HackSpace: Raspberry Pi 5 – what is it?
HS: So over 10, 11 years of development you’re pretty much keeping up with Moore’s law. Nice!
Eben: We’re about on track. Eleven years is about seven iterations of Moore’s law [Gordon Moore’s observation that computing power doubles every 18 months], which is 128, and we posted 130, so I think we are more or less clinging by our fingernails to Moore’s law.
James: It works very nicely as a desktop machine now. I mean, Raspberry Pi 4 did as well.
Eben: We have this idea that there’s a kind of bell curve of demand for computing. There are some people who don’t really want very much performance at all [Eben points to the left-hand side]. And then some people who want as much performance as they can get – extremely high-end gaming, or CAD [Computer Aided Design] programs. [Eben points to the right-hand side.] Some people, say, just want to read a value from a sensor once a minute in a Python script and log it to the network and turn an LED on and off – Raspberry Pi 1 could do this, of course. So, there was a market for Raspberry Pi 1 as a PC, but only for people with relatively modest demands. People’s ambient expectation of how much computing power they need drifts to the right, but it doesn’t charge to the right. So, over the generations, we’ve been catching up. The interesting thing about Raspberry Pi 4 was that it got past that median. I could give it to my parents, to use as a generic PC. They can’t really tell that it’s not a MacBook Pro, because they don’t need a MacBook Pro’s worth of performance. With Raspberry Pi 5, we’re now going back down the other side of the slope: most people will be in the same camp as my parents
A custom-designed system on a chip
HS: What’s new on the Raspberry Pi 5?
Eben: Let’s talk about BCM2712. It was developed by Broadcom. But it was co-developed to a great degree.
James: It’s architecturally pretty similar to BCM2711 [the SoC in Raspberry Pi 4]. It’s got faster Arm cores, more cache, beefed-up fabric, better display pipeline, faster multimedia blocks, like the GPU and image signal processor. We always wanted a user PCI Express port, and it’s nice to squeeze one on. It’s not M.2 or any sort of standard because it just won’t fit on the board. But we will sell adapters to take that and turn it into an M.2.
HS: There are a lot of things that you can plug into PCIe. What are you expecting to work?
James: Almost everything! BCM2711 had PCI Express root complex that wasn’t fully compliant with all the corners of the standard – I mean, PCI Express is a very complicated standard. Devices like PCs just sort of put all the features in and they beat them up to make sure the corner cases work. This isn’t always the case with [Arm hardware]. But Broadcom, with our help, have done a much better job of trying to make the new root complex more standards-compliant, so it should work with more devices.
Eben: That includes things like Coral [the Google AI Accelerator]. [PCIe on the Raspberry Pi 4] would always work with things like NVMe [solid-state drives] but 64-bit register accesses didn’t work due to a bug in BCM2711, and that permanently meant you could not connect to the Coral. I think honestly, if people at Google could have worked a little bit harder to help us debug what was going on there, we probably could have made it work. [But on Raspberry Pi 5] Coral should work. Whether giant PC graphics cards will work, and Jeff [Geerling, a YouTuber famous for trying to hook PCI graphics cards to Raspberry Pi] will have a good experience: search me, I don’t know. There will be fewer bugs, and there are more supported features. But your guess is as good as mine as to whether that collection of features will be sufficient to resuscitate an AMD graphics card.
HS: And what about the processing cores?
Eben: We’ve gone from a quad Arm Cortex-A72 at 1.8GHz [on the Raspberry Pi 4], with a 1MB shared L2 cache, to a quad Cortex-A76 at 2.4GHz. Each core has its own 512kB L2 cache, and they share a 2MB L3 cache, so they’ve got a lot more cache memory – you’ve got a total of 4MB of cache in the system. They’re architecturally more performant – so they’re about 50% faster clock-for-clock, maybe a little more than that – and they’re at a higher clock frequency too. And those effects stack together for about a two-and-a-half times performance [improvement over Raspberry Pi 4].
Power for the peripherals
HS: The power system has had an overhaul. Can you tell us about it?
James: We’ve worked with Dialog (now Renesas) to build this chunky power management chip. You can see there are a lot of inductors around it. These little black guys are inductors, and these four grey things at the top are inductors. It’s a quad-phase 18–20 amp switcher, which supplies the core of the chip.
Eben: You have two-ish megahertz PWM [pulse-width modulation], so every half a microsecond you connect the five volts to the output node for a fraction of half a microsecond. The more load it’s under, the greater the duty cycle of the PWM. And that’s how most of the switchers work, how the rails work. But with the quad-phase rail, you have four of these machines all feeding the same node through their own inductors, and those periods are offset by an eighth of a microsecond, so they are at 90-degree phase offsets. And that means that you can respond faster [to changes in the board’s power demand]. At launch, Raspberry Pi 4 had a single-phase core supply. But the newer boards have dual-phase, and this is now quad-phase. Going from single-phase to dual-phase is what enables the 1.8GHz operating point on Raspberry Pi 400 and the more modern Raspberry Pi 4s.
HS: And, of course, the big news is that there’s now a power switch!
James: We’ve got a power switch, which works much like a PC or laptop power button. It has a soft and hard power-off modes: if you touch the button when it’s booted, it will tell Linux to shut down; if you hold the button for long enough, it powers off hard, by cutting the supply rails. The other thing that this power management chip does is it talks ‘PD’ to USB power supplies [Power Distribution – the standard for negotiating with the type of USB-C power supply used by higher-power devices]. USB PD provides a serial interface to the USB power supply that a chip can talk to and say, ‘Hey, what voltages and currents can you give me?’, and then it can choose one. [Raspberry Pi 5] will always choose 5 V, but we now have the ability particularly to talk to the supply to check that it can support 5 amps. [Ed note: Raspberry Pi is launching a 5 V, 5 A power supply, but currently most other USB PD power supplies won’t support this mode.] If you don’t have a five-amp supply [for example, the Raspberry Pi 4 power supply was 3 A], it restricts the USB current output quite aggressively, to 600 milliamps maximum, instead of about 1.1 on the Raspberry Pi 4. So, you can run your mouse and keyboard, happily, but it won’t boot from USB mass-storage devices: it will decline to do that by default.
Eben: This is one of these really difficult engineering things that we had a lot of discussion about. When we say the board can consume 12 watts, we mean that it can consume 12 watts if you craft a horrible use case for it that’s deliberately designed to do nothing useful, but to consume 12 watts. So, we have 600 milliamps and a 1.5 amp mode for USB. Raspberry Pi 5 selects between them on the basis of whether it detects a 5-amp power supply or not. Or, you can stick something in config.txt that overrides it. And the vast majority of people with an existing [3-amp] power supply will plug it in, set the override in config.txt and just forget about it.
James: It’s still five volts, because we don’t want to do power conversion from, say, nine volts, which would be what most people use to get more power into the board. They get more voltage in, and then they convert it into five volts. We don’t do that because it’s costly in silicon, and it’s costly in wasted energy, which just ends up heating the board up. We’ve done the more Raspberry Pi thing, which is make a supply that can drive a five-amp load at five volts, which isn’t a standard PD mode, but you can negotiate it.
Eben: We encourage people to buy the new supply!
Making robots redundant
HS: The bottom of the board is a bit less spiky than before – what’s changed there?
James: The soldering technology is different. We’ve taken out one step – one big step. Instead of having robot arms put these [through-hole components in] and splashing molten solder underneath, we use surface-mount.
HS: You say surface-mount, but it does still go into the holes?
James: That’s right. The SMT – surface-mount technology – process is that you take a screen with little holes and you squeegee your [solder] paste over the top. The pick-and-place machine puts the components down. Then you run it through an oven which melts the solder and everything sticks. You do that for the back, and then traditionally you do that for the top without any of the things that go through the board. And then you do those through‑hole parts as a final step, with wave or selective soldering. But we’re no longer doing that last step. Part of the reason is that now the pick-and-place machines can pick up bigger things, because we’ve got these new heads, but also we’ve worked with the connector manufacturers to trim the pins and put higher temperature plastics in various places, and iterated on the footprints and the hole sizes and the paste, so you now paste everything on the top layer, the pick-and-place machine does every single component, and then you just put it in the oven. With through-hole things, the paste on the top melts and it goes into the holes. Job done!
Eben: Except … we discovered a fringe benefit of having these pins sticking through the board, which is that, along with the SD card connector, they protect the bottom of the board against mechanical damage.
HS: Is that what the little loops are for?
James: Yes, these things are to stop the capacitors smashing on the desk.
HS: So, there’s a surface-mounted through-hole pin?
Eben: Ha! Yes! Two of them!
HS: I’d assumed they were ground test points.
James: They are sort of useful if you’re prodding things. They are grounded.
Eben: It really is quite flat. [To James] Did you pick ones that are the same height?
James: Yes, they’re the same height as the SD card holder.
The first Raspberry Pi chip makes an appearance
HS: Can you tell us about the new chip on the board? Is it called Project Y or RP1?
James: Project Y is our internal name for RP1. It was roughly a $15m programme. It’s what would be called a southbridge in a PC – so it’s the I/O control. It talks to the Broadcom chip over four-lane PCI. That’s the big slug of wires that go from one chip to the other.
HS: So obviously on Raspberry Pi 4 there wasn’t a Project Y, there were a bunch of other chips that did an equivalent job.
Eben: There were actually two places where I/O functionality was provided on Raspberry Pi 4: some of it was provided by the I/O expander chip, the VL805 USB 3.0 controller and hub, and quite a lot of it was on the core silicon. In this generation, all of that stuff has been schlurped out of the core silicon and integrated into this I/O controller. MIPI is no longer provided by the Broadcom silicon. Analogue television is no longer provided by the Broadcom silicon. GPIO is no longer provided by the Broadcom silicon. USB, Ethernet, those are all provided in our I/O controller. So all that’s left in the Broadcom silicon is an enormous amount of high-speed digital logic – so CPUs, GPUs, the DRAM interface, the HDMI (it’s very high bandwidth so you can’t push that off to an I/O device), and the PCI Express that talks to the I/O control. From an analogue perspective, the core silicon is now much simpler. Why is that good? Because it means you can take that core silicon down to smaller process nodes without worrying about ‘oh, how do I provide 3V3-tolerant ESD-hardened GPIO pads on a three-nanometre device?’ You don’t need to worry about that any more. That’s what we call a chiplet architecture.
James: That’s right. We certainly appear to have reinvented (or pre-invented?) the chiplet architecture. And also, the analogue doesn’t shrink very well into the small geometries, so it costs you more and more as you shrink the digital stuff in the analogue stuff. And it costs you time and engineering to actually shoehorn it on, test it, and make it work every time. HS So, the point of that chip isn’t the I/O itself, it’s to allow the CPU to be faster and better?
Eben: It’s to simplify the core device. It’s better in some respects than before. So a simple example: your VL805 chip is a single USB 3.0 controller with an integrated hub. Yeah, so you have five gigabits of downstream bandwidth. But you only actually have four gigabits of upstream bandwidth. So, you can’t even get all of those five gigabits: it’s shared between the two USB 3.0 ports. But you can’t even get all five gigs because you only have a single PCI Express, so you have four shared between both ports. [Raspberry Pi 5] has four-lane PCI Express, so you have 16 gigabits of upstream bandwidth, and each USB 3.0 gets a full five gigabits of bandwidth.
HS: And this is the second chip that Raspberry Pi has released. How does it compare with RP2040?
James: It’s a lot more complicated. It’s the first chip that we built the chip team to build back in 2015. It was the start of the programme. It’s why the chip team exists. Which is also why it’s called RP1.
Eben: I’m amused that nobody ever asked us about this. RP2040 has RP2 written on it, and people think that’s short for RP2040. And the chip that was on the Zero has RP3 written on it. People think that that’s somehow a reference to Raspberry Pi 3, because it’s the same die, but it’s not: RP3 is our third chip; RP2 is our second chip; RP1 is our first chip.
Inflation: They fought the law, but the law won
Eben: The Raspberry Pi 5 cost $25 million to make. And it’s composed out of three enormous engineering programmes (BCM2712, RP1, and the DA9091 PMIC), and it itself is an enormous engineering programme. That power supply is itself a vastly larger engineering programme than pretty much anything other than we’ve ever done, apart from Raspberry Pi 4. Raspberry Pi 4 was a few million dollars. It wasn’t cheap. But the Raspberry Pi 4 was probably a $5 million programme for us.
HS: And in terms of cost benefit, I guess, does that pay-off come from selling more units or making more profit per unit?
Eben: It comes from continuing to sell units. It’s a Red Queen’s race, against our competitors, and against the ever more demanding expectations of the public. That’s the thing: you’re running to stand still. It’s to make the same number of units and make the same amount of money per unit, but to keep doing it for another few years. Now, in practice, of course, we’ll make less money per unit on this than we do with the Pi 4 – this is going to be $5 more expensive than the Pi 4, but it costs more than $5 more than a Pi 4 to make. So, we’ll make less money. But I think we will sell more units, because the market’s growing, and we’ve got extra features and performance.
HS: This is the first Raspberry Pi to cost more than $35.
James: Silicon has got more expensive. People are charging us more for silicon.
Eben: And, it’s silicon that is more complex. It’s the first time our cost structure has materially degraded between generations. [Pi] 3+ and the 4 are probably naturally five bucks more than their predecessor [but Raspberry Pi didn’t pass this cost along to the consumer] and the [Pi] 5 is five bucks plus that.
HS: So it’s inflation catching up?
James: It’s the silicon pricing that has really driven it because we expected it to go down, but it’s gone up.
Eben: You [to James] were modelling it 30% down, and it’s actually 30% up.
HS: So, you didn’t start the design cycle saying, ‘Let’s spend an extra ten bucks’?
James: I spent more on connectors because I expected it to be 30% down on the silicon. And we’ve got a switch. All of those add a few cents.
Eben: And we’ve got a power switch! I would pay five bucks for a power switch. We’re way behind inflation. If inflation had stayed at its low level, we’re probably where inflation would have been, except inflation hasn’t done that. It’s not desirable. We really enjoyed swimming against inflation. Who knows, at some point in the future we may find some way to undo some of the damage, but for now we are adrift.
James: I think people value the fact that not much changes between generations. You get more of stuff and it gets faster for the same price – although the price is going up now. I’m pleased with the innovation that’s gone into stuff like the soldering process. This is the most Raspberry Pi [model] ever, in terms of all the stuff that we’ve touched on it. We’ve had a hand in almost everything. It’s quite a triumph, I think. Let’s hope there are no bugs.
You can find out more at the launch announcement at raspberrypi.com/5
To get HackSpace magazine delivered to your door every month, and the ability to order a Raspberry Pi 5 from our reserved stock of computers for subscribers, head to hsmag.cc/subscribe