2037-based products, so [Pi]3, 3+, CM3, CM3+.
So that was good. That was a bit of luck. You know, your inventory can oscillates a bit of the time and we were on the upswing there, um, when it was pretty good when we received the chipset for Pi 4, not a huge amount of finished goods. Um, but that’s fine cause we can convert chips that, um, uh, yeah. What are we, what are we prioritizing, uh, CM4, differentially, prioritizing CM4, I guess, um, is a big thing.
The more popular variance because CM4 is an enormous product line, right? So, um, it’s got a, uh, there are 32 different, uh, there are four, um, DRAM choices, four flash choices and wireless or no wireless, so there are 32 variances of our product. So we prioritize the more popular variants.
Um, we’ve done some work where we’ve, um, we’ve cut minimum order quantities. We’re kind of in a situation where, um, you know, things are tight. You don’t want there to be units sitting on people’s shelves cause they’ve had to buy units from you ahead. A goal is, what’s the goal this year? I think we’ll get there in good shape, but really the goal this year is make sure that every chip ends up in a device that ends up with an end user. We don’t want anything on any shelves anywhere, no chips on the shelves in the factory waiting to be assembled, no finished goods in the factory waiting to ship. Nothing away at a house, nothing at a customer facility stockpiles because someone’s had to buy, um, a minimum number of units. And actually you find is when you’re in this inventory position, you’ve got quite a channel which has inventory, the delta between at least with inventory at the start and, um, inventory added to a channel, um, and a very lean world where you’re literally consuming every chip that comes in immediately and getting it through to a customer. There’s a lot of, there’s like a million units that, between those it’s a one-shot benefit. Right. But then there needs to be a one-shot benefit because you want to, to get through it yet. Right. Um, and so what you find is if you’ve got your underlying supply, um, and that’s secure, cause you’ve got good support, um, uh, and then you see a surge, cause we have some in addition to type of suppliers, you know, surge of demand, um, or you’ll, you can absorb a lot of that surge by being a little bit clever, a little bit leaner, a little bit more responsive in your, in your supply chain. And that’s really where we are this year.
Um, the biggest surprise for me of course, is the extent to which we’re prioritizing CM4. Um, I absolutely have not seen this surge, you could have told me the start of the year we have, we’re going to have component shortages, I’d have said “that’s fine”, you know, we have more CM4 in the budget than we’re ever going to sell. Now we probably sold getting on for the year’s budget in CM4. Um, that’s a shock to me. Uh, and, but when it’s fine, we’ll, you know, we’ll, we’ll manage through it. And just like I said, if there is one thing we’re lucky we were ever good at inventory cause otherwise we would be in trouble.
Yeah. No, that’s definitely… I mean, what’s constantly been surprising me, uh, if you had, if you had given me a call, I would have told you: the demand for the Raspberry Pi, especially the Compute Module and everything outside the core market, the ones that your, your more complete sort of products, uh, cover is amazing. And honestly, the CM, the Compute Module in general, and the CM4 does that even better, um, is a fantastic way.
Like, so this is, this was our thought process, right? We make the Fin, you know, like I, I have it right next to me. It’s one of my favorite things we’ve ever made, but we were thinking, “Hey, we can let those guys do the hard part,” like all of the, you know, the high speed stuff and, you know, like integrating the Silicon and all that stuff. Like we don’t, we’re not a hardware company, but we can build a, the rest around that based on what we know from our customers, where they keep failing or whatever, uh, we can, we can spend more than what you guys would optimally spend for your use case or whatever, like gold plate it, but the, you do the hard work for us. So we can just take that and know that it’s going to work and adding the wireless on that actually for me is huge because we had to certify for 5GHz. And like, if you could just do that for everybody. Thanks, please. (laughter)
I mean, it was one of the things we did with the core product, with 3+, is we’re, uh, an FCC-certified module. Um, so with 3+, we did [Pi]Zero W and PI3, which are chip-on-board, un-canned wireless. And then what we did with 3, from 3+ onwards is we put a can over the wifi. And what that lets you be is an FCC module. Um, so just like you would go buy a module from, uh, us or someone, the entirety of a Raspberry Pi 3, the entirety of a CM4 wireless version, um, is an even, uh, sticks out around the side of those, the can is a square, it’s a centimeter square, and it’s got a large amount of board sticking out around the can. But from the FCC perspective, it’s a wireless module.
And what that means is, as you know, you don’t need to do, you don’t need to re-qualify the wireless components. You obviously need to do incremental, um, testing for intentional emissions, but you can use our conformance. You can ride on our conformance with it for the wireless side, there’s enormous savings in this. It’s 300-400,000 pounds on conforming a product because we can form every– that’s the other thing we do– we conform everywhere. So, you know, if you’re in Columbia, um, and you want a, uh, and you, you want to use this, you’ll, there’ll be a piece of paper somewhere out in the filing cabinet, which is your “Colombian CM4” certificate. Um, you know, we all, we really are pretty much everywhere. We hardly ever say, “no, this country is not economic for us. We’re not going to conform that.”
So, in that context, obviously the modularity is very important to us and just trying to find ways, you know, you mentioned, you know, being a hardware company and whatnot: what’s the core proposition for Raspberry Pi in the embedded hardware space? We let software companies not have to become unintentional hardware companies.
We see this most in digital signage, and our clients are actually originally there. And there were a lot of companies making hardware only because they had all the, and all of that differentiation was in software. And they were only making hardware because they needed a cost-effective platform to run a stable cost effective platform to run their software on. We walk up and everyone’s like, “Hey, okay, we can, we can get out of this space now!”
So you see a lot of say, um, thin client products, which are transparently a Raspberry Pi 4 in a box. Um, what they’re running is the differentiator software provided by the vendor. Um, so there’s still loads of scope for differentiation, but people aren’t doing this kind of spurious engineering, this unnecessary engineering.
ALEX: Yeah. It’s, it’s actually fascinating. We’ve, uh, we’ve been in the space for, for a while and obviously, you know, interacting with, uh, with, with companies and, and other vendors as well, that, and some of the, sort of the big names in the industry let’s call it that, um, have been, you know, trying to put out single board are some, some more successful than others. But I’ve seen a very specific change in the perception of Raspberry Pi itself, right? In the beginning, it’s almost like the, what was the quote? Like, “first they ignore you, then laugh at you, then they fight you, and then you win.” Uh, but I remember specifically at a, at a CES where I was talking to a vendor and they were saying, “well, you know, we want to build a board for when you know, those people that are prototyping Raspberry Pis are going to get serious and they’re going to move to this.”
EBEN: You know, there’s a board for that. (laughter)
(laughter) But it was fascinating to me because it was almost this, even though it sounded kind of bad, it was also a step of progress because they at least were acknowledging that there was something happening. In the beginning, it was unthinkable, right? Then you saw sort of the form factor leaking over and, you know, the acknowledgement that there’s something going on and now we kind of see it. Yeah. Just going everywhere itself. And, you know, nobody blinks an eyelid because what else would you do?
It’s just better made than any other product. So that’s the, you know, there’s always this, I remember people talking about the space shuttle, right? And the space shuttle did a hundred ish flights all over a hundred flights. Any other airplane, that’s a test program right now to be fair. It never ceased to be an experimental vehicle because it just didn’t fly enough. Um, and you see this with these, you get people who say, “Hey, we would like the industrial Raspberry Pi.” Uh, and you know, these are companies, often they make 10,000 boards a year. Um, and you know, they’re this tiny scale. And you know, 10,000 is the number where we check that we can build a product. We press the button, wait until the machine has spat out 10,000 Raspberry Pis, and then we have a look at them and see if we like them, and are totally prepared to landfill them if we don’t like them.
And uh we get a little bit more, usually we build 100,000 to make sure, but you know for functional use, there are PTV ones, there are DVT ones, design validation test ones, is 100 units. Um, and then a PVT production validation test run is five to 10,000 units. Um, and so you have people who are saying that the industrial Raspberry Pi who were building numbers of units per year, which are kind of comparable to the number we accidentally build if we lean on the button for a little bit longer than we intended to.
And that’s, you know, volume drives out defects. You know, we just go with defects, um, and volume drives them out because if you have, you know, we built, um, or we sold upwards of 7 million Raspberry Pis last year, we did 2.1 million in the first quarter of this year. So we’re at an annualized rate of pretty much 8.5 million. Um, and if you’re building that right, and you have any significant population of defects, then you’ll just end up with a massive room full of defective units.
And so, you know, we’re always pushing on, we’re always pushing on that. Uh, I’m probably taking it to kind of the next level as well with CM4. CM4 is the first product where we are, um, regularly, um, pushing units that we believe to be perfect through third-party teardown. You know, we’ve always done our own work and we’ve also worked with our manufacturing partners for that in-house, uh, in-house um, quality management. But these are the ones where we have units, which are passed everything that we know, um, how to, uh, perfect, uh, pushing through external test, uh, through external tear down, and then looking at the output of that and looking to see whether we can improve the production process. That’s, that’s a, that’s a, that’s a new step. Um, that’s, that’s really a new step for us. So, you know, always trying to find you just try to be paranoid. I don’t think people are paranoid enough. Everybody around me is just paranoid.
I think it would have the, they have the quote for the video now. I don’t think people are paranoid enough!
I mean, if you’re shipping like 7 million of something a year, you should be paranoid! You shouldn’t be sleeping, you know, you should be worrying about what’s wrong. Um, you know, what’s wrong at, you know, uh, 10 PPM, a hundred PPM, you know, you should be worrying about those things, let alone a percent… a percent! Can you imagine if you had a percent defect on 7 million units? 70,000 units. That’s mostly, that was 20 pallets of units. It would just be a nightmare.
Yeah, no, I mean, exactly right. When you get to the 1% of the 1% of being significant, I mean…
I used to have hair. I used to have hair. (laughter) I went through a brief period of having hair and then I, long before Raspberry Pi… Maybe some stuff around the side has gone.
(laughter) Yes. Yeah. The whites here (pointing at beard) have been acquired due to hardware.
I feel like at least for a long time, I was able to tell myself, at least the white hair was falling out, and everything that was left was brown. And then actually down here, not so much now looking like Hampton [inaudible]. (laughter)
That’s awesome. Uh, and, the point you’re making is it’s fascinating to me because we’ve also seen it. And we’ve also seen it be very counterintuitive to newcomers, which is that the… it’s kind of the thing you said about the child’s bedroom and the industrial, you know, set up. Or, um, the, you would expect that the requirements would be diverging, right? Um, in, in sort of ways where you’re like, well, this couldn’t possibly be used here. This couldn’t possibly be used there.
Um, but, and we see this on the software side as well, which I’m kind of, I think I’ve stopped now for, but for the longest time, it was kind of waiting for the requirement to drop that would make it, you know, incompatible, like, uh, for our, for our product. But in general, uh, the space where you say like, well, you know, if you have this, you know, strict bifurcation, but we, what we keep seeing is somebody asking for something.
And this is by the way, how we do support, we always try to give the best support of everybody, because you don’t know if somebody is going to come with a question, right? And are going to be like, “oh, you know, what about this thing?” Um, and if you ignore that, then you’re going to get it from a huge customer on a huge timeline, and you’re either going to have to abandon that, or just run like crazy, and you’re going to be like, “why didn’t I listen to that guy in the basement who had this thing?”, and we said like, “nobody cares.” And now I’m running like crazy a madman.
Always better to fix these things when it’s not your biggest customer.
Yeah, exactly. But yeah, that consistency is just, uh, really, I mean, it almost feels abnormal. Like, you know, that things are so just building a platform like this should be, should have a consistent, you know, it’s almost as like a platonic ideal, and we’re kind of reaching for it, and it doesn’t matter where you’re coming from. Like you’re always drawing the same shape or something. I don’t, I don’t have an explanation for it.
Well, it’s strange. And of course, you know, you can, it’s important to price divergence. It’s important to price it… if you diverge, you basically divide your software team in two. And so, you know, it’s easy to see, and I do this all the time. I have to restrain myself because of course I have this access to this amazing engineering team that can do almost anything. Um, and so it’s really tempting to come in the morning and say, “Right, lads, here are the 10 new things we’re going to do.”
Um, and I have to, I I’m, I’m lucky and having guys like James Adams, and Gordon Hollingworth around me to beat me over the head when I do this. Um, but you know, um, uh, leadership strategy is about, um, the things you don’t do. And one of the things that we really try not to do is to bifurcate the platform.
It’s one of the reasons why we’re lucky that we’ve had this series of, um, one, two, three, four, five chips now. So we had 2708. Okay. 2835, 2709, 2836, 2710 twice. So we had the A1 design that went into, um, Pi 3. Um, and then the, uh, CM3, and then the B’s era design that went into, um, uh, Pi3+ and CM3+.
The difference between those is that they are, the second one is, um, flip chip. So the first ones are wired onto chips, where your chip sits on a substrate, and then you have little bond wires that go from the top, from the die onto the substrate with, if you think about that shape, it’s kind of a coil. Um, so it’s inductively terrible.
At the very highest frequencies, it’s very hard to get, um, uh, energy, um, onto the die. And so we kind of had this hard, upper frequency limit of 1.2 GHz. Um, and then, um, the, um, the B zero chip is a flip-chip. Um, so you, so you have, instead, you have a bumped dial, you flip the die over and bond it down. So there’s no… none of these inducting bond wire things. Um, it’s all packaged with the heat spreader as well. So it’s thermally, thermally very performant.
Um, so we had that twice, effectively that, that chip appears twice. And in fact, we love it so much, we’ve actually rolled it back onto the Pi 3 as well, cause it’s pin compatible, it’s back on Pi 3. The only product now that takes A1 is CM 3. Um, so we have that chip twice, which was kind of fun.
Then we have 2711, which is quite a big change. All of those ones are 40nm chips, progressively larger ARM cores. Progressively clever packaging technology. Um, but all in 40 nanometers. Um, and you put about a 15x performance range across there in terms of ARM performance, um, starting with your single-core ARM11, going to your quad-core A53. Single-core ARM11 700 megahertz, quad A53, um 1.4GHz, so you’ve got twice the clock speed, four times the number of cores, and actually quite a lot of uplift in terms of construction for those two architectures. You’ve got architectural IPC improvement. Um, so you’ve got those guys, then you’ve got 2711. 2711 was kind of, uh, the first full implementation, um, of, of the basic chip concept. Um, but on 28 nanometers… and that’s, uh, obviously that’s going to be the, the big jumps that you’ve got the out of all the cores, then you’ve got the [A]72s… is about another 3x.
So you’ve got about a 40x, um, delta, across the line, but they’re all fundamentally the same thing, right? Where 2711 has grown to more stuff. Uh, but fundamentally, all of these are the same architecture and, therefore, same software team. We still run a 32-bit, we are going to transition more to a 60, we have a prototype 64-bit OS. We’ll transition all probably in the next 12 months, our attention really moves across to that. Um, but we’re still shipping a 32-bit userland.
We’re still shipping an ARM V6 32-bit userland. What does that mean? It means that it’ll run on a Raspberry Pi Zero, which has an ARM11, and it’ll run on a Raspberry Pi 1, it’ll run on a Raspberry Pi alpha board. So I’m sure if you are one of the 50 people who has a Raspberry Pi alpha board from 2011, today’s operating system release will boot on that board. And in fact there were one or two good people from the community who will complain if it doesn’t.
And that’s an enormous level of a from a customer perspective, it’s knowing this level of continuity from an engineering development perspective. It avoids divergence in your, um, engineering. And places where there is divergence, are the places we have the most challenge, or where we have to spend the most money. So for example, the graphics technology, the multimedia 3D acceleration technology boost VideoCore in the 40nm line to the VideoCore VI in the, um, in the 28nm line. Um, and of course what you will do end up doing is paying two driver teams. So we end up spending more money or where you could spend all of that money on making one of those two architectures better. Instead, you split your attention a bit like CM 4– you could afford to do this sometimes, but if you made a habit of it, it wouldn’t go.
Yeah. So it’s about finding the right moment to kind of make the jump. Which, actually, brings me to, uh, something that a lot of people have been curious about, uh, which is, you know, most people know you’re very passionate about semiconductor, uh, design of the integrated circuits and all of that. Uh, so you must have noticed somewhere, uh, unless you’ve been hiding under a rock, which you haven’t, that the RISC-V sort of, uh, movement is getting steam. So, I’m sure lots and lots of people are curious about, you know, your thoughts, and again, not necessarily specific to the Raspberry Pi, as you mentioned. There’s, you know, these kinds of changes are, uh, you know, something that probably would, uh, require a lot of thought, but, um, just more broadly, I mean, just as a person who’s interested in the industry.
It’s, well, it’s an interesting development. How do I feel about it? I think that was a good thing. And so it was nice to have more choice. I think there are some substantial challenges ahead of it. Um, because there’s so much investment in the two dominant architectures: x86 and ARM, on the software side, primarily. So there are, in fact, ARM is much less developed than x86. So if you go and look in the source for FFMPEG or something, what you’ll find on the x86 side is you’ll find accelerators fast paths for every generation of Intel, um, multimedia instruction from them, um, uh, MMX through the reason stuff, ASC2 to whatever. Um, there’s more reasons I need to update my thinking about this, because at least one generation more of this anyway, with the Intel pipe, you have to go and do this.
Actually took a long time for ARM. It’s still not done. ARM is still on 30-year-old instructions– AARCH32s, a 30-year-old instruction set architecture, um, and it will come to the end of its life, which you might see now, before people finished doing this work. Um, so, so I think that that’s, that’s a challenge.
I think that there’s a lower barrier to entry in microcomponents. There was a lower bar of entry because the tooling is, you know, the tooling is only about working GCC pretty much. Um, you’re okay. So that’s, that’s, that’s, that may be somewhere where we see, um, some, some other RISC-V stuff. The instruction set standardization on licensing, um, you can make. There’s a pair of articles called “ the case for open instruction set architecture” and “a case for a license to the structures that architectural professors” written by, um, one of the RISC-V people, and the second one was written by a couple of ARM guys and was in Linley (the Linley Group) on the website.
And well, the open guys argue, we can have the flexibility to do what you want. Um, the ARM guys argue that having someone who curates, um, the instruction set in countries using a variety of legal mechanisms to control what goes into the instruction set creates a much more, um, uniform platform, uh, which can serve… um, you get more innovation when you have a platform that people we rely on. So you don’t end up fragmenting: “Oh, those processes have those interesting instructions that make you do this thing! Oh, I’ll go use, I’ll use those machines.” Um, so, it’s finally balanced actually now. So I don’t really have an opinion there. I think it’s a finely balanced argument.
No, it’s fascinating. And I mean, the way you put it almost brought to my mind, like arguments about like Wikipedia versus a cyclo PDF in Britannica, sort of, of like 10 to 20 years ago, which might indicate it could go one way. But, the fragmentation that you mentioned on the other side is definitely…
Platforms for innovation, these are platforms, these are shared platforms for innovation, and you have to have some centripetal force, um, that pulls the platform together. And the question is, is there, uh, is the RISC-V centripetal force? ARM centripetal force? Intel was incredibly powerful because it’s basically just Intel and AMD making the processors. The ARM force is a little bit less powerful, uh, and then the RISC-V force is less powerful than that. And is the RISC-V force powerful enough, um, to build a common platform for innovation?
Yeah. That’s yeah. Um, I mean the one thing we discuss internally at balena, the natural thing as an engineer is to always want to offer more choices, and more freedom, and more options. But especially when you start to think about things from the whole stack perspective, I tend to model it as a budget. You have a freedom budget, and where are you going to spend it?
Because if you actually start spending it at the, at the bottom layers and everything’s fragmented, there’s less you can do on top. Right now, you’re fighting all the way up. Um, so yeah, I think that’s something that we probably, uh, should be better understood that there’s no free lunch in terms of just giving, you know, knobs to people and just do everything in any way you want, for sure.
I think the thing you’ve got to remember is that they’re all really high-end, um, RISC-V cores available. So, you know, they are Cortex A7 equivalent, maybe at a push A53. A53 is an alarmingly good call actually for an older core. So the interesting, interesting thing about how the on-calls evolved is that you got the in-order and out-of-order lines. So you’ve got kind of three places where they developed cores: they’re making them in Cambridge, they’re making them in Sophia-Antipolis, South of France, and then making them in Austin. And each of them has a kind of a distinctive competence.
So, the Cambridge [ones] were the in-order cores. The Cortex A7, for example, Cortex A5, Cortex A7. Sophia(-Antipolis) were the earliest out-of-order cores. Austin A8, actually, but Sophia A9, the canonical, A9 for a long time was the canonical big arm. All right. Um, and that’s now an out-of-order core. And then Austin would do, because what was interesting with the Cambridge team was eventually they kind of caught up with, by designing a really, really, really, really well-optimized in-order core. A53 kind of caught up with A9. Um, and so A53 is a really good core. Um, and it’s probably the best in-order core. I mean, 55 adds a little bit to it, but 53 is probably the best in-order core that’s ever been made by mankind. Um, it, it pushes a long by the time, uh, Intel got to 53-levels of performance, they’ve gone out of order. They come Pentium Pro, Pentium II. Um, and so it was really fascinating that ARM discovered new reserves of in-order, um, performance with 53.
So 53 is a great call. You might possibly get a push, be able to go out and get a 53-class, um, um, RISC-V core. Uh, now maybe higher-end cores will come. So sort of A72. We’re using A72 at the moment. I don’t think I can currently get a competitive RISC-V core for a Raspberry Pi product.
Now, maybe they’ll arrive, but they are very substantial pieces of work. Um, they’re not knocked out by grad students. Um, you know, these ARM cores, the A70x series cores have hundreds of engineers of work each. Um, and some of these engineer years from a running start. Not, “Hey, I’m going to sit down and run his processor.” It’s like, “I’ve been writing a processor a year for a decade. And I’m now going to spend a hundred engineer years to make next year’s processor.
They tend to do this now. So they tend to, it’s kind of, you see it as having 50 engineers for two years or something, or a hundred engineers for two years. Um, so, it’s incredibly expensive. It’s extremely expensive. And that means that you’re not going to end up with, I do not believe for an instant that you are going to end up with, unless some eccentric billionaire endows an organization that’s composed of very clever people, can’t just be any organization has to be a very capable organization, endows an organization with the intention of making a disruptively high-performance RISC-V processor. Any future high-performance RISC-V processor will be a proprietary product, either have an in-house team. And we may see that we may see that out of China, um, or it will be a licensed proprietary core, like you would see like SiFive.
Um, and so that means actually kind of on some level, what have you gained? What you might end up getting is they’re going to cost the same as ARM cores. They’re going to be as proprietary as ARM cores. You’re going to gain the ability to have other, to migrate to another core from another vendor, which runs the same instructions set which you can’t do at the moment with ARM cores, but that could be end up being a very theoretical increment of, uh, freedom. Particularly, if you probably can’t necessarily sustain multiple organizations doing this. And therefore, you may find yes, you have, you know, um, you know, it will be a disappointing kind of, it’d be just one kind of freedom.
So maybe what we need to do is hope that the eccentric billionaire (laughter) decides that they want to kind of teleport us to the end state, just kind of approaching the end state on the fabrication side as more, as Moore’s Law runs out. Um, somebody decides that they wish to just create the omega processor of the, kind of, the end-of-time out-of-order the processor. Um, uh, let’s go do that. And I’m going to spend a hundred million, probably a hundred million dollars, well spent.
Yeah, the A53 of out-of-order.
Well, I was thinking kind of the A76, A78 maybe. Yeah. A real case would be the A53 of out-of-order so which would kind of be like a A76 or A78. Right. Um, so maybe that might happen, but I think that’s probably what we’re talking about here. I’m not sure. I don’t think we’re talking about this spontaneous flowering of freedom.
Yeah. I mean, I imagine the, the sort of the optimistic case would be that you will have a sort of multiple parties sort of pushing each other and contributing at the same time. So this “co-op-petition” creating unforeseen-like breakthroughs, but that’s theoretical, right? Like as you say, I think the interesting thing is to see it and I think, I mean for me, and I think for people who are listening to this, uh, we, we get from you the much more sort of grounded perspective of, like, let’s see how this plays out, because it’s easy to get excited about like this new thing that’s coming out. Uh, but there’s, yeah, hardware is hard. People aren’t paranoid enough.
Well, I’m a pragmatist, you know, I, I, you know, I mean, you see this with CM4. Like if, if RISC-V ever rocked up that I could license to stick in a chip, um, and it had demonstrable benefits for my users over the, the cores that I’m using, I would, and I could justify and they were demonstrable enough that I could justify the bifurcation of supporting these, for a period, two architectures, then I’d be gone in an instant.
Um, I think I’m just… you know this question of like, “will this thing happen spontaneously?” Well, look at GPUs. So GPUs implement open standards. You know, Khronos do an amazing job of standardizing, um, of standardizing graphics. For months, I used to spend a lot of time at Khronos, very much enjoyed it. Uh, you can go join Khronos, go download the OpenGL ES3.2 spec or the Vulcan 1.2 spec and you can build a GPU. It’s not like an instruction set architecture. It’s not, um, it’s not as tightly defined as an instruction set architecture. It does what it does. It’s the operations you have to perform that are specified at a more abstract level.
I know it’s not the proprietary lock-in that comes with most devices. Um, is there, out there, a GPU, which is freely available, which is remotely competitive? Well, is there one out there at all, which is standards– I’m not aware of any standards-compliant, open GPU. And to the extent that anything exists and even isn’t standards compliant, is there anything that is remotely competitive with what you can get for with ARM Mali, SGX, Imagination Technologies, VideoCore, or the stuff you get from AMD or Nvidia um, that just isn’t anything. And there’s, and there’s no, all of the barriers that prevent you from building a high-performance CPU architecture do not apply to GPU. Um, uh, that’s nothing. So that’s probably my analogy. That’s what I’m thinking in my mind about the likelihood of stuff spontaneously happening.
Cool. Um, okay. So we’re, we’re headed towards the top of the hour, so I want to make sure we, uh, you know, we’re, uh, carefully using your, your very generously provided time. Uh, one thing I think, maybe a slightly lighter topic, but, um, what are you, are you hacking or something in your spare time? Are you doing something that, um, are you working on something that’s, uh, you know, when you’re not thinking about anything else? Where does your mind flow to like naturally?
No, I have, I have a four-year-old and a, well, tomorrow, one-year-old. My son’s first birthday is tomorrow. I have a four-year-old and a one-year-old. Um, and I run the world’s most exciting computer company.
That’s right. That’s amazing.
I don’t have spare time. (laughter)
You know, I’ve been known to do a bit of retro computer programming, um, before my son was born. Um, I quite enjoy, um, BBC Microboss. I don’t know if you’ve come across this. So it’s a Twitter bot that you can tweet a BBC basic program to, and it will send you a video of it running under emulation.
They have a wonderful system that dispatches it to AWS to Graviton-based AWS Lambda. It’s quite a nice little ARM world there. Um, and so they have a version of, uh, BBC micro emulator that runs at about an equivalent speed, about six gigahertz, is equivalent to about a six gigahertz BBC micro. BBC micro is a 2MHz platform. So you can send it and it will see, you can record you and it’ll run an hour. So you can either get a few seconds of video, from 30 seconds, or if you want to do some sort of effect, which takes a long time to render, like, yeah, I have a box set or something. You can tweet it with a little rocket emojis and it will run it for an hour. And then so, we do a screenshot of what you got after an hour. And so this is kind of fun because you have 280 characters to do something clever.
I do a little bit of, a little bit of offline BBC micro stuff. Um, I’ve been trying to play with shader toy recently. I don’t really have the time to build the kind of infrastructure required to do really nice shader toys. So it’s quite nice cause you kind of sit down and just flail it away at it for 10 minutes and come out with something pretty, uh, and then stick it onto Twitter. So that’s quite fun. But my technical life is very, um, attenuated at the moment, um, by the responsibilities to Raspberry Pi and my responsibilities to a family.
Sure, sure. I mean, actually the, um, the, another question in the same vein: what hard problem is sort of on your mind? Again, in the macro, you don’t have to go into the details, but, uh, is there a sort of a macro issue that you’re sort of trying to breakthrough?
In the world?
In the world or…?
I mean, in the Raspberry Pi world.
Well, I mean, the supply chain is everything this year we’ll, we’ll do all right. But, it feels like you’re, you’re trying to, and uh… Mike Buffham, he’s the Chief Commercial Officer, spends basically all of his time now doing supply chain. It’s not his job, but it is this year.
So we are, um, we are, um, we’re confronted with lots of supply chain challenges and it feels like you’re just trying not to die this month, and it’s gonna feel like that all year. So that certainly soaks up a lot of my attention. Um, other things… it’s fun to think about future Raspberry Pis, but a long way away from that, yet. I mean, Raspberry Pi 4, but into the fact it’s an enormous uplift over the Raspberry Pi 3, is correspondingly probably a four-year platform. Probably less through the lifespan of Raspberry Pi 4 at the moment.
So there’s that, um, I think we’ve been about accessories. You know, we love to sell accessories for Raspberry Pi, but um, but also really like just doing generic accessories stuff. So sort of thinking about what new things we could add on to a Raspberry Pi to make it more useful to people. Um, [we] did some great stuff last year with audio.
Um, we brought these IQ audio brand that we acquired and then brought that into our product line. So we now have some really nice high-fidelity first-party audio accessories.
[CAT JUMPS INTO EBEN’S FRAME]
And here comes the cat.
(laughter) Guest star!
[inaudible cat excitement from everyone!]
(laughter) That’s awesome.
I was like, he’s not going to jump… he’s on the sofa over here, and I’m like, he’s not going to jump because that’s going to go really badly wrong.
I think there is some law about cats and podcasts. Or if it isn’t, we should, we should make it up because it definitely is there.
If he, or my daughter, or my son, who’s about to be behind us– there he goes [son passes by camera background]. Um, I want to participate in podcasts or business meetings or whatever. It’s been a glorious thing, actually, about this year. Uh, you know, there’ve been many, many, many terrible things, but becoming the fact that everyone has had to acknowledge that they are not a business robot and that they have families, and cats, and children, and cooking going on that you can hear in the background. You can hear all of this stuff because we’re all in our homes, it has been, I think it’s a positive. Yeah. It’s a small positive to go with the substantial negatives.
Yeah. It’s a fascinating point you’re raising because it’s, yeah, it’s always felt disembodied, right. That you sort of, you know, enter the door of your workplace and all of a sudden it’s almost like you, the old IBM man suits sort of, you know, drapes over you and you are an interchangeable…
So, I used to work for IBM.
(laughter) I’ve heard there are jingles playing at a specific time of the day and all of that.
A lot of the stuff, the interesting thing is a lot of the things that we associate with Japanese corporate culture, you know, sort of the mysterious rather stereotypical things. So all the stereotypical things, the Japanese corporate culture do anthems, and you know company loyalty stuff were actually IBM things because after the second world war, um, the Americans sent people to Japan to help Japan, um, with their economy. Um, so lots of these, lots of these things are actually the product of, um, how IBM saw itself. In the night, we found a copy of the IBM hymnbook, um, the IBM company songbook when we were there. You know, it’s like, you know, these, like, hymns to IBM’s great inventors and stuff.
Um, I used to work with… I have a funny story. I bought a suit to go work for IBM, a blue suit. I went with my father, just leaving school. I was just between university school and university. I left school at 17, went down the next day and bought a blue suit. I went to IBM in my blue suit. First day in, I noticed there was a guy in the corner who… everyone was wearing suits, and there was a guy in the corner wearing ripped jeans, a t-shirt, and a Microsoft hat with the windows 95 start button, and at this point this was when IBM and Microsoft were feuding. It was perfect.
He was wearing this hat. Um, and I was like, and I noticed that everybody was giving him enormous amounts of respect and like it like, cause he was a great programmer. This guy’s, name was Graham.
He now actually works for us.
In fact, he actually wrote the SDK for Pico, uh, Graham Sanderson, for the RP 2014 Pico platform. And he, um, and I was like, hang on, everyone likes him. Next day, I came in ripped t-shirt jeans, ripped jeans, t-shirt, and borrowed his hat.
Wore his hat around all day, nearly got fired.
Um, but you know, there is that IBM-suit-wearing thing that was the mid-90s, and it was still very, very strong at that point, that taboo, against it. And I think, you know, we, you know, I never turned the zoom background on, you know. I won’t, I will not turn the zoom background on when I’m on because I want people now I’m in my house.
No, it’s, I mean, the suit is kind of funny because I think right now, if somebody showed up in a suit, that would be the act of rebellion. (laughter)
Well, we’d like sweatpants where like, you know, there’s going to be in six months. So it was, I was reading a Matt Levine column for Bloomberg. Oh, he was saying, there’s now going to be for bankers, there’ll be a six-month window where you can turn up to a business meeting in sweatpants. (laughter)
And even if you look like a stooge, you can say, “lol pandemic.” Um, after six months, you, if you turn up in sweatpants…
It’ll clamp down again, yeah…
You know, make the most of your six months of fun bonding around inappropriate clothing choices in business meetings.
But it’s, it’s the, the broader point though, I think is, is one that’s rarely made around how, um, we are all real human beings. And I think with the pandemic, as you say, the many, many bad things have come out of it, um, the fact that we’ve all had to acknowledge that, you know, we are, you know, unified beings with like, we don’t actually forget everything the moment we walk in the door or out. Um, and we have to find ways for work to, to acknowledge that I think is, uh, definitely something I think we’re going to hopefully carry with us and not after the “six-month timeline” again, like get amnesia and like… (
I mean, that’s it, you know, just trying to freeze in your mind how it feels, how it feels at the moment or how it felt. Um, and just trying to take away one or two quality, good things. I’ve been running a little more. So I’ll got a little more exercise. I live out in the countryside and I have been getting all my exercise as a result of this.
I’m trying to make time in my day because I don’t know about you, but I find it’s the first thing to go, and it shouldn’t go, because you save some time by not exercising, but then you end up that your productivity declines more because your not feeling locked down…
EBEN’S FAMILY MEMBER:
I’ve been having more babies (jokingly, carrying son).
Yeah! Here’s lockdown (pointing at son). 364-day-old lockdown!
You know, what’s funny, my son is turning two tomorrow as well.
He is? Oh happy birthday!
So, it’s a very similar situation. Yeah.
Yeah. And so, so, you know, so it’s those things as well, let’s make sure that you take something positive with us.
Yeah. For sure. (baby screaming in background)
I can’t think of a better, better line to close on. So, I’m just going to, like, choose this time and draw a line and uh, you know, quit while we’re ahead.
That’s great. Thanks. Thanks so much.
Wonderful Well, thank you very much, indeed. That’s been great.
(fade out to the wonderful sounds of parenthood…)