NVIDIA's Current Linux Driver Is Hungry For vRAM This Holiday

[phoronix]

Phoronix: NVIDIA's Current Linux Driver Is Hungry For vRAM This Holiday

With a NVIDIA Linux developer having confirmed a current driver performance regression affecting driver releases since the 378 series and not being worked around until the yet-to-be-released 390.xx beta driver, I decided to carry out some tests...

http://www.phoronix.com/scan.php?page=news_item&px=NVIDIA-Memory-Problem

[xorbe]

If they have any sort of version control, I fail to see how they can't pinpoint the effect fairly rapidly with a binary search. Sounds like an attempt to drive hw sales to me.

[Michael]

If they have any sort of version control, I fail to see how they can't pinpoint the effect fairly rapidly with a binary search. Sounds like an attempt to drive hw sales to me.

What's your hypothesis on 'drive [hardware] sales' when this issue occurs even with their latest generation hardware available?

[OneBitUser]

It would have been a better idea to test this issue with cards like say the GTX 1060 3GB and GTX 970, or 780 Ti.
I also do not know if DX:MD is the game to test stuff like this... aren't there even more vRAM-intensive games at your disposal?

[benklett]

It would have been a better idea to test this issue with cards like say the GTX 1060 3GB and GTX 970, or 780 Ti.

Did you read the last sentence?

I am still testing other graphics cards, but given the range of tests being used (the ones on this page are just a sampling), it's taking a fair amount of time. I should have more data completed during the weekend.

[OneBitUser]

Did you read the last sentence?

Yes, so what?
I was not saying that all those cards should have been tested already. I just think it would have made a lot more sense to perform initial testing on a card that was far more likely to be affected by the issue.
Testing the effects of issues caused by too high vRAM usage on an 8GB card just does not feel right.

[benklett]

I just think it would have made a lot more sense to perform initial testing on a card that was far more likely to be affected by the issue.

I think the same way. But I also think you have to have at least one card, which does have enough vRAM to make sure it is not affected.

[Tomin]

If they have any sort of version control, I fail to see how they can't pinpoint the effect fairly rapidly with a binary search. Sounds like an attempt to drive hw sales to me.

They said they know which change is the culprit and they will have a workaround, but they also want to fix it properly.

https://devtalk.nvidia.com/default/topic/1026874/linux/huge-performance-losses-with-newer-nvidia-drivers/post/5223375/#5223375

We've been tracking it internally as bug 1963500. There was a change, introduced in our r378 branch, to the logic of allocation of certain textures, but it apparently exposed a bug in our memory manager.
Our next release branch, r390, will carry a workaround, and we're still working on finding and fixing the root cause.

I understand this so that they know why the driver was working well before. They know what they can do right now to workaround it, but they also know that there is another bug elsewhere that they would want to fix so that they don't need the workaround. I think they know what they are doing here.

[L_A_G]

My first thought was that this could explain the minor performance regressions I've been seeing with my GTX 970, but upon closer inspection it seems like this was introduced too recently to be the culprit. Still, I wonder if this is a universal bug or just a Linux-specific bug introduced due to changes made to cross-platform code without consideration for the effect on Linux.

[oleyska]

What's your hypothesis on 'drive [hardware] sales' when this issue occurs even with their latest generation hardware available?

All users with GTX970 (Me was included) had some nice issues above 3.5gb vram usage.