GTX 970 3,5GB + 0,5GB with 56 ROP and less cache than advertised.

Now nvidia have come clean they are on major damage limitation.

I wonder how they will deal with it and how they will address the angry consumers as most people including myself bought these cards for high res gaming and future proofing.

This is the true test for nvidia, how they deal with this mistake will show what kind of company they truly are

I can agreed with that too
 
Straight from the Lion's mouth (in this case, Jonah Alben who is the senior vice president of GPU engineering at Nvidia):
..GeForce GTX 970 features 56 raster operating pipelines (ROPs) and 1792KB L2 cache, not 64 ROPs and 2MB of cache, as previously reported. In a GeForce GTX 980, each block of L2/ROPs directly communicate through a 32-bit portion of the GM204 memory interface and then to a 512MB section of on-board memory. Since the model GTX 970 lacks part of L2 and ROPs, it cannot effectively use 512MB of 4096MB of onboard GDDR5 memory.
I got this from Kitguru, who got it from PCPerspective.
 
Straight from the Lion's mouth (in this case, Jonah Alben who is the senior vice president of GPU engineering at Nvidia):

I got this from Kitguru, who got it from PCPerspective.

Yep and the memory bandwidth calculated and advertised on the side of your box is based on those original specs and the assumption that all the memory was running at the same speed.:mad:
 
It gets better, I just ran gpu z just to see and guess what, according to gpu z my 970 has 64 Rop's?????

So did the PR team write the bios as well
 
It gets better, I just ran gpu z just to see and guess what, according to gpu z my 970 has 64 Rop's?????

So did the PR team write the bios as well

That is very interesting, especially when this miscommunication is flying around and that Nvidia have admitted that the 64 ROPs was not accurate. Could there be more to this coverup than we know? ;)

Anyone else with a 970 who can check their ROPs in GPU-Z. Does GPU Shark also report ROPs?
 
TBH unless Nvidia do something to repay all the GTX 970 owners which they frankly have cheated, I'll be moving to AMD when I next upgrade.
 
remember this is the first generation of Nvidia GPUs which could have a bit of l2 cache disabled in this way.

If somebody didn't notice this this time Nvidia could have done it again and not told us.
 
I'm not excusing Nvidia or anything, they have lied and should have been honest about things.

But the 970 is a 4GB card it has 3.5GB VRAM and a separate, but slower 500MB of VRAM. Regardless of specs and what they said it had and what it actually has, the performance is still the same today as it was last week.
The real question is does the 500MB lower speed VRAM actually affect the performance of the card and if so by how much?

The way the 970 works is it will use the 3.5GB first and when it goes over the 3.5GB it will access the other 500MB which is slower. So what we need to find out is when the 500MB is accessed, does it actually have an impact on the performance and how big of a problem is it?

There has been no proper testing done yet, so I'll wait to see what places like Pcper and Anand have to say once they have done some proper testing and confirm how much of an impact this actually has before drawing a proper conclusion.

Doesn't change the fact Nvidia lied though, I have a 970 myself so it directly impacts me.
 
To be fair if it's physically on the card they haven't really 'lied' at all. It would be like accusing them of lying because the Titan Z doesn't have 12GB of Vram. It does. You just can't use it.*

It doesn't change what I think about nVidia at all, it is what it is. They always were lets say interesting with the truth. Most of the graphs they present don't start form 0 or have a linear scale to visually mis-represent improvements etc. everything they actually say on the other hand always seems very controlled and considered.


*well you can but you can't appreciate it because you just have two copies of the same thing :D.

JR
 
Last edited:
I'm not excusing Nvidia or anything, they have lied and should have been honest about things.

But the 970 is a 4GB card it has 3.5GB VRAM and a separate, but slower 500MB of VRAM. Regardless of specs and what they said it had and what it actually has, the performance is still the same today as it was last week.
The real question is does the 500MB lower speed VRAM actually affect the performance of the card and if so by how much?

The way the 970 works is it will use the 3.5GB first and when it goes over the 3.5GB it will access the other 500MB which is slower. So what we need to find out is when the 500MB is accessed, does it actually have an impact on the performance and how big of a problem is it?

There has been no proper testing done yet, so I'll wait to see what places like Pcper and Anand have to say once they have done some proper testing and confirm how much of an impact this actually has before drawing a proper conclusion.

Doesn't change the fact Nvidia lied though, I have a 970 myself so it directly impacts me.

Yeah, I am also an owner of a GTX 970, and while performance is the same as at was when it launched, had I known what I know now, I'd have reconsidered the purchase.

I hope Nvidia do do something to address this, just acknowledging the problem isn't enough. They lied to us, what we paid for isn't what we have received!

We don't get 4GB or ram at Nvidia's specified speeds, we get 3.5GB (of somewhat) slower Vram and 0.5GB of a ram cache, which while it is much faster than system memory isn't what we paid for.

Then there is the issue of the smaller amount of ROP units.

While what we buy is very similar to the GTX 970 that was advertised, it still wasn't what we were sold.
 
Last edited:
The fact they lied about the memory bandwidth and number of rop's is the most compelling thing. Not only did they outright lie but it's also encoded in the bios which leads me to believe they had no intention of telling anyone and only when end users noticed their gpu's performance taking a nose dive when the last 500mb of vram was used and began to look into it themselves and found the anomaly did nvidia shit a brick and make a statement
 
I'm not excusing Nvidia or anything, they have lied and should have been honest about things.

But the 970 is a 4GB card it has 3.5GB VRAM and a separate, but slower 500MB of VRAM. Regardless of specs and what they said it had and what it actually has, the performance is still the same today as it was last week.
The real question is does the 500MB lower speed VRAM actually affect the performance of the card and if so by how much?

The way the 970 works is it will use the 3.5GB first and when it goes over the 3.5GB it will access the other 500MB which is slower. So what we need to find out is when the 500MB is accessed, does it actually have an impact on the performance and how big of a problem is it?

There has been no proper testing done yet, so I'll wait to see what places like Pcper and Anand have to say once they have done some proper testing and confirm how much of an impact this actually has before drawing a proper conclusion.

Doesn't change the fact Nvidia lied though, I have a 970 myself so it directly impacts me.

Testing has been done. Check guru3d. There is even highlighted screenshot of the last cluster showing a rapid drop in the bandwidth. Someone have also tested the vram bandwidth test by allocating the first 1gb vram immediately then running the bench.

WYP
You are actually getting 3.5gb of fast ram on one partition and 500mb of ram on a second partition which has to share the 7th and 8th memory controller with the 7th L2 so yes slower bandwidth. I'm not defending them, but they haven't really lied to us about anything except the ROPs.

I hope they are hammered for this. I'm sick of their overpriced products, ... which I still paid for :(

Taken from another site to show others exactly what the issue is incase it hasnt been understood. I'm using the diagram as I do not have a 970 to bench my own results

index.php


Now to understand what you are seeing, (for the non tech heads) pay attention to the bandwidth on the far right of the 970 benchmark. The dive in bandwidth begins at 3.2gb because of the 300mb allocation that windows AERO uses. Each chunk represents 128mb of DRAM.
 
Last edited:
Back
Top