1000-Core Computer

[Rodrigo]

http://www.pcworld.com/article/21524...l_wbx_h_crawl1

Suppose that such a machine became commercially available at a price within reach of individual buyers. What might the implications be for GIMPS? Or is there less than meets the eye, here?

Rodrigo

[Batalov]

For certain tasks, FPGAs are very effective and had been used for years, certainly more than a decade in more than prototypes.
For example, we've been using a sequence alignment server by TimeLogic since 1999 till... that company went bankrupt and was bought for cents (and in this capacity, it still exists; their ancient rival Paracel was first bought by Celera and then buried. Paracel used ASICs).
Even as both were doing fairly well, I've seen literally dozens of posters at bioinformatics conferences, every year, with eager students reporting how well the sequence alignments could be implemented in FPGAs... (granted, wikipedia or google searches were so much harder back then ).

There must be some fairly exciting implementation details in that researchers' work, but "Scottish Researchers Claim 1000-Core Processor" is not a good title. The MORA and Gannet are interesting things but there's nothing about them in the article.

[Rodrigo]

Fascinating -- thanks for the additional information! I didn't know that this had already been done before.

Rodrigo

[jasonp]

It's *all* been done before; you just have to find it, and/or make it cost-effective.

[Rodrigo]

Originally Posted by jasonp

It's *all* been done before; you just have to find it, and/or make it cost-effective.

OK, so the question then becomes -- assuming that it were even possible to make a processor like this affordable to individual users, would it deliver as much processing power as the "1000 cores" description suggests? Or are we basically talking about an otherwise ordinary modern processor that's been divided into 1000 little pieces, each delivering a tiny fraction of the available total processing power?

I guess what I'm really curious about is whether GIMPS would hypothetically be able to make use of something like this (with the appropriate programming). Imagine what the Prime95 worker window screen might look like...

Rodrigo

[bsquared]

Originally Posted by Rodrigo

OK, so the question then becomes -- assuming that it were even possible to make a processor like this affordable to individual users, would it deliver as much processing power as the "1000 cores" description suggests? Or are we basically talking about an otherwise ordinary modern processor that's been divided into 1000 little pieces, each delivering a tiny fraction of the available total processing power?

You'll note that the 1000 core FPGA did something only 20 times faster than a modern processor. So cores != performance benefit, and in general the answers to your questions are: it depends and no.

For very specific jobs, FPGA's can be configured to do that job (and only that job) very well. So you can get many more times the performance of a modern processor with something that in no way resembles the internals of a modern processor (other than the basics along the lines of "it has I/O and logic functions").

The performance "multiplier" depends on many things: the degree to which I/O bandwidth is a bottleneck, how parallelizable the job is, and the clock rate achievable with a given design, to name a few.

Originally Posted by Rodrigo

I guess what I'm really curious about is whether GIMPS would hypothetically be able to make use of something like this (with the appropriate programming). Imagine what the Prime95 worker window screen might look like...

Rodrigo

Sure, I suppose it could. Is the result worth the investment and effort? That's another question.

[xilman]

Originally Posted by bsquared

For very specific jobs, FPGA's can be configured to do that job (and only that job) very well. So you can get many more times the performance of a modern processor with something that in no way resembles the internals of a modern processor (other than the basics along the lines of "it has I/O and logic functions").

Indeed. I know of a system built out of FPGAs. It has about 200 boards, each of which contain 16 FPGAs, each of which are about 500 times the speed of a half-way decent quad-core cpu. Do the multiplication for yourself.

What I haven't told you yet is that that system can do only one thing, and that one thing is so specialized that virtually no-one wants to do it at all, let alone quickly.


Paul

[bsquared]

Originally Posted by xilman

Indeed. I know of a system built out of FPGAs. It has about 200 boards, each of which contain 16 FPGAs, each of which are about 500 times the speed of a half-way decent quad-core cpu. Do the multiplication for yourself.

What I haven't told you yet is that that system can do only one thing, and that one thing is so specialized that virtually no-one wants to do it at all, let alone quickly.


Paul

One-upping you would be pointless, but I have similar stories

[Rodrigo]

Thanks bquared, your analysis jibes with what I suspected but didn't have the technical expertise to believe with any intensity.

Rodrigo

[xilman]

Originally Posted by bsquared

One-upping you would be pointless, but I have similar stories

I'm sure you do. Please feel free to one-up me if you wish. Public willy-waving is often amusing to by-standers. However, I will not post any more details about the system I've just described, other than it is still in existence, still computing and that I have played a significant role in its construction and use.

Paul

[Rodrigo]

Originally Posted by xilman

I'm sure you do. Please feel free to one-up me if you wish. Public willy-waving is often amusing to by-standers. However, I will not post any more details about the system I've just described, other than it is still in existence, still computing and that I have played a significant role in its construction and use.

Paul

Just want to say that as a bystander, I wouldn't view such comparisons as public willy-waving, but as a chance to hear about some of these systems working in practice.

R