And the winner is....
The electric company. Seriously, what does it cost to run the top 10 machines for 24 hours?
Looking at the semi-annual Top 500 list of supercomputers, you would never know the world was battling recession. Supercomputer centers keep chucking out old tech and rolling in new tech at the same feverish pace. The fall edition of the list was announced this morning at the SC09 supercomputing trade show in Portland, Oregon …
I recon that's the sort of hardware that we will need for M$ windoze 8 (AKA windoze 7 SP2) for the Lip-smackin', gut-wreching, RSI-inducing, eyestrain-causing, mind-blowin', heartburn-causin', RAM-gobblin', de-motivatin', paris-wankin', transformer-buzzin', high-costin', fast-failin', CPU-fizzin' Pepsi sponsored desktop themes.
Paris, 'cos I mentioned her somewhere in that lot
It was only 1996 or early 1997 that saw the last _single cpu_ super on the top500 list. It also seems that MIPS per MHz has dropped to half what it was then in the upper end (nevermind what has happened elsewhere). Part of the commoditisation drive perhaps.
If your number crunching needs are specialised enough you can already get a front seat for a dime by going the FPGA route, but that obviously won't earn you a top500 list mention. I think we'll see more and more of that reconfigurable computing thing now that GPUs and GPGPUs have become acceptable. DSPs and FPGAs are not too far from that. Going forward.... Software updateable general computing cores, anyone?
In fact, some of that already exist(ed), but hidden under the rug: transmeta.
Are you talking ? At least it's flash on Linux. For your embarrassment, Windows 7 will not allow me to play Minesweeper (pretty intensive computing task isn't it?) because I don't have graphics acceleration. That's why you don't see much MS stuff in that top 500 list.
Mr. Morgan, I disagree in part with your statement:
" Provided supercomputing centers can shift their codes to hybrid architectures, the price/performance of multicore x64 processors and their related GPUs is probably the horse to bet on. Exotic machines may have seen their heydays already."
In my book, CPU/GPU machines ARE still exotic - because of the coding issues involved. While I haven't written any commercial applications, I have spent a fair bit of time learning CUDA, and the demands and limitations of the GPU architecture are probably going to continue to be different enough from a CPU programming model that programmers will continue to have a lot to learn at the very least. The GPU vendors recognize that, and are working wonders towards making the GPUs easier to program - but at the end of the day, it is STILL a SIMD-type programming model, and that's different enough from "normal" programming to be hard for most people - and it only works well on certain types of problems. A hybrid machine running problems that don't SIMD well is hardly faster than it's CPUs.
The one area that CPUs and GPUs both are not exotic is, thankfully, their economics. They both are driven by "consumer"-lead investment models, and in that way they differenentiate themselves from other hybrid solutions such as Cell, FFGAs, vector co-processors, etc. That is why they will very probably win. The only question for the next batch of HPC machines will be the ratio of CPUs versus GPUs, and how exactly they are connected...but even then, I suspect we will have to call them "exotic" to program...at least to do it well.
So, 402 of the top 500 use Intel chips. 42 AMD and 52 IBM Power. That leaves 4 systems -- are they using FPGAs or GPGPU to the extent the CPUs don't do the main work.. do they use SPARC.. some huge ARM cluster.. or what? I'm too lazy to look at the list and find out...
I must say, the box with 850mhz PowerPC 450s, that box must have a *lot* of cores to be nearly as fast as the 2.6ghz AMD 6-core box with almost 99,000 cores in it.