Barton und Sandra

Den folgenden Beitrag hab ich zur rechten Zeit gefunden - nachdem ich die Resultate meines 2500+ beim Sisoft-Sandra-Cpu-Bench gesehen habe. (sollte folgendes schon xmal oder öfter hier erwähnt worden sein, dann bitte ich um Nachsicht)



Ok, this was from an OcUK forum thread that I responded to and the subject is why the arithmetic was slower on the barton at '2500' speed compared to a XP2600:

My current thinking one this subject is as follows:

The athimetic, pipelines etc within the CPU core will run at the CPU clock rate.

A faster OC'd t'bred athlon's athimetic speed will be higher - right up to it's 256Kb L2 cache, at which point it stalls and has to wait for the memory to supply information.. In essence it's running faster than a Barton within the cache space..

The Barton works on a slightly different principle..

The Barton's athimetic speed is slightly slower - however it can run at that speed for longer due to it's 512Kb L2 cache, at which point it too stalls back to memory speed if has to access memory.


So how come it's faster???

Well the tbred simply can't maintain the breakneck speed for very long - so it's actually sitting idle waiting on memory for a considerable portion of time. So that speed is wasted after a certain point..

A Barton can run at a slower speed but it spends less time idling because of the cache, so overall it performs more work than the tbred. It's slightly slower speed is not wasted.


Hence for numeric computation that fits into 256Kb a tbred will be faster, however with windows and 90% of computations.. the processed size of information is larger than 256Kb.. so that is where the Barton wins - sustained processing.

One extremely good example of this is the OcUK SETI benchmark, not to blow my own trumpet but the two Bartons that have entered SET results are 2nd and 4th..

Setbenchmark results table can be found here:
http://www.orifice.pwp.blueyonder.co.uk/

Hope that helps.. Like the rabbit and the tortoise - raw speed isn't everything!

So just because Sandra says your arithmetic is slower.. in a real number crunch.. whammo! You're laughing with that 512Kb cache.

Btw - this also explains why a Barton runs hotter than a tbred.. a tbred is idle more so it's not going to produce as much heat as the working Barton..

I really wish it had a 4Mb cache like the SGIs..

gepostet am 03-18-2003 10:56 in amdforums.com von NickK