Andreas Olofsson is a man who knows his chips. Educated at the University of Pennsylvania, where he gained a BA in Physics, a BS in Electrical Engineering and an MS in the same discipline, Olofsson worked for semiconductor giant Analog Devices for nearly a decade designing highly parallel digital signal processors (DSPs).
Growing tired of working for someone else, Olofsson would leave Analog Devices in 2008 – bringing a few of his colleagues along with him – to found Adapteva, a privately held semiconductor company of just five people.
Adapteva might be small, but the firm has a big aim in mind: to revolutionise the semiconductor industry with its extremely low-power yet highly parallel processor architecture.
Since successfully raising Series A funding in 2009, Adapteva has launched a series of remarkable chips based around its Epiphany highly parallel architecture, including the Epiphany-III 16-core chip and its bigger brother the Epiphany-IV 64-core chip. Despite drawing around 2W under load, the chips promise masses of power for parallel processing tasks – and to prove it, Olofsson is looking to sell a miniature ‘supercomputer’ based around his design to anyone who wants it, for just $99 (around £62 excluding taxes.)
Dubbed Parallella, the board mates a dual- core ARM-based Cortex-A9 processor with an Epiphany-based co-processor. The result, Olofsson claims, is a device no larger than the highly successful Raspberry Pi, but with the sort of processing power that has people sitting up and taking notice.
Seeking $750,000 of cash on the crowd- funding platform Kickstarter for his project, Olofsson promises to release all the details of the Parallella project – board design, chip documentation and compiler source code – under permissive open source licences.
Linux User & Developer: What makes the Epiphany architecture your company builds special, compared to existing architectures such as x86 and ARM?
Andreas Olofsson: “We don’t run an operating system, but we are really good at real-time processing, maths acceleration, and the kind of thing that the ARM and Intel processors can’t handle today very energy efficiently. We’ve been building chips over the last few years, four chips so far, and the latest one came out this summer: the 28nm, 64-core Epiphany-IV.
“It’s a 64-core co-processor that’s going to sit next to an ARM or Intel host, running at 800MHz, consuming less than 2W for the full chip. That’s about 50GHz of CPU performance – that’s the way we like to count it, because it’s 64 real RISC [reduced instruction set computing] cores that can run a lot of different applications written in C or C++ that are being launched from the host.
“One interesting application was the OpenCV library, which is a very popular computer vision library developed by Intel over the years. We took that library and ran it on an x86 processor for face detection. Then we took the inner kernel of the face detection and brought it over to our platform, and started accelerating it and running it at much, much lower power. So basically let the OpenCV machine vision library do all the camera interface and everything like that, all the high-level application, and then we do the brute-force acceleration.
“If I go through what we’ve done over the last year, besides getting the 28nm chip out and verifying it and having it work, it’s really working on our platform, working on our compiler, our debuggers (we put out an OpenCL compiler to allow people to do parallel programming very easily on our platform) and then also building our board offering portfolio to our partners. So, really, the last year has all been about beta customers, early adopters, to get to the point where now we’re ready to bring it to the mass developer audience.”
LUD: What made you decide to launch a Kickstarter project (the first semiconductor firm to do so) offering a highly parallel development board for just $99?
AO: “Until today, our kits have cost thousands of dollars and we’ve had hundreds of people interested in our technology, but they couldn’t afford it. We feel that if we can make it priced at $99 for a parallel computer, now that should be cheap enough that anybody could have it.
“A processor architecture is only as strong as its community. That’s been shown time after time: if you can’t get a few thousand users involved in your architecture, to build software and infrastructure, you’re going to have a very hard time surviving – especially with parallel computing where you’re up against the old way of doing things, the old software ways, the single-threaded way. We feel that we have a technology which is very compelling from an energy efficiency standpoint, but if we really want to have long-term success we need to grow our community very quickly, and it can be very hard to do that if we keep the price very high.
“We’re never going to get the volume up selling onesies and twosies to R&D labs. We really need a bigger audience. One of our biggest development costs is mask sets: at 28nm, they’re millions of dollars, at 65nm they’re hundreds of thousands of dollars. If we can get, through the Kickstarter project, some pre- purchases, we can use those pre-purchases to basically buy mask sets and bring the cost of the chip down to something very, very attractive.
“We just needed some kind of platform to do the launch on, and [Kickstarter is] going to be a way for collecting people’s orders. I think they’ve done a great job of setting that up, they’re a good escrow platform, but it wouldn’t be the only one. For us, it’s easy to use, we don’t have to worry about payment. For us it’s the contract: the all- or-nothing kind of funding really works for us, because the only way we can pull off this project is if we get certain funding levels. For example, to get the $99 16-core version, we would need to raise $750,000 to make that and to not lose a lot of money per board shipped, and it’s that kind of all-or-nothing – people can give their credit-card numbers safely and know that they won’t be charged unless we hit our funding goal.”
LUD: The $99 price point you’re targeting is extremely competitive. How sustainable do you think that is? – is that something you’re going to be able to stick to after the Kickstarter project has finished and maintain as a product?
AO: “With the right manufacturing partners, absolutely. It comes back to our silicon-area efficiency: we are extremely small, in terms of our chip size. Our 64-core processor, at 28nm, is only 10mm2 – so that’s about 3.5mm by 3.5mm. Compare that to large GPUs and large microprocessors, which are hundreds of square millimetres… Just imagine: they’re hundreds of square millimetres, and what they sell for or cost to manufacture, and we’re only ten.
“The majority of the cost of a chip is the silicon cost. So we know that we have a very cost-competitive solution. We have a pretty good track record of delivering on time and within budget, but it’s a challenge to bring a product to market for $99 and ship it in high volume. If this is a huge success, finding the manufacturing partner that can produce in those kinds of volumes and committing to that very quickly is our plan.”
LUD: Why is it so important to you that the Parallella board is priced within the reach of the mass market?
AO: “For me, personally, the reason I’m an engineer is to build stuff that other people use – that’s the dream of a chip designer, that you build a platform and it’s a blank canvas, and then other people will make amazing things with that. That’s the reason we’re doing it.
“Our big goal is that this is going to be as successful as the Raspberry Pi. That’s the end-goal. But addressing a different market, obviously. We’re not at $35 or $25, we’re at $99, but with vastly more performance.”
LUD: One of your target markets for the Parallella is education, much like the Raspberry Pi. How important is it that universities teach the next generation of programmers to harness the parallel processing capabilities found in modern hardware?
AO: “I think it’s a huge challenge and problem right now. I mean, everybody knows that the future’s parallel. What GPUs are showing, and what we are showing, is that parallel’s not even the future – it’s now. There are massively parallel systems right now that could give a huge boost to applications, but there’s nobody who has the energy and the know-how to rewrite a lot of applications for that.
“In the future, it’s just going to get worse. Single-threaded processors are saturating, and I think there is agreement on that, so you need to go to heterogeneous computing, and to do that you need to educate from scratch all the new programmers who come out. The curriculum needs to change immediately, because we’re losing time.
“All the people out there who are experts at single-threaded programming and who maybe haven’t taken the plunge to parallel programming, they need to be educated and retrained and gotten up to speed. It’s a big context switch for the mind, to go from serial to parallel programming, and really the only way to get fundamentally higher speed-ups is to rewrite your program code from scratch. There’s really no magic bullet.”
LUD: If programs needs to be rewritten to take advantage of the Epiphany co-processor, can you see the Parallella board appealing to a mass-market outside developers looking to experiment with the architecture?
AO: “It’s a little bit chicken-and-egg. Obviously, once the development community is there, there is going to be the software to do all kinds of things. The platform itself is going to have a dual- core Cortex-A9 on it running, initially, Ubuntu – so it is a fully fledged computer, shipping.
“To use the power of our co-processor, for that people are going to have to develop software applications. One of the things we can envision is maybe early adopters will buy the platform as part of the Kickstarter project, but that there won’t be any software to run on it yet – so they’re going to have to wait until some of the developers develop the software, but they’re still going to have a platform that’s going to be great, it just won’t have the full offering of software on it.
“People are going to get a board shipping, and when they get it it’s going to be a computer. It just won’t have the really flashy, high-performance applications that will come later. There are lots of cheap platforms out there running dual-core A9s, so that’s not the reason they’re going to buy our platform. People are going to be excited about our platform because of the openness, because of the parallel computing, and that’s still something that needs more development.”
LUD: You’ve mentioned the advantage of openness – could you elaborate on your plans to open up the Epiphany platform as part of the Parallella project?
AO: “Part of enabling the community is – if this project works and if it gets funding – we’re going to be opening up all documentation. No more secrets about what our architecture and chip can do. All the software development tools are going to be open-sourced and given away for free, so that’s kind of what we give to the community. If the community wants to give back, well, you know, we hope they will take the platform and open-source their software and show off what they can do with it, but that’s up to them.
“If you look at companies that traditionally have a very broad horizontal following, lots of markets – companies like Altera, Xilinx, Texas Instruments, and to some extent Analog Devices – they’re selling to lots of small customers, and they’ve done very well doing that. The applications that have come out of that are very innovative.
“BeagleBoard, PandaBoard, Arduino are examples of [open] platforms with very big followings, and I would urge any semiconductor company to open up their platform as much as possible. Why not? It really only hurts the end- user community to close up the platforms and keep it secret. Some of the companies do that because it’s very competitive, right? It’s a dogfight between different chip companies to eke out a profit, but it doesn’t help the end-user at all.”
For details on Parallella, including a video of a prototype running Ubuntu, visit the project’s Kickstarter page.