ARM - kickstarter project

firejuggler · 2012-10-25, 20:06

It came to my attention only today
http://www.kickstarter.com/projects/adapteva/parallella-a-supercomputer-for-everyone

jasonp · 2012-10-26, 13:01

It's a very nice idea, and has a lot of promise if they scale up the core processing element. At least they publish the architecture manual for the chip.

firejuggler · 2012-10-26, 20:15

adding a pdf on matrix multiplication on the multiprocessor.
http://www.adapteva.com/wp-content/u...g-Matrices.pdf

firejuggler · 2012-10-27, 01:03

Project is funded... and still 21 hours to go.
Meaning that, for a mere hundred buck you will havea nice chip in May 2013.

fivemack · 2012-10-27, 08:06

It depends what you mean by nice: for $100 you're getting a hard-to-use devboard which would use as much electricity as a Haswell to do a lot less computing than a Haswell in a model infinitely more inconvenient to program than a GPU.

(I read the white paper and it crystallised my concerns about the project; equations used for decoration rather than illumination, uncritical Moore's-Law boosterism, and comparing a tiled decomposition on tiled hardware to a test-of-TLB-bandwidth-disguished-as-an-SGEMM on fairly ancient x86)

LaurV · 2012-10-27, 13:33

I also read the paper and I am a bit on the bitchy side here. Working in an electronics factory those papers seem to me to be more advertizing than realistic, so I would vote with fivemack here. In an year or two, I would pay 7-8 times for a Tesla (k20?) that would "waste" 10 times more electricity and do 20-30-50 times more calculus, not talking about easy to program and ready-made software (like CudaLucas)... And as some guys who know better said on this forum, AMD is not sleeping during this time too... Other companies like Tilera for example, tried these tricks too, but their performance is lousy. It is not "piece of cake" to put all those cores to work, they need pipes, I mean PIPES, to the memory and surroundings, and lots of resources, to rival modern CPUs and GPUs. Putting millions of cores on a chip means nothing, a core is just few fast registers and ALU, not very useful or efficient by itself, unless filled with lots of registers, at least, or fast access to a storage facility (memory). I know it is possible to exit any labyrinth carrying only two stones with you, and running a lot and very fast, but it would be much faster walking slowly, with a pocket full of stones, or even a map. Till they won't solve memory access, playing with the map would be difficult, unless each core would have space enough to store its own map.

They may be efficient for some pure-parallel stuff, systolic calculus, whatever (we had a semester of systolic calculus, but I hardly can recall any algorithms except simple things), but I doubt they will be more efficient then a modern GPU (modern at THAT time), for the same price and consumption.

firejuggler · 2012-10-27, 13:45

I wasn't promoting it. I was merelly pointing the fact it existed. Your decision.
Thanks for the attention and the time you put into it.

bsquared · 2012-10-27, 14:17

I'm skeptical too, but I supported it anyway. Worst case and all of the occupants of my unused hardware bin down in my basement get another friend and I get a mildly interesting story to tell. Worst worst case I see nothing at all. Either way I'm only out $100 that I probably would have lost in the stock market anyway. And just possibly it will be really cool.

ldesnogu · 2012-11-01, 00:08

Might be fun especially if the FPGA tools are available

2012-10-25, 20:06	#1
firejuggler Apr 2010 Over the rainbow A28₁₆ Posts	ARM - kickstarter project It came to my attention only today http://www.kickstarter.com/projects/adapteva/parallella-a-supercomputer-for-everyone

2012-10-27, 08:06	#5
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 7²·131 Posts	It depends what you mean by nice: for $100 you're getting a hard-to-use devboard which would use as much electricity as a Haswell to do a lot less computing than a Haswell in a model infinitely more inconvenient to program than a GPU. (I read the white paper and it crystallised my concerns about the project; equations used for decoration rather than illumination, uncritical Moore's-Law boosterism, and comparing a tiled decomposition on tiled hardware to a test-of-TLB-bandwidth-disguished-as-an-SGEMM on fairly ancient x86) Last fiddled with by fivemack on 2012-10-27 at 08:06

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Neutron Kickstarter	Primeinator	Hardware	2	2014-10-20 08:19
Special project #3b - Project 400	schickel	Aliquot Sequences	307	2011-10-28 01:29
Special project #3a - Project 300	schickel	Aliquot Sequences	29	2011-08-12 17:45
Possible new project for RPS	robert44444uk	Riesel Prime Search	1	2010-04-30 22:01
psp-project.de down	opyrt	Prime Sierpinski Project	6	2010-04-20 10:51

2012-10-26, 13:01	#2
jasonp Tribal Bullet Oct 2004 3541₁₀ Posts	It's a very nice idea, and has a lot of promise if they scale up the core processing element. At least they publish the architecture manual for the chip.

2012-10-26, 20:15	#3
firejuggler Apr 2010 Over the rainbow 2³×5²×13 Posts	adding a pdf on matrix multiplication on the multiprocessor. http://www.adapteva.com/wp-content/u...g-Matrices.pdf

2012-10-27, 01:03	#4
firejuggler Apr 2010 Over the rainbow 2³×5²×13 Posts	Project is funded... and still 21 hours to go. Meaning that, for a mere hundred buck you will havea nice chip in May 2013.

2012-10-27, 13:33	#6
LaurV Romulan Interpreter Jun 2011 Thailand 7×1,373 Posts	I also read the paper and I am a bit on the bitchy side here. Working in an electronics factory those papers seem to me to be more advertizing than realistic, so I would vote with fivemack here. In an year or two, I would pay 7-8 times for a Tesla (k20?) that would "waste" 10 times more electricity and do 20-30-50 times more calculus, not talking about easy to program and ready-made software (like CudaLucas)... And as some guys who know better said on this forum, AMD is not sleeping during this time too... Other companies like Tilera for example, tried these tricks too, but their performance is lousy. It is not "piece of cake" to put all those cores to work, they need pipes, I mean PIPES, to the memory and surroundings, and lots of resources, to rival modern CPUs and GPUs. Putting millions of cores on a chip means nothing, a core is just few fast registers and ALU, not very useful or efficient by itself, unless filled with lots of registers, at least, or fast access to a storage facility (memory). I know it is possible to exit any labyrinth carrying only two stones with you, and running a lot and very fast, but it would be much faster walking slowly, with a pocket full of stones, or even a map. Till they won't solve memory access, playing with the map would be difficult, unless each core would have space enough to store its own map. They may be efficient for some pure-parallel stuff, systolic calculus, whatever (we had a semester of systolic calculus, but I hardly can recall any algorithms except simple things), but I doubt they will be more efficient then a modern GPU (modern at THAT time), for the same price and consumption.

2012-10-27, 13:45	#7
firejuggler Apr 2010 Over the rainbow 2600₁₀ Posts	I wasn't promoting it. I was merelly pointing the fact it existed. Your decision. Thanks for the attention and the time you put into it.

2012-10-27, 14:17	#8
bsquared "Ben" Feb 2007 3·1,171 Posts	I'm skeptical too, but I supported it anyway. Worst case and all of the occupants of my unused hardware bin down in my basement get another friend and I get a mildly interesting story to tell. Worst worst case I see nothing at all. Either way I'm only out $100 that I probably would have lost in the stock market anyway. And just possibly it will be really cool.

2012-11-01, 00:08	#9
ldesnogu Jan 2008 France 226₁₆ Posts	Might be fun especially if the FPGA tools are available