Parallella: A Supercomputer For Everyone

[wallarug]

I found this article on a new project that is taking place...it came up under Google news when I searched Raspberry Pi.

http://www.kickstarter.com/projects/ada ... f=category

Here is an extract from the website:

Parallella Computer Specifications

The following list shows the major components planned for the Parallella computer:

Dual-core ARM A9 CPU
Epiphany Multicore Accelerator (16 or 64 cores)
1GB RAM
MicroSD Card
USB 2.0 (two)
Two general purpose expansion connectors
Ethernet 10/100/1000
HDMI connection
Ships with Ubuntu OS
Ships with free open source Epiphany development tools that include C compiler, multicore debugger, Eclipse IDE, OpenCL SDK/compiler, and run time libraries.
Dimensions are 3.4'' x 2.1''

Once completed, the Parallella computer should deliver up to 45 GHz of equivalent CPU performance on a board the size of a credit card

This seems like a similar project to the Raspberry Pi. I was wondering if anyone was going to support this project. Looks cool for $99 and $450.

[mikerr]

"45GHz of performance" ?

That's probably just adding up clock speeds of all the processors, I'd be more interested if they quoted GFLOPs or (Drystone) MIPS

E.g.

ARM (Pi) : 965 MIPS 1.2 GLOP (24 GFLOP GPU)
Intel Atom : 4000 MIPS 2 GFLOP
Intel i7 : 128,000 MIPS 100 GFLOP (Graphics card GPUs can generally do 500+ GFLOPs)

[wallarug]

"45GHz of performance" ?

That's probably just adding up clock speeds of all the processors, I'd be more interested if they quoted GFLOPs or (Drystone) MIPS

E.g.

ARM (Pi) : 965 MIPS 1.2 GLOP (24 GFLOP GPU)
Intel Atom : 4000 MIPS 2 GFLOP
Intel i7 : 128,000 MIPS 100 GFLOP (Graphics card GPUs can generally do 500+ GFLOPs)

they do quote GFLOP:

...A9 based SOC and a 16-core Epiphany-III co-processor chip for $99. The 16-core board Parallella should achieve 13 GHz and 26 GFLOPS of equivalent CPU performance. If the Parallella Kickstarter project reaches a stretch goal of $3M, a higher performance Parallella board will be released that includes the pin-compatible 64-core Epiphany-IV chip. The Epiphany-IV based board should achieve a peak performance of 45 GHz and 90 GFLOPS and will be available to all backers who pledge $199 or more. The Parallella computers will ship with a...

http://www.parallella.org/2012/09/27/pa ... -everyone/

[Gert van Loo]

The 3D processor in the Raspberry-Pi has 24 Gflops.
If you add the GPU (especially in vector mode) and ARM you get close to 27 Gflops. But then for half the price.
Then we can throw in a 2Gbits/sec. input port (CSI) and an output port (DSI) which allows you to build some simple CPU network topologies.

[ssam]

The difference compared to a GPU is that the 64 cores are actual useful independent cores. A GPU is only fast when each core is doing the same operation (just on different data). The Epiphany architecture seems more like the xeon phi, just without the baggage of x86.

calling the 16 and 64 core versions a super computer is a bit of a stretch, but its a lot of cpu power for an embedded board. The 1000 core versions will be quite something though.

[aaa801]

Ive pledged $119 to get one of these if it reaches its target
Should be interesting for distributed maths etc

[axelt]

I've just discovered this project. The original poster didn't mention the best bit:

The Parallella platform will be built on the following principles:

* Open Access: Absolutely no NDAs or special access needed! All architecture and SDK documents will be published on the web as soon as the Kickstarter project is funded.
* Open Source: The Parallella platform will be based on free open source development tools and libraries. All board design files will be provided as open source once the Parallella boards are released.

As has been discussed multiple times on the Raspberry Pi Forums, the GPU on the RasPi and practically all comparable ARM boards are encumbered by NDAs, limited or unavailable datasheets, binary blobs, closed source libraries etc. etc. The Parallella project seems pretty unique in offering a fully open platform, without any closed-source/proprietary bits. For that reason alone the project seems worth supporting.

Originally they were only going to release the documentation when the project was fully funded, but they have now released it ahead of plan to drum up more interest in the projects: http://www.kickstarter.com/projects/ada ... sts/323691.

Now all we need is the boards! I would urge everybody to back this worthwhile project.

[AlThomas]

1 Day left... check it out if you haven't yet! They are very close...

http://www.kickstarter.com/projects/ada ... e?ref=live

[jimjimjim]

I've been sitting on the fence for a long while because:

• Never used kickstarter before
• Dislike amazon, and it's the only form of payment kickstarter accepts
• It looked like they wouldn't hit the $750k goal

Things have changed now though, they're only $50k off the goal with over a day left, and pledges seem to be accelerating if anything (probably partially because of fence sitters like me ). Finally bit the bullet and pledged. If you ignore the bit where they quote 45 GHZ, and try to draw up comparisons to raspberry pi in the faq to try and gain mass appeal, it looks like a promising project (just bad promotion on those two points). Certainly worth a punt at $119 for a board with a 16 core co-processor to play with.

edit: Make that only $45k off the goal, $5k down in the space of 20 minutes

[aaa801]

27k now

[jimjimjim]

It's like ebay in reverse, you hit refresh every 5 seconds wanting the number to go higher

[Wizard]

YES!

They actually made it - and by only a few hours to spare... Talk about cliffhanger
This is the future for sure and now we have affordable hw within reach.

[wallarug]

That's great!

[bounce]

Not to bring this thread up from the dead, but can someone explain the difference between this and Raspberry Pi? I know Parallella uses more than 1 core, but does this mean it's harder to program? All I need is something to build a media center and if Raspberry Pi is suitable for that (I've read reports video can be shaky?) and browsing I'll just stick with it. Thanks

[wallarug]

Not to bring this thread up from the dead, but can someone explain the difference between this and Raspberry Pi? I know Parallella uses more than 1 core, but does this mean it's harder to program? All I need is something to build a media center and if Raspberry Pi is suitable for that (I've read reports video can be shaky?) and browsing I'll just stick with it. Thanks

Raspberry Pi uses one core.... that is right.

Parallella has 16 cores and 64 cores in the more powerful version.

This means that you can give each core a job to do and it will do it more efficiently compared to the raspberry pi.

From a coding Point of view, all the software has to be written in ARMv7 compared to Raspberry Pi ARMv6. This allows people to run the popular Ubuntu distro of Linux.

Basically, The raspberry pi is your go for a cheap media server and you can find lots of information about it on the forums here. The reports of it being shaking are isolated and it works fine for me.

On another thing: What has happened to this? Has it been released?

[rurwin]

There's updates on the Kickstarter site. They are at an advanced stage of getting the hardware design of the board finalised.

I doubt this will be a plug and play media system. In a few months then maybe the software support will be there, but from the start, it wont. It's also four times the cost of the RaspPi.

[bounce]

Very informative. Thanks.

[Gert van Loo]

I am most skeptical about the software. You can't "just" take video encode/decode and distribute it over 64 cores.
The BCM2835 has many hundreds of years in SW in it. They will need a company with 60 top engineers running for eight years to get near what we have.

[wallarug]

The BCM2835 has many hundreds of years in SW in it.

Computers have only been around for 35 years (max). (I know what you mean though )

[rurwin]

35 years? More like 64.

@gert, If it hits the FOSS imagination it could get far more than 60 people working on it. I doubt they'll be very interested in building a videocore of any great strength, but they might. On the other hand, 16 general-purpose cores are unlikely to compete with RaspPi.

[dmytty]

I am most skeptical about the software. You can't "just" take video encode/decode and distribute it over 64 cores.
The BCM2835 has many hundreds of years in SW in it. They will need a company with 60 top engineers running for eight years to get near what we have.

Gert made the assertion that RPi is 1/4 the cost of the Parallella, but 'cheap' is relative. As Gert correctly points out, there are man years of development in the SoC. However, that actually makes the case for open source.

As a developer, would I choose a platform that allows me to freely tap into the man years of development (Parallella) and modify it to my needs, or choose a platform that requires me to duplicate the man years of development (RPi)? What's quicker and cheaper to prototype, develop, and sell in volumes of 1000's? In short while, Paralella at $100 will bring with it a lot of open source code and knowledge.

FUD (Fear, Uncertainty, Doubt) is the usual way the established interests attack new ideas. Broadcom is certainly an established company, and RPi is just a cheap board - and making cheaper hardware is hardly a new idea.

No offense Gert, but your attitude towards open source is very predictable and 'corporate'. RPi gets no Open CL because it 'takes too long' and Broadcomm won't get it's money back. RPi gets just 1 camera module for the same reasons. As the community has tried to push RPi in new directions, Broadcomm has responded with the money argument.

The solution seems easy: Broadcomm ought to just bump the RPi cost by $10 and open source the software!

If Parallella gets it's business going (seems like it now), in a few years it - or a similar FOSS platform - could very well be the platform of just about every academic, hobbyist, and small commercial project out there.

Will RPi just be a cheap XBMC box (running older versions) because they stay closed source?

[mahjongg]

The solution seems easy: Broadcomm ought to just bump the RPi cost by $10 and open source the software!

You are kidding aren't you? When talking about naive and impractical ideas, this is a prime example!
Bumping the cost of the PI by $10 will kill off the PI fast, also the software isn't theirs, so they cannot give it away anyway. It would also cost a thousand times more than what you are suggesting if they had it to give away.

By the way, I really hope that Parallella will have as much of an impact on the market as the PI has been, it deserves at least that, for to their thinking out of the box.

[gritz]

I am most skeptical about the software. You can't "just" take video encode/decode and distribute it over 64 cores.
The BCM2835 has many hundreds of years in SW in it. They will need a company with 60 top engineers running for eight years to get near what we have.

Gert made the assertion that RPi is 1/4 the cost of the Parallella, but 'cheap' is relative. As Gert correctly points out, there are man years of development in the SoC. However, that actually makes the case for open source.

As a developer, would I choose a platform that allows me to freely tap into the man years of development (Parallella) and modify it to my needs, or choose a platform that requires me to duplicate the man years of development (RPi)? What's quicker and cheaper to prototype, develop, and sell in volumes of 1000's? In short while, Paralella at $100 will bring with it a lot of open source code and knowledge.

FUD (Fear, Uncertainty, Doubt) is the usual way the established interests attack new ideas. Broadcom is certainly an established company, and RPi is just a cheap board - and making cheaper hardware is hardly a new idea.

No offense Gert, but your attitude towards open source is very predictable and 'corporate'. RPi gets no Open CL because it 'takes too long' and Broadcomm won't get it's money back. RPi gets just 1 camera module for the same reasons. As the community has tried to push RPi in new directions, Broadcomm has responded with the money argument.

The solution seems easy: Broadcomm ought to just bump the RPi cost by $10 and open source the software!

If Parallella gets it's business going (seems like it now), in a few years it - or a similar FOSS platform - could very well be the platform of just about every academic, hobbyist, and small commercial project out there.

In an ideal world, perhaps. But just because something is open source it doesn't naturally follow that anyone will actually be able to understand and exploit it. It's also possible that "handing something over to the community" means that responsibility for bug fixing and implementation of core features passes from the manufacturer / developer - which is where ownership of these tasks belongs - to the customer. Orphaned product, anyone?

It depends on whether one has the skill and the free time to dedicate to a project - it's folly to expect that the other guys will pick up the slack. If anyone is considering buying into any work-in-progress (open or closed) with the express intention of having it do useful work (rather than as an amusing diversion, or a calculated risk) then they would do well to buy it for what they can make it do now, not what it might do at some point in the future. Community development has it's place (e.g. beta testing in the widest variety of scenarios, getting the right balance re, features / bloat, making a U.I. ergonomic and intuitive), but I'm not sure that working on core functionality is the job of the community, unless e.g.the piece of hardware in question is very cheap.

Then there's the raft of issues created when something is Designed By Committee.

[rurwin]

Regarding VideoCore, it is never going to be open-sourced. There are multiple reasons for that, including that, unless developers were trained by Broadcom, they couldn't program it. Not every project will work as open-source.

However open-source does not mean community led. There are many open-source projects out there that are led by companies who maintain overall control.

There seems to be a mind-set in the "corporate" world that open-source is wishy-washy and closed source is perfect. That's far from the truth. Consider the case when you find a bug in a third-party library. In my experience it is slightly easier to interest an open-source project to fix it. But what happens if nobody wants to? Maybe it's an old version of the library, maybe fixing it would break stuff that the developers think is more valuable. Maybe yours is a silly edge-case affecting only you. With open or closed source that happens far more often than anyone who has never found such a bug might believe. In fact in both cases it's more the rule than the exception. With closed-source, that's it; find a work-around, or port all your work to another version or another project. With open-source you've got the option of fixing it yourself, or paying someone to fix it for you. That is hugely valuable. That is man-years of cost that can be avoided.

Even if you don't actually go to the trouble of fixing a bug, just understanding it is highly valuable. One simple example that I only managed to characterise perfectly because I knew libc libraries are all very similar:

Be careful when using the isxxx() libc functions (isdigit, isalpha, isprint etc.). In a debug build these will assert if their parameter is not in the range 0-127, maybe 0-126. The assertion reads:

Debug Assertion Failed!
Program: <whatever>
Line:68

Expression: (unsigned)(c+1)<=256
<Abort> <Retry> <Ignore>

None of the buttons do anything useful. Whatever it says, and this may very well be a Microsoft bug, it causes problems when the top (0x80) bit is set in tested characters.

The BSD libc implementation uses a table look-up for all these functions. If the character is outside the range of the table then the function returns false, with a comment that an assertion may be considered. It seems Microsoft have decided it was a good idea, at least in the debug build. To avoid this assertion it is vital that only valid, 7-bit ASCII characters are passed to the isxxx() functions.

[dmytty]

The solution seems easy: Broadcomm ought to just bump the RPi cost by $10 and open source the software!

You are kidding aren't you? When talking about naive and impractical ideas, this is a prime example!
Bumping the cost of the PI by $10 will kill off the PI fast, also the software isn't theirs, so they cannot give it away anyway. It would also cost a thousand times more than what you are suggesting if they had it to give away.

By the way, I really hope that Parallella will have as much of an impact on the market as the PI has been, it deserves at least that, for to their thinking out of the box.

Not kidding. Here's my math...

According to Techcrunch about a million RPi boards were sold by end of 2012
http://techcrunch.com/2013/01/12/one-mi ... ce-launch/

Per Gert's numbers... 60 engineers x 8 years = 480 man years
Figure $50k/man year (hey, lotsa this stuff is now done in the non-Western countries)
480 x $50m = $24 million

So...2.4 million RPi's at the suggested $10 software cost (ie $35-45 total including the software 'fee' for open sourcing) would have covered the cost of all the Videocore and SoC development that Gert cites.

So...at this point RPi would already be about halfway to paying off all the embedded costs...and that's assuming RPi didn't sell even more boards due to being an entirely open source system.

This development cost also assumes that there would be zero open source contribution, that is Broadcom would simply shoulder all the development costs and give the code for the $10 premium.

And here's the 'soft profit' part of the argument. Think of all the universities that would basically adopt RPi as their educational platform. There would be thousands of engineers graduating every year with loads of knowledge about Broadcom and RPi and they would be hired into startups to create new boards, apps, etc all based on the Broadcom SoC! All those startups looking for a great embedded board would prototype and develop and then order thousands of RPi's for their launch. In other words, RPi could be a viable board for many commercial projects. The increased sales volumes could increase Broadcom profits + bring down the RPi 'software' cost.

What's naive and impractical about that?

Edit: I think the real reasoning for this is the elbows that all the bigwigs like to rub. Parallella does not have any big OEM customers [yet], so they can freely stomp on toes at the cocktail parties where all these board level decisions are made. Broadcom customers would likely be pissed off if some college dorm startup had an embedded board + software that equaled or surpassed their own.

Of course, those same college dorm startups are what gives this industry it's life...not the MBA's.