• Quick note - the problem with Youtube videos not embedding on the forum appears to have been fixed, thanks to ZiprHead. If you do still see problems let me know.

Supercomputers!

Cuddles

Penultimate Amazing
Joined
Jul 28, 2006
Messages
18,840
We don't seem to have a general thread for these, but I always find the Top500 list and associated discussions fairly interesting. The latest Top500 can be found here, and there's an article about it here.

Some of the trends are quite interesting. In particular, GPUs/co-processors have been expected to really start taking over. However, after shooting up from nothing over the course of a few years, their use hasn't actually increased at all for the last two years. Also interesting is that 10 years ago Intel was an occasional bit player, with most of the field split between 5 or so big players and a bunch of proprietary purpose-built machines taking the rest. Intel now holds over 4/5 of the CPU market.

Finally, the first graph on the Reg article amuses me. 20 years ago, the combined total computing power of the 500 most powerful computers on the planet was 1.17 Tflops. A mid-range consumer graphics card from the last generation can now get over 1.5 Tflops.
 
I'm pleasantly surprised how well the UK computers come out in this:

The European Centre for Medium Range Weather Forecasting (The Met Office to most people) have an identical double entry at #52 and #51.
HECToR the UK's older Cray machine in Edinburgh at #50.
DIRAC, the newer UK IBM machine also in Edinburgh at #27.
And heading the UK pack at #23 and #19, the two research council machines in Swindon and Daresbury respectively. Good work UKPLC!
 
Well, those of us who do DSP for a living, or did, have noticed the utility of video processors. Harvard architecture systems, even with pipelines, are remarkably useful for many DSP functions, in particular direct convolution. There is something to be said for non-pipelined processors for IIR filters, FFT's, and the like, of course, but given the power in a modern video card...

The first computer I used at Bell Labs was 800kips, and 200kflops. And it had memory maxed out, a whole 32kwords. (not bytes) The start of the coding for cellphones that you hear nowdays every day was done on those machines. When we got the first array processor it was a real kick (5.5mflops).Then we went to 88mflops and 64 mB. Now that change in memory size was a blessing, that machine is partially responsible for the existence of MP3, for instance. (I wrote the first AT&T perceptual codec as a test program from the Alliant FX8 Fortran FX compiler, among other things. It had been fallow for some time due to lack of memory.)

Now? I don't think there's a chip out that slow, to say the least.

Of course the new massively parallel machines are for problems that can be partitioned, like weather, fission, etc. I can't help wonder if we will see some new understanding of star formation and galaxy formation soon from some of these extremely large parallel systems.
 
Last edited:
I have Top500 bookmarked. The Tianhe-2 has been at the top for many months now. I love that it has a petabyte of RAM. Oh, and 3 megacores.

~~ Paul
 
Well, those of us who do DSP for a living, or did, have noticed the utility of video processors. Harvard architecture systems, even with pipelines, are remarkably useful for many DSP functions, in particular direct convolution. There is something to be said for non-pipelined processors for IIR filters, FFT's, and the like, of course, but given the power in a modern video card...

Although right on cue there's some interesting news from Intel. They've already taken a reasonable chunk out of Nvidia's market share in the co-processor end of things, so they could well be on to something. Still much more parallel than traditional CPUs, but at the same time very different from the GPU based approach.
 
Finally, the first graph on the Reg article amuses me. 20 years ago, the combined total computing power of the 500 most powerful computers on the planet was 1.17 Tflops. A mid-range consumer graphics card from the last generation can now get over 1.5 Tflops.
Thanks, now my wife are having an argument about this. I think that's a hilariously awesome stat.

My wife, who spent the past year building out 3-D rendering machines, says "mid-range only by grouping several together; high-end maybe a single card if you have the right board and slot for it".

Can you resolve our dispute with some makes/models that you have in mind?
 
The web pages give interesting summary information, but there's much more detail in the spreadsheet. One thing I notice is how many--like 97% --are running various flavours of Linux: CentOS, SuSE, Red Hat Enterprise, Scientific Linux, and many just "Linux." There are 10 running AIX, which is Unix-like OS, and one runs SuperUX, which is BSD based and thus another Unix-like OS.

Only two are running Windows. I wonder much the Windows licenses cost for 30,000 cores? :D
 
Thanks, now my wife are having an argument about this. I think that's a hilariously awesome stat.

My wife, who spent the past year building out 3-D rendering machines, says "mid-range only by grouping several together; high-end maybe a single card if you have the right board and slot for it".

Can you resolve our dispute with some makes/models that you have in mind?

http://en.wikipedia.org/wiki/GeForce_600_Series#GeForce_600_.286xx.29_series

GTX650 Ti and up. The 600 series is last year's generation and the 650 is far from top of the range, the 800 series will be out soon. In addition, note that the PS4 uses a single Radeon GPU that is stated to get 1.84 tflops and the Xbone has a similar GPU stated to be 1.31 tflops. Neither of them have particularly impressive raw specs compared to many PC GPUs. For example, the Xbone GPU has 768 cores running at 853MHz, while the GTX650 Ti has the same number of cores running anywhere from 900MHz to over 1GHz. Obviously there are more details involved than just that, and actually making use of all the raw power is easier said than done (and is one area where consoles have a distinct advantage). But the raw power is certainly there.

Which, conveniently enough, actually links back to the main thread topic with this article about computational efficiency and how it should actually be measured.
 
We don't seem to have a general thread for these, but I always find the Top500 list and associated discussions fairly interesting. The latest Top500 can be found here, and there's an article about it here.

Some of the trends are quite interesting. In particular, GPUs/co-processors have been expected to really start taking over. However, after shooting up from nothing over the course of a few years, their use hasn't actually increased at all for the last two years. Also interesting is that 10 years ago Intel was an occasional bit player, with most of the field split between 5 or so big players and a bunch of proprietary purpose-built machines taking the rest. Intel now holds over 4/5 of the CPU market.
Finally, the first graph on the Reg article amuses me. 20 years ago, the combined total computing power of the 500 most powerful computers on the planet was 1.17 Tflops. A mid-range consumer graphics card from the last generation can now get over 1.5 Tflops.
The trend from highly specialised vector systems (like the Crap Y-MP I used a bit back in the early nineties) to the highly parallel systems using COTS components today. Brute force rather than optimisation in many ways.

That said I think my current tablet is more powerful than that Cray.
 

Nice to meet you! I was a VMS guy for Fermilab long ago. (Also RSX-11 and RSX-15)
 
You know, I would never have bet on *nix as being the future of computing. VMS seemed more likely.

I worked in the VMS development group at Digital for three years. While VMS was clearly better than Unix (:D), it was also clear that an OS that uses less system resources has certain advantages. It was also clear that VMS was stuck on VAXen, while Unix was free to wander the hills.

~~ Paul
 
I'll soon be shutting down my VMS cluster that has given many years of faithful service. Does anyone recall where the off switch is?
 

Back
Top Bottom