discussion The GTX 970's Memory Explained & Tested (Comment Thread)

[SKYMTL]

1) Would performance improve with a firmware update to make this a 3.5GB card by removing the 500mb partition?
#1 seems answered by the simple fact that the PCIe allocation is still slower then the on card 500mb (plus with optimizations on read/write operations, and striping) that it is fine with the extra ram.

That's pretty true since the GTX 970 CAN utilize that extra partition to its benefit. Had NVIDIA NOT adapted their drivers' load balancing algorithms to their revised memory layout or had done so poorly, we would have had a completely different conversation. However, removing the 500MB would be detrimental to overall performance, particularly after the 3.5GB allocation is saturated.

As for the folks calling this "L3 cache", I'd respectfully disagree with that terminology. Unlike cache, the extra 500MB partition is accessible for direct read / write operations rather than acting strictly like a pool of shared resources. That's a fundamental difference and points towards the GTX 970 being a full 4GB card....albeit with two separate memory partitions.

2) If the answer is 'yes' to 1, should nVidia be lynched?
#2 is...questionable. As this is the first time I can think of nVidia getting thrown under the bus in a long time I'm inclined to give them the benefit of the doubt. Now, should something like this happen again...Lynching should occur

I completely avoided addressing this in the article and with good reason. Editorializing at this point in time does absolutely nothing other than adding fuel to the fire, especially when so many publications are simply talking out of their collective asses (with the exception of Anandtech, PCPer, Hardware.fr, Tech Report and a few others) without having a clue what's going on, simply to up their pageviews. We now know (most) of what went on, there has been a frank discussion about the architecture and nothing has really changed about the GTX 970's performance.

With that being said, and like I discussed in the article, its completely unacceptable that it took a mob of pissed off folks to get NVIDIA to realize their mistake. They have some serious egg to wash off their collective faces.

3) If the answer is 'no' to 1, do I give a crap?
#3 kinda got answered by #2 and #1. My card works, and works quite well. If I get into newer games that require more vram at 1440p and I start to see issues...well...I'll revisit the lynching solution

If you run into a game that needs more than 4GB of VRAM, the only thing that will save you from the PCIe bottleneck is a 6GB+ card....aka: whatever NVIDIA is launching in the next few quarters.

What I think would be fair (to 'put this to bed') is if nVidia provided reviewers with said firmware update...but as I said in #1, I'm not sold that would actually prove anything.

Why cripple performance?

 

[Creig]

NVIDIA says their omission of these elements boils down to a miscommunication rather than any intent to mislead. We’re inclined to believe them since absolutely nothing has been gained by keeping it from us; the GTX 970’s performance hasn’t changed and neither has its overall value to gamers.

I disagree on all three counts.

First, I find it impossible that in the four months since the 970's release that nobody out of the 8,800 of employees at Nvidia happened to notice that the specs given to reviewers were incorrect. I'm sorry, but the "miscommunication" excuse simply does not hold water.

Second, Nvidia had plenty to gain by keeping this design flaw quiet. Money. If it had been known at the outset that the last 500MB of the 970 would cause slowdowns, stuttering and any other number of problems, they would have lost sales. Especially through the Christmas season. Right now you can go to nearly any tech website forum and find people who are upset by this and are wanting to return their 970's in favor of either 980's or 290/290X's.

Third, the overall value has most definitely changed. The 970 is now being labeled as a 3.5GB card instead of a 4.0GB card due to the issues that arise when attempting to use that last 500MB. 970's are now off the list for people wishing to use SLI/CrossFire at higher resolutions. And as future games will undoubtedly continue to utilize more and more memory, the 970 may simply not be able to keep up with the 980 or 290/290X's true 4GB of RAM.

This isn't the first time in the negative spotlight for Nvidia and I suspect they will employ the "keep quiet and eventually it will all go away" strategy that they've perfected over the years.

 

[SKYMTL]

First, I find it impossible that in the four months since the 970's release that nobody out of the 8,800 of employees at Nvidia happened to notice that the specs given to reviewers were incorrect. I'm sorry, but the "miscommunication" excuse simply does not hold water.

It really boils down to a "he said / she said" situation. Innocent until proven guilty is the way I approach this and there's no proof of malicious intent. In the court of public opinion though, emotion usually wins out.

Second, Nvidia had plenty to gain by keeping this design flaw quiet. Money. If it had been known at the outset that the last 500MB of the 970 would cause slowdowns, stuttering and any other number of problems, they would have lost sales. Especially through the Christmas season. Right now you can go to nearly any tech website forum and find people who are upset by this and are wanting to return their 970's in favor of either 980's or 290/290X's.

Knowing about this issue hasn't changed the GTX 970's actual performance one iota. Check out all the reviews including our SLI review which looked at frametime data. How would knowing about this 500MB partition affected our opinion or the public's opinion of the card from day one? From what I see, pretty much everyone is trying REALLY hard to find issues and that's the crux of the situation; finding one of those rare scenarios is extremely difficult and usually requires resolution scaling, odd DSR settings or something else.

As for the folks who want their money back, I think that's a non-starter. It's in human nature to jump on a bandwagon and that's exactly what is happening here. While we can rag on NVIDIA all we want for waiting five months to come forth with this information, we can also point a finger at all those consumers who bought the card, bragged about their newfound in-game performance and are now suddenly searching for virtually non-existent problems while turning around and demanding their money back. It's ridiculous IMO.

Third, the overall value has most definitely changed. The 970 is now being labeled as a 3.5GB card instead of a 4.0GB card due to the issues that arise when attempting to use that last 500MB. 970's are now off the list for people wishing to use SLI/CrossFire at higher resolutions. And as future games will undoubtedly continue to utilize more and more memory, the 970 may simply not be able to keep up with the 980 or 290/290X's true 4GB of RAM.

I'll play devil's advocate here for a second. With the advent of DX12, resources will be handled in a completely different way with the texture memory playing less of a role than in previous generations. Under no circumstance is the GTX 970 a 3.5GB card, that much has been proven (see my post above), but even if it was, future titles that utilize the new API or PROPERLY use DX11 rather than being a crappy console port will continue to effectively manage resources and hit levels well under 3.5GB @ 4K.

 

[10e]

So what this boils down to in my "idiot layman's terms" is that 3.5GB of memory runs at full speed, and a 500MB partition runs at a reduced speed, correct?

To me it's not a big deal. I do believe nVidia should have been more forthcoming about how they cut down the 970 versus the 980, and I'm a bit disappointed because I have generally held that company in relatively higher regard than their (once) Canadian based counterparts, but at the end of the day I'll keep my GTX 970s and still keep running them, as I've generally been happy with them in single card or SLi configurations. But I still have some R9 290s that I will try in similar situations in Crossfire as a comparison.

Good article, SKY and thanks for giving a straightforward (as possible) report on your investigations.

This does make my GTX 780 6GB card look a lot better than it once did though.

 

[Ernimus Prime]

It MemoryGate for Nvidia!!!!

 

[Fragman]

It really boils down to a "he said / she said" situation. Innocent until proven guilty is the way I approach this and there's no proof of malicious intent. In the court of public opinion though, emotion usually wins out.

This is an unscruplous company we are talking about which fought a class action lawsuit and even when settling with consumers denied any wrongdoing - the GPU bumpgate scandal. The GPU bump failures were due to an engineering mistake and Nvidia just never owned up to it. so people are not going to give Nvidia the benefit of doubt given they have a bad history.

http://www.canadiannvidiasettlement...greement - Fully Executed (June 12, 2013).pdf

forget emotion and apply a bit of logic. When a high tech product is designed and manufactured the specs go from the engineering team to the marketing and not vice versa. marketing knows nothing until they are briefed about the product and given a spec sheet. Even then marketing does not know any of the in depth technical details if they are required to answer questions like how the VP of engineering is doing now after this memory issue got found. So what ever specs was given to marketing was decided by engineering. If marketing assumed the specs and engineering failed to correct them before launch why did it take more than 4 months after launch. See when Nvidia has no answers to perfectly logical questions obviously the public is going to call it a bluff. Actually in this case your emotions seem to be affecting your judgement.

Knowing about this issue hasn't changed the GTX 970's actual performance one iota. Check out all the reviews including our SLI review which looked at frametime data. How would knowing about this 500MB partition affected our opinion or the public's opinion of the card from day one? From what I see, pretty much everyone is trying REALLY hard to find issues and that's the crux of the situation; finding one of those rare scenarios is extremely difficult and usually requires resolution scaling, odd DSR settings or something else.

It has affected performance from day one especially for people who bought GTX 970 SLI for 1440p and 4K gaming and played at settings which hit VRAM usage between > 3.5 GB and <= 4GB. Just because you did not test at those settings does not mean the performance issues did not exist. btw your GTX 970 SLI test was done on games released before 2014 (except Thief). 2014 was the first full year of the next gen consoles and you could see quite a few titles like Middle Earth, COD AW, AC Unity, Lords of the fallen easily cross the 3.5 GB limit even at 1440p.

http://www.hardwarecanucks.com/foru...-nvidia-gtx-970-sli-performance-review-5.html

As for the folks who want their money back, I think that's a non-starter. It's in human nature to jump on a bandwagon and that's exactly what is happening here. While we can rag on NVIDIA all we want for waiting five months to come forth with this information, we can also point a finger at all those consumers who bought the card, bragged about their newfound in-game performance and are now suddenly searching for virtually non-existent problems while turning around and demanding their money back. It's ridiculous IMO.

How do you think the issue came to light. Gamers noticed the performance issues when the last 0.5 GB was used in single and SLI configs. The forums even at Nvidia saw such complaints earlier than the CUDA program which confirmed the memory issue. That program was only a reaffirmation of the problems faced by consumers and gamers.

I'll play devil's advocate here for a second. With the advent of DX12, resources will be handled in a completely different way with the texture memory playing less of a role than in previous generations. Under no circumstance is the GTX 970 a 3.5GB card, that much has been proven (see my post above), but even if it was, future titles that utilize the new API or PROPERLY use DX11 rather than being a crappy console port will continue to effectively manage resources and hit levels well under 3.5GB @ 4K.

The fact is we are talking about both performance in the present and the future. GTX 970 is a 3.5GB + 0.5 GB card. The bandwidth is not 224 GB/s. That is something which everyone now understands.

 

[SKYMTL]

I'll keep on repeating this: the peak bandwidth is an effective 224GB/s.

 

[Fragman]

I'll keep on repeating this: the peak bandwidth is an effective 224GB/s.

Funnily Nvidia's spec page for GTX 970 does not even mention the word "Peak" when they mention Memory Bandwidth. :biggrin:

GeForce GTX 970 | Specifications | GeForce

 

[SKYMTL]

And they should.....just like all other vendors.

While they may think it is a foregone conclusion that folks realize a memory bus will never operate at 100% bandwidth all the time, it is something that should have been indicated.

 

[Dzzope]

Anytime anyone says bandwidth = xxx, they mean peak.

While I will agree thet 5 months is far too long for them to have come clean, no-one can say whether they initially mis-lead or if it was truly miscommunication. I think they kept quiet after realizing in order to not cause a scandal.. Afterall as said above sooo many lorded this cards performance...
As to the performance / refunds / those saying it's only 3.5GB... Learn to read.