Discussion Forums

$99 ARM supercomputer

14 replies [Last post]
sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732

There is a new Kickstarter project that wants to create a $99 ARM-based raspberry Pi-like supercomputer. Initially it will use a 16-core Adapteva ARM chip. The goal of the Kickstarter campaign is to get enough money so that they can design and create a 64-core version that will have 90 GFLOPs of performance, use 5W of power, and sell for $199. 

For more information see:

The SETI-angle of this inexpensive Parallella supercomputer is three-fold:

  1. Imagine setiQuest hobbyists being able to buy a 64-core model to use for SonATA code/algorithm development? This would solve a couple problems the open source aspect of setiQuest had.
  2. The Dell server donation in the Hat Creek server room was brand new 3 years ago but now they are about 2 generations old which is quite a lot in the world of Moore's Law. Those Dell servers also consume a huge amount of electricity and produce an enormous amount of heat, hence installation of the new air conditioner. Imagine replacing the Dell servers with many of these 64-core Parallella boards?
  3. Kickstarter has forever changed the Internet-based landscape with new expectations for social funding. How can the SETI Institute use Kickstarter to aid its funding efforts? How has Kickstarter changed what people expect to get from their funding? Both physically and from a communications standpoint. The whole process is very open. What did the setistars.org donors get? A new air conditioner which is great but what else?  I don't know. I think more open communication of how the setistars.org money is benefiting the SETI Institute would help improve donors willingness to donate money in the future.

I recently attended an Intel Xeon Phi training class. The Xeon Phi is a 50-core MIC (Many Integrated Core) PCIe card that has about 10X the performance but will also be about 20X the cost and consume about 60X the amount of electricity (heat). I learned that a major problem with MIC is programming the parallel beast. Intel is solving this with their cluster software toolkit.  The development tools for the Parallella board will be an issue. 

Dave Robinson
Dave Robinson's picture
Offline
Joined: 2010-04-29
Posts: 196
Interesting Thought

 Who said

"all this talk about SIMD, GPU's, multi-processor motherboards, and distributed computing is a bit premature."  But it is good to see you coming round to my way of thinkingsmiley

I have been watching with interest a University group who are trying to generate an Ethernet connected array of Raspberry Pi's whose performance promises to be really fast, and at 25GBP per processor seems to be really cost effective. But the architecture you seem to be describing looks even better. I guess if you can integrate one ARM chip onto a small part of an ASIC or even an FPGA then there is no reason why you shouldn't be able to integrate a reasonable number together with a suitable interconnect. I must admit that I like the ARM structure. For a while I had been using a Rabbit processor for my Embedded experiments, but have recently switched to using an ARM based 'stamp' processor running FORTH, the ease of accessing I2C and tons of parallel IO make it a great machine for that. (using FORTH is simply a nostalgic idiosyncrasy of mine).

I have been toying with the idea of getting a Raspberry Pi, but at the price's you quote, if they are doing a good prototyping system, I will think seriously consider going for a multicore ARM. Maybe SETIquest should branch out into designing high performance hardware structures for processing the ATA data.

Regards

Dave Robinson

sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732
I still think algorithm

I still think algorithm development is best done at a higher level. Focusing too early on implementation and optimizations is a mistake IMO.

My primary interest in this 64-core ARM board along with all my recent SIMD, GPU, multi-processor CPU, ... talk is in implementing these technologies into baudline. I thought I'd share this info with setiQuest because it is relevant on several levels: help the ATA data-center cost effectively follow Moore's Law and the changing funding landscape.

Keep in mind that Kickstarter is not a store, this product does not currently exist, and it might never get funded. It all depends on how many people believe in this project and kick-in some money.  The project is backed by a commercial company so it has a good probability of actually being built if the funding levels are reached but $750,000 is a lot of money. I'm not sure they will reach it and I'm not familiar enough with Kickstarter to know what happens if they don't.  Also on a GFLOPs level how does this compare cost-wise to a low end GPU? No one has been able to do any benchmarks since the product doesn't exist yet.  In spite of that, I think this project is a great idea and I wish them luck.

sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732
This article on The Register

This Register article has some more details and new information about the many-core Adapteva-Kickstarter ARM chip:

For 2018 they are planning a lower end 1000 core, 1 TFLOP, 2W chip and a higher end 64000 core, 100 TFLOP, 100W chip. This amount of computing power will be amazing but these chips are going to be a major pain to write software for. OpenCL was mentioned and that may be the best route to go.

The Kickstarter campaign has so far pledged $244,672 of a $750,000 goal with 19 days to go.

Dave Robinson
Dave Robinson's picture
Offline
Joined: 2010-04-29
Posts: 196
DIY - well nearly!!!

Not quite $99  but sounds like a lot of fun

http://www.geek.com/articles/chips/supercomputer-built-with-raspberry-pi-and-lego-20120912/

Regards

Dave Robinson

sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732
That makes for an interesting

That makes for an interesting comparison.  The Lego Raspberry Pi "super computer" uses 64 rPi boards, has Ethernet connectivity, and cost about $4000. The $199 Adapteva 64-core board is planned to cost $199. Both have 64-cores but what I think is key for the Adapteva is that the on-die connectivity will be extremely fast.  A super computer is only as fast as its connectivity ...

Kickerstarter funding update for the Adapteva-Parallela project: 2792 backers, $439565 pledged, 4 days to go. It doesn't look like they are going to hit their $750000 target. ):

sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732
Kickerstarter funding update

Kickerstarter funding update for the Adapteva-Parallela project:

3625 backers, $644655 pledged, 28 hours to go. It looks like they might reach their $750000 goal.

zigzag
zigzag's picture
Offline
Joined: 2012-10-23
Posts: 19
I want one of the 64 core

I want one of the 64 core boards! All those cores would be fun to program. cool It would be an awesome SETI development platform. The SETI Institute should buy a whole bunch of 'em and give them to people for educational purposes! cheeky

sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732
At first I thought it was

At first I thought it was crazy but that's actually a very good idea. Seeding developers with cool hardware is a fantastic way to promote creative development.  Things might be very different around here if some of that TED wish money was used for this sort of endeavor when setiQuest launched. Unfortunately that was several years ago and the SETI Institute doesn't have any money for the setiQuest project. This would be a great donation opportunity. It doesn't hurt to ask. I'll bring it up at the community meeting next week.

zigzag
zigzag's picture
Offline
Joined: 2012-10-23
Posts: 19
No problem. I'm glad to have

No problem. I'm glad to have helped. angel

The Parallela Project got more than their goal, currently at $870,775 with 2 hours to go. Now they need $2.13M before they make the 64-core boards. Not sure when that funding goal ends.

Bummer to hear about the SETI Institute's financial problems. Where did all that TED money go?

sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732
You should really ask that

You should really ask that question over at this thread:

http://setiquest.org/forum/topic/ted-prize-now-1m

The TED wish prize was much smaller than it is now.

zigzag
zigzag's picture
Offline
Joined: 2012-10-23
Posts: 19
OK. I just asked my question

OK. I just asked my question over there.  I don't understand why you wouldn't answer my question here. angry

sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732
C / C++ / OpenCL

I'm a C and pthreads fan myself but OpenCL may be the opportune way to program the 16-core and 64-core chips. A side benefit is the developed OpenCL code can also be run on a multi-core CPU, GPU, or FPGA. I just read this article about OpenCL which mentioned that Altera wants to support OpenCL programming on its FPGA's.

http://www.techdesignforums.com/embedded/technique/opencl-portable-code-multicore/

The article is well worth a read.  I'm a little concerned that Adapteva / Parallella is focusing on the hardware and leaving software as an open source / community effort.  From my perspective, the hardware seems like the easy part.

sigblips
sigblips's picture
Offline
Joined: 2010-04-20
Posts: 732
Just to put the planned $199

To put the planned $199 Adapteva 64-core, 5W, 90 GFLOPs board into perspective, the newest iPad 4 has a GPU capable of 77 GFLOPs:

http://www.anandtech.com/show/6426/ipad-4-gpu-performance-analyzed-power...

I just read that PowerVR GPUs are OpenCL compliant. Hmmm.

zigzag
zigzag's picture
Offline
Joined: 2012-10-23
Posts: 19
An iPad 4. Really?  Whoa !!!

An iPad 4. Really?  Whoa !!! that is a heckuva lot of GFLOPs. cheeky