GTX 970 3.5 Ram issue

Question

So we have now had it confirmed that the specification of the 970 GTX was wrong. These specifications were released by Nvidia and these were the specifications printed on the box of my MSI Gaming 970 4G. Namely that card has 4G of Vram running on a 256 bit bus. In reality we now know that only 7/8th (3.5G) of the 4G vram operates at the expected 224 Gb/sec memory bandwidth. We also know that the cards performance will seriously degrade in scenarios where a game tries to access that last 0.5G of vram.
Games increasingly used up more Vram when using high res texture packs. Also running higher resolutions and sli profiles require much more Vram and can easily exceed 3.5G.
What are MSI's thoughts on this issue?
The new MSI 970 Me100 editions, is this a true 4G card or will it also be limited to an effective memory pool of 3.5G??

Answer 1

Quote from: Airjarhead on 01-February-15, 07:13:57
Sure, Nvidia, MSI, EVGA, Asus, etc. mislead us with the whole 3.5GB+512MB issue, but I don't think it was an egregious lie.  It was more a marketing issue.  After all, it is 4GB of total RAM, it just doesn't function as we would expect 4GB to function (as with GTX 980).  On this matter, I can see the manufacturer's/board partner's/retailer's point.
It is also true that the GTX 970 is still the same card (benchmark wise), that we bought after seeing the very first reviews.
HOWEVER...the following specifications prove a lie that can not be explained away with Marketing terms or benchmarks:
Advertised ROPs: 64
Actual ROPS: 56
Advertised L2 Cache: 2MB
Actual L2 Cache: 1.75MB
This is false advertising.  This is the lie.  This is why I am angry, and I do not understand why any user would defend Nvidia/board partner's/retailer's for this.
I don't think people is defending nVIDIA or AIBs in the first place. I have absolute no doubt anyone that bought the card liked to find the card was advertised with wrong specs. I know I am p***ed at nVIDIA - and nVIDIA alone -, because what they did. They now say there was a "missunderstanding between PR and engineering" and I believe them; what I find hard to believe is that not a single engineer noticed the PR mess before January 2015. But, this is a different discussion and I'm only speculating.
What I don't accept (and I think that is what some owners are having issues with), is the card having a flawed design; that is not working as intended. I defend nVIDIA engineers decision to cut the L2 cache so die-harvesting GTX 970 capable GPUs would increase, thus reducing the price. And at this moment, knowing the card's true specs I'm even more astonished with my GTX 970 performance. It really does more with less. And does it for €350 (that was how much it cost me). Plus, it completely destrois the GTX 780 I had before (and sold for €300 after using it for 18 months).
For nVIDIA to come out with a "new" GTX 970 with same ROP and L2 cache as GTX 980, they have to price it higher because is harder to harvest dies that will fill those specs.
And here's the ugly truth: if such a card cames out, it will stutter exactly the same as present GTX 970 stutters. Why? Because it will have the same CUDA CORES as the present GTX 970. The card will get out of "cuda core" power long before getting out of VRAM. The GTX 970 as is, has 1664 CUDAs that can handle with 52 pixels per clock, while the ROPs can handle 56 pixels per clock. Do you see where the bottleneck is?
So, having ROPs being capable of handling 64 pixels per clock would do anything for the GTX 970 unless the CUDA count increases - in which case, we will end up with a GTX 980.