You may not have heard of Adapteva, but the
company makes some bold claims about its potential to revolutionise the mobile
market. According to the company’s self-promotional activities, it’s developed
world’s most efficient multi-core processor architecture. That may sound like
vapourware, but Adapteva has already released products, including a 16-core
chip based on a 65mm production process and an impressive 64 core 28nm chip.
The
tiny Epiphany cores are assembled in a gird-like fashion to create a many-core,
low-power processor
Not content with selling a few chips,
however, the company has revealed plans for an ultra-compact microcomputer such
as the Raspberry Pi, but offering a claimed 50GHz of CPU-equivalent processing
power. Better still, it plans to sell the board to enthusiasts for just $99,
and has turned to crowd-funding platform Kickstarter to see its dreams become
reality.
Bare bones supercomputing
The Parallella board envisioned by
Adapteva’s chief executive Andreas Olofsson and his team is certainly
impressive. With a dual-core Zynq ARM Cortex-A9 processor running at 1GHz, 1GB
of RAM and gigabit Ethernet, there’s plenty of power for general-purpose
computing, while a pair of 48-in general-purpose input-output (GPIO) headers
provides support for talking to external hardware.
Inevitably, the board draws comparisons to
the ARM-based Raspberry Pi. ‘There are lots of cheap platforms out there
running dual-core A9s, so that’s not the reason people are going to buy our
platform,’ Olofsson admitted. ‘People are going to be excited about our
platform because of the openness and the parallel computing. The platform
itself is going to have a dual-core Cortex-A9 on it, initially running Ubuntu.’
This is a respectable basic computer in itself, but the really big deal is that
it also has Adapteva’s co-processor – the Epiphany.
This highly parallel, power-efficient chip
can be quickly programmed in C or C++ and, offering 16 cores on the $99 version
of the board and 64 on the $199 version, it promises the level of
parallel-processing performance normally associated with desktop graphics
cards. ‘We don’t run an operating system, but we’ve 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. It has 64-cores running
at 800NHz, consuming less than 2W for the full-chip. That’s around 50GHz of CPU
performance; this is 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.’
While multiplying the number of cores by
the clock speed to reach a high figure is perhaps questionable, Adapteva has
another metric to trot out in these scenarios: the latest Epiphany-IV
architecture boasts performance of 70 gigaflops per watt. It’s hard to compare
this to anything else at the moment, as high-end graphics cards are clearly set
up very differently from the Adapteva board, but a top-end Nvidia Kepler GPU
can manage around 20 gigaflops per watt.
The
company’s first 28nm part, the Epiphany-IV, packs 64 cores into just 10mm2
of sillicon
The community problem
So if the Epiphany architecture is so
impressive, why aren’t we already using it? Olofsson readily admits that uptake
of the chip has been slow outside niche markets. ‘A processor architecture is
only as strong as its community,’ he explains. ‘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
single-threaded way.’
That’s where Parallella comes in. ‘We feel
we have technology that’s 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. It can be very hard to do that if we keep the price high. 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 price it at $99 for a parallel computer, it should be cheap enough for
anybody to have it.
‘For me, personally, the reason I’m an
engineer is to build stuff that other people use,’ Olofsson adds. ‘That’s the
dream of a chip designer: you build a platform – a blank canvas – and then
other people will make amazing things with that.’
Open source
Olofsson has also solemnly promised to
ensure that technical documentation, complier source code and so forth are
available publicly under a permissive open-source licence – something that’s
relatively race in the semiconductor world, where companies such as AMD, Intel,
Nvidia and Broadcom keep their details locked under onerous non-disclosure
agreements.
‘It only hurts the end-user community to
close the platforms and keep it secret,’ Olofsson claims. ‘Some 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.
‘If you look at companies that
traditionally have a very broad horizontal following with lots of markets –
companies such as Altera, Xilinx, Texas Instruments, and to some extent, Analog
Devices – they’re selling to lots of small customers, and they’ve done very
well with that. The applications that have come out of that are very
innovative: Beagleboard, Pandaboard and Arduino are examples of platforms with
very big followings. I would urge an semiconductor company to open up their
platform as much as possible.’
Parallel education
‘Everybody knows that the future’s
parallel,’ Olofsson enthuses. ‘What GPUs are showing, and what we’re showing,
is that parallel isn’t 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 with the energy and knowledge to re-write a lot of applications for it.
In the future, it’s just going to get worse. Single-threaded processors are
saturating, and I think there’s agreement on that, so you need to go to
heterogeneous computing, and to do that, you need to educate all the new
programmers who come out from scratch.’
With a $99 board, Olofsson argues that it
will be considerably easier for schools and universities to teach many-core
processing concepts, bridging what he claims is a significant skills gap in the
industry.
Funding parallella
Kickstarter, or at the very least, the
concept of crowd funding, is key to Parallela’s potential. ‘To get the costs
down to a reasonable point, we need to increase the volume,’ admits Olofsson. ‘We’ve
never going to get the volume up selling onesies and twosies to R&D labs.
We need a larger audience. One of our biggest development costs are mask sets;
at 28nm, they cost millions of dollars, at 65nm they’re hundreds of thousands
of dollars, and if we can get some pre-purchases through Kickstarter, we can
use those pre-purchases to basicablly buy mask sets and bring the cost of the
chip down to something very attractive.’
Once the upfront costs are dealt with,
Olofsson claims that the Epiphany co-processors found on the Parallella boards
will be very cheap to manufacture. ‘We’re extremely small in terms of out chip
size. Our 64-core processor, at 28nm, is only 10mm2 – around 3.5 x
3.5mm. Compare that to large GPUs and large microprocessors; they’re hundreds
of square millimetres just think what they sell for or costs to manufacture.
The majority of the cost of a chip is the silicon, so this is very
cost-competitive.’
Success in sight
‘Our big goal is for this to be as successful
as the Raspberry Pi,’ claims Olofsson, ‘but addressing a different market,
obviously. We’re not at $35 or $25, we’ve at $99, but with vastly more
performance.’
