Quad core vs Cray supercomputer from the '90s

[JEDI]

Have we surpassed the old supercomputers yet?

 

[nerp]

You basically just have to dig up an old newspaper article or two from the time and confirm what the MFLOPs of whatever Cray was being written about. Then do the math based on a processor from today. I'd love to spit out the numbers but I don't have reference materials handy.

 

[sutahz]

My question is do you already have numbers or some stats to show one way or the other?
Alas, my vote is for yes.

 

[PolymerTim]

I found a pretty helpful link for the discussion on the website that tracks the top 500 most powerful supercomputers.

http://www.top500.org/lists/20...erformance_development

Luckily, right on their front page was a graph showing the history of the top 500. Looks like around the early to mid 90's the best supercomputer ranged around 100-200 GFlops.

Now maybe you guys can fill in the rest of the easy info that I don't know: what do typical quad cores do today?

 

[Foxery]

Keep in mind that supercomputers on this list "cheat" by having 1000+ CPUs. Free, public distributed computing projects like Folding@Home and SETI@Home make both their power and their budgets look stupid

The quick answer is a 3GHz processor means 3 GFLOPs, x4 cores = 12 GFLOPs in your case at home. This assumes you're doing simple math problems and there are probably a hundred reasons why it's a horrible estimate, but you get the general idea. Today's desktop machines would probably show up on the list from 1990. 18 years of new technology is no joke...

 

[bryanW1995]

hmmmm, so you're saying that 4x 3ghz 04 is the same as one 3ghz Q9650 today? I think not...

 

[aigomorla]

Originally posted by: bryanW1995
hmmmm, so you're saying that 4x 3ghz 04 is the same as one 3ghz Q9650 today? I think not...

lol.... i was going to stretch your statemnt more and pull up smithfields.

3ghz smithfield = abacass compared to a 3ghz c2d. Well maybe not that extreme.

 

[taltamir]

Ah, but they make video cards WITHOUT the ports meant to be used for scientific calculations... (thanks to the programmable shader language allowing it to run generic code)...

And video cards have surpassed teraflop already. (not all, but the 3870x2 breaks teraflop... and well as the 4 GPU tesla from nvidia, which is designed specifically for calculations)

http://rajanand.wordpress.com/...la-teraflop-computing/

 

[exar333]

Yeah, but how many FPS would a 1990's Cray get in Crysis? Thats the question that most readers are wondering!

 

[Lord Banshee]

Originally posted by: Foxery
Keep in mind that supercomputers on this list "cheat" by having 1000+ CPUs. Free, public distributed computing projects like Folding@Home and SETI@Home make both their power and their budgets look stupid

The quick answer is a 3GHz processor means 3 GFLOPs, x4 cores = 12 GFLOPs in your case at home. This assumes you're doing simple math problems and there are probably a hundred reasons why it's a horrible estimate, but you get the general idea. Today's desktop machines would probably show up on the list from 1990. 18 years of new technology is no joke...

you forgot that these newer processor has multiple FP units, if instructions are aligned correctly. The core2duo has 4 FP units i think so it would be 48GFLOPs. But GFLOPS is really a very very bad way to determine performance. Real code does not align FP instructions 100% of the time, and not all code are floating point.

(i think)

 

[Foxery]

I said it was a bad estimate In reality, modern architectures involve pipelines and code caching and predictive branching, and SSE instructions, but some of this is offset by doing more complicated calculations than we might have tried 18 years ago.

The OP was very unspecific about how to compare them. Then again, this is AT... you're going to get specifics anyway!

 

[Griswold]

Originally posted by: taltamir
Ah, but they make video cards WITHOUT the ports meant to be used for scientific calculations... (thanks to the programmable shader language allowing it to run generic code)...

And video cards have surpassed teraflop already. (not all, but the 3870x2 breaks teraflop... and well as the 4 GPU tesla from nvidia, which is designed specifically for calculations)

http://rajanand.wordpress.com/...la-teraflop-computing/

Thats single precision. The figure will drop by ~50% when you apply double precision - and that is what most scientific workloads require.

 

[Rebel44]

Cray T90 which was released in 1995 has Max. 57.6Gflps and QX9650 @3Ghz has ~50Gflops so if you OC this CPU it would easily win.