by Pete Lomas via Wired Magazine, opinion column
“A raspberry ‘pie’?” “You want to build HOW few and sell them for WHAT price?” We had always wanted to build Raspberry Pi in the U.K., but early conversations suggested this was a vain hope. The Foundation had zero credibility in the manufacturing arena, our price point meant margins were way under acceptable levels for contract manufacturers, and volumes were simply too low.
The original plan was to build just 1000 units for new undergraduates at Cambridge University. We figured we could kickstart that ourselves: if we sold each at $35 and it cost $36 to make, well, we would lose $1000. That was worth it, we thought, to engage with keen early adopters who would help us with the debugging, documentation, and education.
We desperately needed a sustainable model. As a U.K.-registered charity, we could not fund demand by normal loan methods. Nor could we rely on our own money to subsidize production at a level over 200 times what we had initially expected. The Foundation was already running on donations by the trustees, some of whom had remortgaged their own houses to provide funds.But instead, our small, part-time only team was faced with a success disaster. Just three weeks before launch, initial demand was well beyond 200,000 units. And at one point on launch day we were even trendier than Lady Gaga.
What we learned is that you have to sell out (a little) to sell (a lot). I’d argue that many makers have to do this when they want to scale. Going from 0 to 1 million units projected in just our first year meant we had to make a number of tradeoffs – from where to manufacture and how to work with partners without alienating our core community, to what features we chose (or didn’t choose) for the Raspberry Pi in the first place.
Manufacturing: You Can’t Have Your Pie and Eat It Too
We had set the target prices at $25 and $35 for Raspberry Pi before we had finalized our design. Sure, we had a prototype, but we had to hone it because manufacturing has a range of tradeoffs that overlap with the design choices.
Ideally, you want to frame manufacturing volume at the outset so the most appropriate technology can be linked with the appropriate manufacturing processes. High volumes mean efficient processes and discounts on components, resulting in tantalisingly low manufacturing costs. Low volume processes recognize the difficulty of not being able to spread out costs – but they replace specialized capital equipment with higher labor costs instead.
Yet we always believed it was possible to manufacture the Pi in the U.K., where labor costs for electronics assembly at the printed circuit board (PCB) level are low enough as long as we sourced the components internationally and reached sufficiently large volumes.Raspberry Pi needed a modest volume and a minimal cost point. The charitable model helped us negotiate component pricing not normally available to a small business, but the only way we could realistically get both was by priming the pump in China, supported by personal loans. So we negotiated with a Shenzhen-based contract electronic manufacturer to build the initial 10,000 to 20,000 units.
Enter the partners. Premier Farnell and RS/Allied had the buying power to keep component prices low, the global presence to handle the logistics, and the financial muscle to make it all happen immediately. And when we wanted to come back to the U.K., Sony’s Technology Center in Wales could handle the technical package-on-package assembly.
So partnering would be the shortest route to getting Raspberry Pi into our community’s hands as quickly as possible. It also meant we could help create 30 new manufacturing jobs in the U.K.
But it meant we were faced with a dilemma. How could we enable hacking while preventing cloning?
Holding back the schematics altogether troubled us. Not being open would impede people’s ability to interface and hack the hardware – defeating the very goals we had set out to accomplish with Raspberry Pi in the first place. Because our remit is education in the broadest sense, we wanted – needed – to provide completely open access to the hardware. And we didn’t want to alienate the devoted hacking and open source community that had fueled early interest and would provide much of the future development.
Well, you don’t get owt for nowt as they say in North England – you don’t get something for nothing. So we decided to publish the schematics, but hold back the detailed Bill of Materials (BOM) and physical PCB design or “Gerbers“ for a limited amount of time. After all, hardware is just one part of our overall plans. The schematics alone don’t provide enough information to clone the Pi without expending considerable effort re-laying the PCB and figuring out the exact part used in each location.But if other manufacturers copied the design, our partners would lose their investment, which was approaching several million dollars. They were spending this time and money optimizing their processes for manufacturing and distributing our product, while exploring component alternatives to meet the aggressive cost targets. Those component choices weren’t arbitrary: we had to get the selection and underlying layout just right because earlier Pis had failed when the amount of supporting electronics overburdened the design and inevitably the cost.
But if you want to hook up other electronics, the schematics are a great help. Especially since hooking up other stuff to the Pi – whether you’re inside or outside the classroom – is the whole point. How else would a whole new generation grow up actively building with computers, instead of just passively playing on them?
Design: Yes, You Can Have Your Pie and Eat It Too
“Houston, we have a problem.” Well, we never said that (apparently the Apollo 13 astronauts never did, either), but the power-sequence scene in the Apollo 13 movie really captures the final design process for us. The guys on the ground are constrained by time, power, and only materials available in space; they’re frantically adding some essential function and then having to cut something else; they’re trying to keep it all under weight. The Raspberry Pi team wasn’t constrained by time, but as interest grew it certainly felt like it. And we had plenty of other constraints.
While we agreed on the vision of educating the next generation of programmers and engineers, we disagreed on the best way to get there. We all came from different backgrounds – hardware, software, ICT, electronics, hacking, academia, you name it – and with those different perspectives, we were holed up in a room for hours trying to decide what design tradeoffs to make.Price was critical. The financial risks of over-enthusiastically hacking and accidentally breaking the Pi had to be acceptable. But even though it was supposed to be a bare-bones computer, the Pi had to be fully functional and sufficiently powerful. In a world of iThis and X-that, an 800 x 480 display and 8-bit processor just weren’t going to cut it.
Here’s how we decided what to include and exclude. I think we would have gotten it very wrong if we had left some things off. And even after all that, we still encountered a really low point when we realized we might not be able to manufacture the design we ended up with….
Form factor. Someone threw out that the single-board computer should be “as big as a credit card”. Being an engineer, I took this literally. Turns out credit cards are perfect for multiple Lego bricks (which has led to some inventive case and stack designs). More importantly, this also saves on expensive laminate.
Processing. We had to have enough processing power to run Linux, supporting Python, Scratch, and other educational software. A 700MHz ARM CPU provided just enough power; any more would impact the BOM cost. For application processing, we went with the Broadcom BC2835 chip. This ended up hardening up a lot of the other specifications (specs) for us, because the chip defines the ultimate architecture.
Memory. The hardware guys in the room thought 128MB would be enough for bare metal programming. But the Linux software demanded more: 256MB. We figured this was enough to get people going, and thanks to Moore’s Law, the same budget would get us more RAM in a couple of years. For non-volatile memory, we had an SD card reader (the Pi has no memory otherwise), and of course, that would also allow for exchangeable environments.
Multimedia. A key consideration here, as with the other specs, was keeping the Pi geographically inclusive. Many regions of the world cannot access or afford HDMI, so we needed a less expensive way to get video and audio: composite and PWM. Meanwhile, digital cameras everywhere meant the Pi needed to support high-resolution graphics while remaining relatively low-power, and the VideoCore IV GPU gave us those capabilities.
Connectivity. Surprisingly, this was the biggest source of debate: How important was it to have internet connectivity? Turns out, it’s absolutely key: Without it, you can’t connect to the XBMC open source media player or get version updates via GitHub. So the $35 Pi had to include Ethernet, and USB ports would be used for all the basic peripherals.
I/O. But perhaps the most important feature of all was providing access to the General Purpose Input/Output (GPIO). The GPIO was key to unlocking the hardware so successfully developed in the Arduino ecosystem: The stuff people could add to and embed with the Raspberry Pi. The GPIO also meant features that ended up on the cutting room floor could be added back by our
ever-inventive community.
ever-inventive community.
Without all of these specs, a hardware community could not grow around the Raspberry Pi. The pieces of the Pi finally seemed to fit together just perfectly.
The problem? We couldn’t manufacture it anywhere near our target cost. It turned out that there was a looming problem routing the PCB to accommodate the design.
And while there are special high-density interconnect (HDI) techniques for densely layered PCBs, those would just reduce yield and increase processing steps not to mention the costs. We panicked when we realized we might have to increase prices to accommodate the HDI technology.This wasn’t just a minor technical detail: The solution to the problem meant our budget would be well and truly blown. The core of the design involves wiring up all the peripheral connectors, support chips, and five power supplies – most of which have to fit under the processor. But we had 253 connections to bring out (the BGA escape) in an area much smaller than the size of a dime.
But failure was not an option – we simply had to find a way.
After much fiddling with all the possible variables, and a few frantic calls to our PCB manufacturers, we had our Eureka moment. What if we could steal the idea of “blind micro vias” from high-density interconnects, but apply it cheaply enough for the Pi design?
Instead of going through all the PCB layers, we made human-hair sized holes (micro vias) that go through only the first couple of layers (blind) – saving just enough space on the other layers for wiring up the other PCB components. At high volumes, these holes could be made quickly and efficiently with a laser. And it only cost a few cents extra. Making these tradeoffs resulted in a relatively simple six-layer board that didn’t compromise power distribution. And it enabled manufacturing at scale.
Raspberry Pi has been a project seven years in the making. It’s often felt like the closer we got to our goals, the further away they seemed. But we’re having the time of our lives. We’ve only been selling for six months, and we’ve sold around half a million units already. Universities are giving Pis to their freshmen as a gift on arrival; 7-year-old kids are sending us videos of their adventures writing Scratch games.
So in the end, I guess you can have the hard and soft little bits of everything.
