Parallella, the $99 supercomputer

[Edvard]

They've reached their Kickstarter goal, so I hope the goods appear soon, this looks simply delicious.

Parallella: A SuperComputer for Everyone

Inspired by great hardware communities like Raspberry Pi and Arduino, we see a critical need for a truly open, high-performance computing platform that will close the knowledge gap in parallel programing. The goal of the Parallella project is to democratize access to parallel computing.

http://www.kickstarter.com/projects/adapteva/parallella-a-supercomputer-for-everyone
http://www.parallella.org/

from teh intarwebs

[f0dder]

Claiming that $99 gives you a "supercomputer" is IMHO a bit of a marketing stretch, but the project is pretty interesting - and the architecture seems interesting. Will be interesting to see what real-world performance is like (including perf/$ and perf/Watt), not least compared to GPUs and Intel's Xeon Phi.

Alluring that you can program it in "standard C++", but you obviously still have to be able to parallelize your code (each core is relatively slow), and I wonder to which degree you have to be locality-of-reference aware (and to which degree you have to be Parallella-architecture-aware, possibly reducing the portability), given the small amount of per-node memory, the shared memory and the mest structure.

[Edvard]

From the specs and interviews with the designer, it's a 64-core Epiphany co-processor sitting next to a dual-core ARM CPU, so I wonder how much is managed by the system, and how much is bare metal. It's got expected virtual speeds up to 50GHz.  I know, in modern computing terms, Gigahertz is a trivial benchmark, but for a chip that consumes less than 2 watts, it's impressive.  
Apparently though, the object isn't exactly speed, but functionality; an inexpensive platform for learning how to program for parallel computing, almost like a hardware emulator of more serious iron, to make it easier for students to get into parallel and multi-threading concepts now, just when it's starting to grow.

From an interview in Linux User & Developer magazine:
http://www.linuxuser.co.uk/interviews/parallella-supercomputing-for-the-masses

...
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.
...

[Edvard]

Oblig:
CYhaL8hY-0Y#!

[f0dder]

From the specs and interviews with the designer, it's a 64-core Epiphany co-processor sitting next to a dual-core ARM CPU, so I wonder how much is managed by the system, and how much is bare metal. It's got expected virtual speeds up to 50GHz.  I know, in modern computing terms, Gigahertz is a trivial benchmark, but for a chip that consumes less than 2 watts, it's impressive.

I'm not saying it's not a nice chip, and it does sound like it packs quite a punch - and I love that they're saying they want to be open about it all. But you're not going to reach that top speed unless you have something that's massively parallel - lots of things are hard to split across threads (and the single-core performance of Parallella is low compared to x86). Other things are hard to do without synchronization which, apart from hard to program correctly, can mean massive performance drops (I hope the shared memory / inter-core communication is very efficient!). Then there's also the thing about GHz by itself being mostly meaningless, you also need to know how many cycles the various instructions take, and an amount of other factors

Apparently though, the object isn't exactly speed, but functionality; an inexpensive platform for learning how to program for parallel computing, almost like a hardware emulator of more serious iron, to make it easier for students to get into parallel and multi-threading concepts now, just when it's starting to grow.

And that's what I wish people would focus on, instead of the silly "supercomputer" claims - your quote from Supercomputing for the masses is spot on the sugar. Parallel computing is important, and reducing what used to take a cluster down to a single chip is awesome! Heck, even if the chip didn't deliver more performance than an octa-core x86, it would still be more usable for teaching scale-out parallelism.

[mahesh2k]

99 bucks computers are not much scalable for manufacturer.

[Edvard]

You mean, not much profit margin?  The guy touting this is the founder and CEO of the company (Adapteva) who designed the co-processor.  He says the biggest hurdle is the initial mask for fabricating the chips.  After that, it's all silicon and overhead, and since the penny-sized coprocessor chip uses 1/10 the silicon of a standard CPU, it's naturally going to be less expensive to manufacture.  Also, consider the Raspberry PI and Arduino at less than $50 per unit, and those projects are doing rather well.  If Mr. Olofsson's idea works out like he plans, I think he'll do OK.

[mahesh2k]

No, I mean if you consider the product for recurring earning, that part of scaling is very low. It's not like apple boxes, where there is profit to be made from every official hardware parts. Not that I want restriction like that but just saying that It becomes harder once you multi-continent with sales.

[Edvard]

Ah, I see.  Yeah, I'm thinking that he's banking on the educational and experimenter angle rather than selling OEM stuff for the end-user; make it so affordable that students and hackers can't afford NOT to get a board and learn some parallel and multi-thread programming, and see what amazing things get done by the hardware geek community as well.  Yes, the "supercomputer" moniker may be a little overblown, but the architecture and framework are there.

I wonder how fast a system one could make with a cluster of these...