Parallella: The $99 Linux supercomputer
Summary: Move over Raspberry Pi, here comes Adapteva's Parallella, a low-cost parallel chip board for Linux supercomputing.
Chip-company Adapteva announced on April 15th at the Linux Collaboration Summit in San Francisco, California, that they've built their first Parallella parallel-processing board for Linux supercomputing, and that they'll be sending them to their 6,300 Kickstarter supporters and other customers by this summer.
Linux has long been the number one supercomputer operating system. But while you could build your own Linux supercomputer using commercial off-the-shelf (COTS) products, it wouldn't be terribly fast. You needed hardware that could support massively parallel computing — the cornerstone of modern supercomputing.
What Adapteva has done is create a credit-card sized parallel-processing board. This comes with a dual-core ARM A9 processor and a 64-core Epiphany Multicore Accelerator chip, along with 1GB of RAM, a microSD card, two USB 2.0 ports, 10/100/1000 Ethernet, and an HDMI connection. If all goes well, by itself, this board should deliver about 90 GFLOPS of performance, or — in terms PC users understand — about the same horse-power as a 45GHz CPU.
This board will use Ubuntu Linux 12.04 for its operating system. To put all this to work, the platform reference design and drivers are now available.
Why would you want a $99 supercomputer?
Well, besides the fact that it would be really cool, Adapteva CEO Andreas Olofsson explained:
Historically, serial processing [conventional computing] improved so quickly that in most applications, there was no need for massively parallel processing. Unfortunately, serial processing performance has now hit a brick wall, and the only practical path to scaling performance in the future is through parallel processing. To make parallel software applications ubiquitous, we will need to make parallel hardware accessible to all programmers, create much more productive parallel programming methods, and convert all serial programmers to parallel programmers.
And of course, Olofsson added, to "make parallel computing accessible to everyone so we can speed up the adoption of parallel processing in the industry", the Parallella had to be created. Olofsson admitted that his company couldn't have done it by itself. The project required, and got, the support of other hardware OEMs, including Xilinx, Analog Devices, Intersil, Micron, Microchip, and Samtec. The companies have enabled Adapteva to bring its first per-production boards to San Francisco, and soon, to its eager programmer customers.
Related stories
Kick off your day with ZDNet's daily email newsletter. It's the freshest tech news and opinion, served hot. Get it.
Talkback
Bitcoin Mining
Bitcoin Mining Machine
I wonder how it does perform on floating point operations, and if it does support CUDA instructions. But those really are a job for the GPU unit, aren't they?
Bitcoin Mining Machine
Bitcoin Mining
Not really...
"266 kilohashes per second, per chip. Not great for mining."
About the same as a Cortex-A8 running at 600mhz.
It would be worthless for bitcoin mining.
Linux: It's over.
Sir Name = dumb-koff
Don't forget the 4th mole Cynical99
It's over Sir Name go find Loverock Davidson and Owlllllnet1 & Cynical99 and have a good cry.
Linux: It's over.
Go spread your poison somewhere else.
Your comments are way off base.
Who am I?
So let me get this straight. In a "good enough" computing era, this is supposed to do something in sales? If I'm a developer, I'm developing for what people have and use now, not in a decade or more.
Your missing the point of the device entirely.
Personaly I immediatly started imagining a cluster of these bad boys set up in a render farm configuration. At the price these are going for it would be far more economical to buy twenty of these at the same price as two decent PCs with ok gpus. And this is coming from a consumer, not a developer.
Supposed to do something in sales? You've nissed the point.
Right now, for less than some i5's, you can buy an 8 core amd processor. I can pretty much guarentee you that that chip will not reach it's full potential as most applications currently consider two threads to be multi-threaded.
We can easily build chips such as these 64 physical core chips, plus hypertheeading right now. These enableassive computing power incredibly efficiently; they can run at lower clock speeds, using less power and producing less heat.
In a clinate of good enough conputing why would you wat this? Because from smaller packages using less power and with smaller cooling solutions needed, everyone wins. Forget the word server for a second. Consider the word tablet. Yes your future tablet cpuld easily run at a slower clock speed using less poser with more cores than your current one, utiliting advancents in fabrication tech as well. But it needs the applications to take advantage.
This project is designed to promote interest in, and development of mega-threaded applications. Not to sell in the millions and take over a market. Not everything is about dominence. Sometimes things are just done to help push things forward.
It is a commercial product
Living up to your screen name there..
My wording wasn't very clear, and I apologise. The purpose of the board is to promote multi read developement. As a stand alone product it's not the kind of endeavour that you would back finacially; it has a much smaller market to appeal to that even raspberry pi, and they haven't just made a motherboard with off the shelf bits on it. As a stand alone product it will likely never pay for the R&d of their accelerator chip and generate significant returns on ivestment. As a stand alone product it will be a commercial flop. This is why the product has been crowd funded.
You are quite correct that the company has additional motives behind development; to promote their chips, their server products, brand image, get publicity, etc. it's what separates them from a charity.
I should have inserted the word "relatively" for you. As in relative to a company that did want to "sell millions and take over a market"
I disagree
One of the core techniques my employer uses right now is completely parallel. One of these small babies could massively speed up our data computation. I can think of many other processes where the parallelism could be harnessed. A small rack filled with these cards could be put to amazing use.
Massively parallel computers have been used for transaction processing,
Banking, for example, could have many thousands of people accessing their accounts, all at the same time, and doing the exact thing, like withdrawing money and updating the accounts, all within the same computer, which contains many processors. The processors all handle their own workload, but the workload could be for the same exact type of transaction, or many different types of transactions.
Each processor could handle a process, or a set of code to handle a transaction. If you had a computer with 1000 processors, all of them could be doing the same thing for 1000 different users, "at the same time" within the computer. Or, the 1000 processors could be handling a thousand different kinds of transactions at the same time, for 1000 different users. Multi-threading on a single processor would not be needed, since the 1000 processors could handle the work-load quite easily. But, you could do multi-threading in a single processor, which would then give your 1000 computer the power of a super-duper-computer.
The problem with massively parallel computing is that, most programmers don't know how to spread the workload efficiently, and that's something that the OS for a computer should automate and take away from programmers.
I don't understand...
Does it only support a minimal set of instructions that would make it unsuitable for anything else?
It is simply a misrepresentation
They received enough negative feedback that they address this issue in their FAQs; however, they simply try to explain away their number and point fingers at others doing something similar to access GPU performance. Unfortunately, they miss one important distinction - the numbers they point fingers at are theoretically possible to attain, whereas, their chip will never be running at 45Ghz.
Just to drive home the absurdity of this notion: using their logic I can take a Nivida K20 GPU with 2496 thread processors and multiply by 745Mhz and get a 1.86 Terahertz CPU equivalent or 1860Ghz...
It's not a 45GHz computer
It would be great if applications were written to use the power of these 64 computers; but most applications were written to run in a single computer, and it's not trivial to parallelize them.
Imagine a very simple program like:
for number=1 to 10:
print number;
In a single computer this produces a very predictable output:
1,2,3,4,5,6,7,8,9,10.
If this program was executed in parallel, it is very difficult to predict the order in which the values are going to be returned; so the output could be something like this:
3,7,5,4,1,2,9,6,8,10.
Now, imagine if the order of the numbers was important -- for instance, the steps in a nuclear facility control system.
Parallel computing is a very difficult subject, and that's why most programs don't use it.