Memory and Core In Sync.

[Hurk]

Is there any advantages for the Memory and Core speeds of the Video Card to be the same. IE:mem/core : 330/330 as opposed to a higher mem clock. IE: 375/330. Is it that big of a difference?

[litghost]

do some testing for you self. i dont think this has been brought up b4. ill do some testing of my own as well.

::UPDATE::i just remebered that the core and the mem use different clock generators, so i dont think there is a way to really "sync" them

[Hurk]

Originally posted by litghost
do some testing for you self. i dont think this has been brought up b4. ill do some testing of my own as well.

My computer is dead. Let me know how your tests turn out.

[Hurk]

Nobody knows?

[q149]

I heard that some FX cards like to be in sync, but i don't think it matters with any other cards.

[Chowdy]

i just keep mine in sync because it looks nice... i'm strange.

My ti4200 came with 250core/222ram. I'd think if there was a major performance boost to having sync'd core and mem,then it'd ship 250/250 or 222/222.

[Infinite Enigma]

i highly doubt it, the cpu sends data as fast as it can to the memory , and the faster it is accepted, or returned, and vice versa, the better. the only in sync advantage is putting the memory at the same speed as the bus that the data is traveling on . even if you got memory at 400mhz on a 333 bus with a mobo, you are likely to perform worse then running the mem at 333 with the bus. this is because the mem not only would have to wait on the bus, or the other way around to send and recieve data each cycle because of the speed difference, but the strobe cycle timings effect it also .

memory flashes at the start of the cycle, processing its load at 400mhz, the bus accepts but only drives it at 333. before the bus has a chance to return and get more , the memory is ready again to flash its 2.0cl cycle, or whatever, but the bus aint there. so it has to wait a whole cycle more, even though the bus is only 66mhz slower, and not an entire 400. so the bus returns just after mem starts a new cycle, but has no data do send, so its waiting, now the bus is waiting. finally, next cycle, the memory can send again.

in sync keeps it all kosher with the bus, but doesnt do crap for the cpu, it is just processing raw data, so the more numbers it can crunch during that cycle , the better, it would just send out less data if you lowered it to the memory speed, resulting in weaker productivity.

[FTC]

Hi,

Let's see : The memory subsystem of a graphics card has to serve two purposes, that is, part of the bandwith goes to the screen refresh circuits and another part goes to the GPU. Given the fact that *usually* a GPU starts being severely degraded when the memory speed is lower than it's own speed, we can conclude that the 'sweet spot' for memory speed is normally around the cpu speed plus the overhead caused by the shared bandwith, which nowadays is low compared to the total available bandwith.

For instance, and supposing you are at 1024x768x32, at 75fps that is 226 mb/second bandwith used by the screen refresh. Now suppose you have a graphics card such as mine, o/c to 300mhz, which gives a bandwith of 9.6 Gb/sec, you can see that the bandwith overhead is of around 2.5%, so my 'theoretical sweet spot' would be to overclock the memory to about 3% more than the core.

Of course this post is full with generalizations... not all cases are the same.

Regards
FTC

[Infinite Enigma]

im always up to learn something new, funny thing is, despite my assumption in my reply , this question actually popped in my head the other day. i have never noticed any real gain though before, but also, wasnt paying attention to the ratio of the 2 speeds like i always did for the cpu and sys ram. infact, i never have really noticed much point gain out of memory at all , mainly cpu, and the mem added a tad. having taken my 5900 from 400/850 to about 435 on the core with no mem increase i noted about 10 points per mhz in 2003. i then bumped the mem up to 900 and over 3 more 5mhz increments and benches in the core, it seemed to remain the same rate of ascent. another 20 mhz, and i gained another 200 points. stock was 5450, mid oc i said was 5800, and now at 455/900, its 6010.

ill have to try the 3% increase on the memory , and put it at a little above 910 and see what happens.

figure it goes without saying, but ill double check , the speed for the ram , is accounted for as being ddr, so if the core is 450, the ram equal is 900 ddr , correct ?

[FTC]

the speed for the ram , is accounted for as being ddr, so if the core is 450, the ram equal is 900 ddr , correct ?


Well, in a sense yes. Most CPUs (and GPUs) are designed so that they can 'eat' (read process) all available bandwith at their own clock rate, but not normally faster so for GPUs designed with DDR in mind this would seem 'natural'. The contrary makes even more sense : Most all GPUs will get penalties if memory clock is slower than core clock... i.e starved instructions because the data is not ready...

**** Just reread your question and notoiced you were really just asking if I meant 600 when talking about 300 (DDR) Mhz... well, yes, of course. ****

Regards
FTC

[Infinite Enigma]

yea, thanks. i know ddr 900 is really clocked at 450 mhz, just moving 2x the data at the start and end cycles. i just wanted to make sure you meant, use the actual mhz of the ram when figuring what to set it aginst the core speed, thanks, gonna try it out.