agent_x007
New member
Hi guys.
I didn't see this topic on site yet, and I appologise if I created second one.
But This Is Important.
Thanks to marketing and engineering team not talking enough - NV got itself into... a peculiar situation.
Few days ago NV confirmed GTX 970 had from the start one of L2 cache chunks (ie. 256kB one) switched off, and (in doing so) it means that 1/8th of total ROP's from GPU is also disabled (there are 56 ROP's in total, NOT 64).
BUT thanks to Maxwell trick, card still can adress the VRAM that would usually be lost due to missing ROP units.
In short : GTX 970 IS still a 4GB VRAM graphics card.
Downside to "trick" ?
Last 512MB of VRAM is slower than the rest (it's got around 1/7th of the "normal bandwidth" - to be more exact).
That's why NV "opted" for 3,5GB + 0,5GB configuration in GTX 970.
Of course, that last 0,5GB is accessed only in 4k + "Ultra" with Anti-Aliasing (ie. when VRAM usage get's over 3,5GB ), so performance drop isn't that bad (comparing GTX 980 4k vs. GTX 970 4k), but still... it can cause problems if OS dosen't know that last bit of memory is slower (and keeps sending important data there).
Maybe Tom could do a Video about this with some thoughts on GTX 970 in general ?
I based all this on pcper.com article (will link it here) : http://www.pcper.com/reviews/Graphi...Full-Memory-Structure-and-Limitations-GTX-970
Other articles about this issue :
Anandtech.com : http://anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation
Hardware.fr (translated to ENG from FR) : https://translate.googleusercontent...t.html&usg=ALkJrhhioO-0LvIVwMt02vN9TS-PEnMARQ
Thank U.
PS. One last thing : "52 ROP's number"
Well, SMM's in GTX 970 can do 52 Pixels/cycle, so that's the number I keep seeing poping out.
But with that logic, GTX 780 would have 32 or 40 ROP's (depending on which SMX's were disabled) and GTX 780 Ti would get 40.
ROP's can do for example AA without SMM's help.
+ Nvidia isn't stupid : They wouldn't put something in GPU if they couldn't use it - lol.
I didn't see this topic on site yet, and I appologise if I created second one.
But This Is Important.
Thanks to marketing and engineering team not talking enough - NV got itself into... a peculiar situation.
Few days ago NV confirmed GTX 970 had from the start one of L2 cache chunks (ie. 256kB one) switched off, and (in doing so) it means that 1/8th of total ROP's from GPU is also disabled (there are 56 ROP's in total, NOT 64).
BUT thanks to Maxwell trick, card still can adress the VRAM that would usually be lost due to missing ROP units.
In short : GTX 970 IS still a 4GB VRAM graphics card.
Downside to "trick" ?
Last 512MB of VRAM is slower than the rest (it's got around 1/7th of the "normal bandwidth" - to be more exact).
That's why NV "opted" for 3,5GB + 0,5GB configuration in GTX 970.
Of course, that last 0,5GB is accessed only in 4k + "Ultra" with Anti-Aliasing (ie. when VRAM usage get's over 3,5GB ), so performance drop isn't that bad (comparing GTX 980 4k vs. GTX 970 4k), but still... it can cause problems if OS dosen't know that last bit of memory is slower (and keeps sending important data there).
Maybe Tom could do a Video about this with some thoughts on GTX 970 in general ?
I based all this on pcper.com article (will link it here) : http://www.pcper.com/reviews/Graphi...Full-Memory-Structure-and-Limitations-GTX-970
Other articles about this issue :
Anandtech.com : http://anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation
Hardware.fr (translated to ENG from FR) : https://translate.googleusercontent...t.html&usg=ALkJrhhioO-0LvIVwMt02vN9TS-PEnMARQ
Thank U.
PS. One last thing : "52 ROP's number"
Well, SMM's in GTX 970 can do 52 Pixels/cycle, so that's the number I keep seeing poping out.
But with that logic, GTX 780 would have 32 or 40 ROP's (depending on which SMX's were disabled) and GTX 780 Ti would get 40.
ROP's can do for example AA without SMM's help.
+ Nvidia isn't stupid : They wouldn't put something in GPU if they couldn't use it - lol.
Last edited: