On 27th October 2012 Adapteva Inc successfully raised almost $900,000 via the Kickstarter crowd funding website. It’s not the largest amount ever raised on Kickstarter, nor is it the most over-subscribed offering. What makes it unusual is being the first such funding success by a chip design company. The money is to be used to manufacture Adapteva’s Epiphany many-core FPU chip and put it on to a $99 single board computer. Known as the Parallella project, the aim is to make “supercomputer” performance available to a much wider market. The image to the right shows the prototype board plugged into a Xilinx Zedboard during development.
Embecosm has been working with Adapteva since 2009, developing the compiler tool chain for their processor architecture. Conventional wisdom has it that modern chip design costs tens of millions of dollars, yet Adapteva are on their fourth generation of silicon — which uses industry-leading 28nm technology — but having required barely $2.5m in funding. Indeed the first silicon was paid for from the founder’s savings.
Make no mistake: this is no ordinary chip. With 64 floating-point processors running at 1 GHz and delivering 100 GFLOPS performance in total, the latest generation of Epiphany is comparable in performance to top end GPUs yet has an energy consumption of just 2 watts. Supercomputer power is now within reach of the battery powered embedded marketplace.
Epiphany processors have been available commercially for nearly two years. Adapteva Inc has been making a steady income selling development boards at around $10,000 each to the defence and other high value markets. Yet you do not make a success of a processor design by selling hundreds of expensive development boards. You need the chip to be adopted in mass-market products, so you can make them by the hundreds of thousands, or millions.
The major chip companies spend a small fortune on marketing activities, trying to persuade product designers to adopt their particular processor. While intelligent engineering meant Adapteva could develop their silicon for a fraction of the usual cost, there was no way, as a four person company, they could invest in the scale of marketing activities of the established chip companies.
It was here that Raspberry Pi gave inspiration. Conceived by a UK-based charitable trust as a teaching platform, it has gone on to become a global success and prototyping platform of choice for certain embedded applications. Around 1 million units have been shipped, and more importantly a host of alternatives have seen their popularity rise – mbed from ARM/NXP, Discovery from ST, BeagleBone from TI and so on.
While there may not be that many hobbyist/hackers/prototypers out there, it is clear there are a great deal more than previously thought. Reducing costs, the ready availability of components over the Internet and ever improving software and tutorial material is making technology increasingly accessible. But far more importantly, in a world of social media, these people are being elevated to the position of opinion formers. They produce the magazine articles, blogs and YouTube videos that a much wider market now relies on.
Another consequence of widespread adoption is that with so many people working on it, Raspberry Pi has found uses its designers never conceived of. Some of these are simply fun, such as a robot arm to deliver beer from the fridge, but others are much more serious, e.g. Raspberry Pi in satellites. The point being the larger the crowd, the greater the chance of innovation. This ability to popularize a new technology through mass availability is what inspired the Parallella project — nothing less than a supercomputer version of a Raspberry Pi. Most important a crowd of new developers will find those new applications where massively parallel, low-power consumption, high performance floating-point computing brings value. Already developers of software-defined radio and UAV technology are looking eagerly at this new platform.
Ultimately, innovation in new product areas is what matters. Kickstarter does not secure Adapteva’s long term success; to do this it needs to have a mass-market product. But with nearly 5,000 developers and many exploring new areas, the chances of success are greatly increased.
Open Source and Partnerships
Central to the success of the Kickstarter funding was an absolute commitment to Open Source. Adapteva have always been strongly committed to such an approach: we developed their GNU compiler tool chain, and the initial pre-silicon version was developed using Verilator models of the hardware.
For Parallella, Adapteva have recognized that their company value lies solely in their unique chip design, and everything else serves only to help sell that design. Therefore everything else must be made as widely available as possible and that means making it free — free in the sense of freedom. Thus the board design, chip interfaces and comprehensive documentation are being made freely available. The goal being that the mass of developers should feel no restriction in what they can do with Parallella.
Alongside this Adapteva has relied on partnerships and a shared commitment in their success, with companies that are typically similar in scale to them; a start-up can experience challenges using a large supplier — they are just too small to secure commitment.
Embecosm is not much older than Adapteva and about the same size. When we agreed to build the GNU tool chain it was our first major contract, and Adapteva knew our success depended upon making a good job for them. That was reflected in our commercial relationship. At this scale, fixed deliverables and fixed price contracts would not work. So Adapteva told us how much they could afford and what their priority list was, and we did the best we could. This made us work very lean, knowing that we had to deliver the maximum result for the effort we had funded.
We delivered a working GNU debugger in just three weeks and our first GNU C compiler was delivered with just two outstanding regression failures in just six weeks. As both companies have grown, we have been able to do more. The GNU C compiler now does an excellent job of vectorizing loops to ensure that both integer and floating-point pipelines are kept full the whole time.
This approach has paid off. Standard benchmarks running on Epiphany have pipeline occupancy that are the envy of the competition, and the resulting performance indices clearly reflect this. The value of any modern processor depends on its tool chain and by working in partnership, we have ensure the Epiphany processor has one of the best there is.
What Could Possibly Go Wrong?
Kickstarter is not a magic bullet: although backers are not equity investors they still have to be kept involved and informed. Having built up a community through the fund-raising process, that community needs to be kept on board as the product is developed, and that has a cost. Kickstarter is also not the cheapest source of funding as they themselves take 5% commission, and the Amazon payments system through which backers pay takes a bit more. So the $900,000 raised translates to closer $800,000 in Adapteva’s bank account.
Backers have been promised Parallella boards – multiple boards in the case of the big backers — and out of that $800,000 Adapteva have to produce around 6,000 boards. It is volume, but not high volume, and it will be a challenge to get manufacturing cost below the target price of $99. Leaving Adapteva just $200,000 to fund all its other needs — its staff, its partners and its operational overheads.
For Adapteva’s partners, their success rubs on to us. At Embecosm we are very proud of the high performance implementation of GCC for the Epiphany chip, and that has served us well in finding new customers.
But making boards is not the sole long term objective. A key goal is to find a major investor to take the momentum from Parallella and to help drive use of the Epiphany processor in mass market products. The mark of success, and indeed the future attractiveness of crowd-funding to chip designers will rest on how effectively that objective is achieved.
This blog post is based on an article first published in “The Ring”, the journal of the Cambridge University Computer Laboratory Ring in January 2013.