SAN JOSE, Calif.–Fabless chip vendor Adapteva Inc. will launch Parallella, a Kickstarter initiative that could fund the development of the startup’s multicore processors and create an open source community for parallel programming.
The startup is asking for $750,000 to pay for a mask set for its 16-core Epiphany chip. If it gets the money it promises to deliver a $99 reference board for the chip.
[Get a 10% discount on ARM TechCon 2012 conference passes by using promo code EDIT. Click here to learn about the show and register.]
If Adapteva, listed among EE Times' Silicon 60 top emerging startups, gets $3 million in the online initiative, it will create a $199 board for its 64-core chip. In addition, it would then release its software development tools, drivers and libraries as open source and publish its chip architecture reference manuals. It also would release Gerber files and schematics for the boards as free open source.
To date, the startup has designed both chips as part of a shuttle program at GlobalFoundries under which the chips use a shared mask set and thus cost several hundred dollars each. Dedicated mask sets would reduce that cost substantially.
Under Kickstarter’s rules, investors have 30 days from the start of the effort on Thursday (Sept. 27) to make an investment. If the company’s minimum goal is not reached by then, Kickstarter reimburses them.
Adapteva describes its designs as based on a homogeneous 2-D mesh of RISC floating point cores. They are geared for use as low cost and low power alternatives to graphics processors for running highly parallel jobs as co-processors to an ARM or x86 host.
"This is about enabling a community to write parallel tools," said Andreas Olofsson, chief executive of Adapteva. "The only way to get access to a competing GPU these days is to sign an NDA," and even then the code remains proprietary, he said.
The Parallela boards would come in two flavors, both using an ARM A9 based SoC running Ubuntu and an open source developer’s kit supporting C, C++ or OpenCL. The 16-core version would deliver 26 GFLOPs, and the 64-core version could provide 90 GFLOPs—a significant leap over alternative boards such as Raspberry Pi.
The startup began sampling its 65 nm, 16-core chip in May. It sampled a 28 nm, 64-core version in July. Both use an OpenCL compiler already available as open source code.
The move is a creative alternative at a time when venture capital is tight for semiconductor startups. "The kind of money we need to create a true processor platform is hard to come by today," said Olofsson.
"The feedback from VCs has been 'show me a customer, and I’ll show you the money.' This way we potentially will show them hundreds of customers," he added.
"If it doesn’t work we will still be fine, but we are impatient and want to get this to the market now," he said. "We don’t have enough traction to get big investors interested in our company, so this is a way to bypass the traditional finance gatekeepers and go directly to people who develop this stuff."
To date, Adapteva has collaborated with researchers at MIT as well as Boston, Northeastern and Halmstad universities.
Please stop calling it a supercomputer.
90 GFLOPS is not even close to 16.32 petaflops (IBM BlueGene/Q - June 2012 fastest Top500)
This is like saying that someone is building a Formula One equivalent race car that can travel 230 miles in 20.7 years (instead of an hour).
A tech reporter should know better, so stop propagating misperceptions....
What this really could become is a small subunit of a massively parallel supercomputer. By itself, it is not a supercomputer!
I agree the title is misleading but I laud the developers efforts.
As an investor, I would not invest in them for the following reasons:
First, the economics don't make sense. Take Nvidia's Tesla as a comparison point:
Nvidia Tesla C2050 (~1TFLOP for ~$1900) = 1.9 $dollar/GFLOP
and Parallella is proposing 90 GFLOPs for $199 = 2.2 $dollars/GFLOP
You can't enter a market place and not be compelling on the price/performance curve as compared to the competition.
Second, You can't add cores without keeping them busy, which means significant memory bandwidth must be supplied. This is the biggest issue when it comes to multi-core computer performance. It's interesting that they say nothing about memory or memory bandwith.
Third, History hasn't been kind to multicore startups. Take the 2008 article:
Five Multicore Chip Startups to Watch
Only one still in existance?
Indeed, power is the key differentiator. This processor was designed for embedded applications.
Nvidia C2050 479W/TFlop
At BlueGene/Q scale that would reduce power consumption to around 1MW (from 8MW), saving around $7M in annual running costs at wholesale electricity prices.