Page 1 of 1

64 cores in an 8x8 array for US$200

Posted: Sat Sep 29, 2012 8:01 am
by sje

Re: 64 cores in an 8x8 array for US$200

Posted: Sat Sep 29, 2012 7:21 pm
by diep
Very interesting link.

Very interesting such cpu's especially for your mobile phone.

What they promise basically seems to be 64 cpu's that are each 800Mhz RISC. They don't offer an instruction set yet.

It's not clear to me whether the entire grid executes the same code at the same time, or whether each individual RISC cpu has its own instruction stream that can run at its own pace.

That's a pretty important question.

I nowhere see mention of 'local instruction caches'.

If it doesn't have that, it still is a very interesting processor, as i know at least 1 workload that would be mighty interesting to run on it.

Yet then the next question is of course, how realistic any of the price projections are. I don't believe them.

They need to produce dozens millions of them to get to that price...

Note that for those who aren't aware of absolute performance - this is a 2 watt cpu - don't compare that with intel i7's.

Something of 2 watt will of course never outperform your latest i7 that eats a 100+ watt (just cpu under full load).

Compare it with ARM and MIPS etc....

p.s. In the end these cpu's always flop when they aren't cheap enough - they always in the end want to sell them for too much!

Clustering 64 of these cpu's would be interesting. Not for $200 a chip of course. Maybe $20 a cpu and total soc costs for each soc in such setup not above entire soc + cpu for above $40.

Maybe then this is interesting - otherwise it isn't.

$200 is too much for this. Even in most optimistic scenario this won't outperform the 2 socket L5420's i have here.

Would be interesting to trial factor on them by the way. Right now works pretty efficient in CUDA for GPU's, yet the candidate generation is not so efficient in such case. Losing factor 2 there redundancy (composites) or so and has to be done on cpu's.

All can be done on this and with 1MB local RAM each cpu that would go fast.

Re: 64 cores in an 8x8 array for US$200

Posted: Sat Sep 29, 2012 7:32 pm
by Dan Andersson
Greenarrays has the GA144 and ZiiLABS the ZMS 5 to 40 already out with models with more cores also the Parallax Propeller offer an eight core chip.

MvH Dan Andersson

Re: 64 cores in an 8x8 array for US$200

Posted: Sat Sep 29, 2012 7:38 pm
by diep
Dan Andersson wrote:Greenarrays has the GA144 and ZiiLABS the ZMS 5 to 40 already out with models with more cores also the Parallax Propeller offer an eight core chip.

MvH Dan Andersson
You refer to something total different. You refer to seomthing that has 128 words for each 'core'.

So that's not even in the same league like this :)

They promise here 64 RISC cores with fully functional RISC instructionset and a C compiler whereas Greenarrays promises something interesting maybe for microcontrollers?

Re: 64 cores in an 8x8 array for US$200

Posted: Sat Sep 29, 2012 11:16 pm
by Dan Andersson
The Grenarrays chip certainly isn't anything like what you are implying. The Parallax neither but it lacks a bit omph.

MvH Dan Andersson

Re: 64 cores in an 8x8 array for US$200

Posted: Sun Sep 30, 2012 4:24 am
by jshriver
Seems fascinating. For $200 I'd gladly get one. Would be interesting if this would be useful in the virtualization world. Can imagine with such low power, this would be great for VPS providers, if memory could be expanded.

-Josh

Re: 64 cores in an 8x8 array for US$200

Posted: Sun Sep 30, 2012 12:14 pm
by Dan Andersson
The RISC cores are FP oriented throwing away any "complicated" integer instructions like multiplication and division ;) Looking it over it becomes obvious they are aimed at Perceptron, SVM, NN applications for recognition and classification tasks like voice, face, writing, surveillance ...

MvH Dan Andersson

Re: 64 cores in an 8x8 array for US$200

Posted: Sun Sep 30, 2012 12:23 pm
by diep
Dan Andersson wrote:The RISC cores are FP oriented throwing away any "complicated" integer instructions like multiplication and division ;) Looking it over it becomes obvious they are aimed at Perceptron, SVM, NN applications for recognition and classification tasks like voice, face, writing, surveillance ...

MvH Dan Andersson
For floating point, multiplication is crucial.

ANN's as well.