970 memory allocation issue revisited

wonder y linustechtips, jayztwocents etc never come out say something bout this particular scandal from nvidia?? i think they knew it earlier that's y keep suggesting SLI for GTX 970.

 

if all 970 users fight for it? we ask for redemption.

 

What redemption? Honestly, does anybody think he'll get money back?

 

So...us SLI users aiming for 1440p gaming are done for?
Interesting how this explains the stuttering I've had in some games.

 

I should quote what neo said about lie is lie and false advertisement is false advertisement. You make excuse just to defend your purchase but at the same time you defend the company that did wrong practice by this justification; " impact is little to none and by the time the impact is large i already upgrade my card" kind of excuse.

 

Speak for yourself. Last I checked, benchmark results are the same -> I'm happy with my card.

And they're not going to give you any money back. Makes no sense. The performance is exactly the same, I doubt people buy cards based on specs instead of benchmarks.

 

you can fight alone, ive gone from a 760 to 970 and am totally happy about it, i have yet to see problems going over 3.5, i even made a video in a previous post showing AC unity working fine at 4gb

i have better things to do in life than fighting for a 0.5 mb partition in a video card

 

Yep, that part momentarily had me confused. Thought that the SMMs could only feed certain allocated portions/addresses of the memory directly by the bit bus (ie, 980 = 4gb/256bit with 16 SMMs = 16bits per SMM, while 970 with 13 SMMs x 16bits = 208 bit = 3328mb, essentially Headds mistaken argument). The crossbar does the job of sorting out the memory distributions of the SMMs. But now thats all clear, I wonder, does the remaining slower part of the memory (512mb) drag the rest down in any way if exceeding 3.5gb usage or does it do a useful job of caching? Meaning would the card have been more effective if it did not have that extra 512mb and just kept as a 3.5gb card? Was making it a 4gb card just for marketing purposes or does it help its functionality as well?

 

this being the first I heard of this ,I can understand some 970 owners up in arms but its not something to worry about.just going by reviews it seems like a really good card.

 

jajajajaj I agree,like wanna buy a bridge? it was the pr's idea most likely

 

Not worried at all, just curious. Even today after all this, would still by the card.

 

Most crazy people won't exactly be nice to if you're trying to challenge their 'religion', haha

 

nvidia must fix this

I've been building pc systems for 15 years and this bug is not a small thing.to all of you who say that we should give it a rest i say: hell no.benchmarks are important but specs are no less.i bought this card knowing i'm getting one set of specs and now i find out that it's revised?no.i can't accept that.nvidia must fix this.period.

 

If we compare with previous generation, the GTX570, it seems Nvidia engineer have try optimize a bit what happend when you disable the SMM / ROP... In previous generation, their arch was absolutely not permitive in this regard, as the ROP/MC was completely tied to the SMM configuration. with Maxwell they have a bit more liberty. ( Far of AMD who can disable any part in any place and still keep the other fully up )

So basically, they have try find a way to use and keep the most over the old solution who was to prefer, disable 2 SMM + the ROPs relative + the MC relative and the memory.. ( hence why we had a 320bit bus + 1.25GB memory on the 570 vs 384bit/1.5GB memory on the 580 )

Well i will point to this article where they have report the "famous" MC/L2 schema too for 980-970M etc.. https://translate.googleusercontent...t.html&usg=ALkJrhh-MhPqRjv5czgbK6qJiP2y5NU3EA

( it report much over Anandtech article, but with a little insight on driver optimization )

This said, for my part, this change nothing for 970 owners, the card still perform the same as when they have buy it, its just a bit bad that Nvidia have not choose to communicate directly on this, it could even have been a marketing argument.

They should have advertize it in the launch, saying what happend when the gpu is cutted down in SMM and push this information for show how they have fix the problem.. you will have end with a 224bit / 3.5GB GPU memory who have an internal optimization for allow it to work closer of a 256bit / 4GB memory ( or the 980 in term of memory ). It was maybe complicate the communication when launch the card and marketing team have decided to keep it simple.

Funny enough i remember have ask a reviewer on the 970-980 launch article if he had badly report the ROP count, as with 3 SMM disabled, i was not seen how the ROP cant be disabled at the same time ( they are in the SMM )

 

As some people trying to downgrade the issue it's as simple as this guy said. Maybe your card is still functioning and you can still play games but this is not what we are supposed to buy. By the way isn't this illegal?
Nvidia must fix this and refund or fix/exchange the 970's people already bought . Changing the spec sheet after the fiasco and lying about the reason this happened in first place does not change a thing. I really don't care if one department made a mistake or another. This an internal issue and we don't care about that. We bought something with the wrong spec sheet from Nvidia and this must be fixed. Period.

 

I'm pretty dam sure cod:aw does not support msaa. So..
Lol?

 

It only slows down another memory chip with which it shares connection to crossbar.
Therefore you either have 3.5GB fast and 0.5G Disabled (unused).
Or you have used all 4 GB, but 3GB fast, and 2x 512MB sharing 1 channel. (here remains to be seen if only last 512MB gets severely affected as it has no priority access or both get slowed)
And it is effectively 224bit bus from crossbar to memory where last 32bits are either dedicated to 3~3.5GB block or shared for full 3~4block reducing performance.

 

They talk about that at PcPer:


The 0.5GB section is slower to the 3.5GB section but not slow enough to cause any slowdowns. You can use up to the 4GB without needing to worry about it.

 

How we can compare a 7Ghz 32bit partition of the memory to the system memory? its pretty easy just divide its effective clock and multiply its bus like that :

512MB 32bit 7Ghz = 512MB 32x4 7Ghz/4 which will give us something about the speed of todays DDR3 systems memory 128bit with 1.75Ghz, according to nvidia this partitin is still 4x faster than systems memory - HOW ?

 

Thanks for the explanation guys. Quick question, would any scenario that causes problems for the 970 over 3.5gb (ie, like Skyrim and exaggerated texture mods) also impact a 3gb 780ti in the same way due to vram limitations?