NVIDIA GTX 970 False Advertising (UPDATE: NVIDIA settles class action lawsuit)

As you may have heard in the news, NVIDIA has been falsely advertising the GTX 970 to have 4GB of memory on a 256Bit bus with a peak transfer rate of 224GB/s. It turns out that the actual transfer rate of the last 500MB of those 4GB is almost a magnitude lower. This can cause significant performance degradation when that memory is accessed.

Memory bandwidth is crucial for an application like Cycles to perform well, however since the data is accessed more or less randomly, the performance hit might be amortized.

It would be interesting if any owners of a GTX 970 could report on any performance degradation when using up those last 500MB in a Cycles scene. You can use a tool like GPU-Z to monitor your memory usage.

UPDATE (Juli 29th, 2016):

NVIDIA has agreed to pay 30$ per device in a settlement of a class action lawsuit. Those who have purchased a GTX970 may be entitled to this sum (after filling out some paperwork).

what about gtx 960 and 980? are these same false ads?

No. The 960 is a different chip, the 970 is a cut-down version of the 980. This problem is apparently related to that “cutting down process”.

Wow. I was going to buy one of these. This is the first I’ve heard of the false advertizing.

Thanks, BeerBaron.

Wow. I just bought one of these… :frowning: Well, I understood that this isn’t much of an issue when the chip is using under 3,5Gb of its memory? If that’s the case, it’s not that bad I guess. I rarely go even over 2GB, because I usually try to take advantage of my GTX770 also. But yeah, a bit disappointing…

It shouldn’t be an issue, the driver will try to allocate in the lower regions first.

I’ve bought one before christimas, and I noticed there was something strange when dealing around 3.4 - 3.7 gb of memory, however nothing terrible and render times were nice.
I think it’s still a worthy purchase, and cost wise it’s really okay specially if you had an AMD just like me before.
For sure It was better to know before about this issues and I think it’s rather unfair advertising, but since the problem does not seem to have huge impact I’ll just deal with it.

nVIDIA what do you expect …
Thats what happens when you have a monopol
They should actually go after them for claiming it has 4GB with the same speed.

Would be cool if someone could devise a benchmark test to really figure out if the render speed drops significantly over 3.5gb. Would have to be two tests that should take the same amount of render time but use different amounts of ram. Hmmm

hello everyone! Here’s Nikola, I just joined the forum yesterday.
I’ve been watching the forum - especially benchmarks threads for a while now, running a igpu and stalking the prices on a 970 when this came out.
I’m not a blender artist, but just can’t wait to start learning, been saving for a while so I’d love to hear your info on the issue, I don’t know if it’s a game changer, and like @blackdot I also think some benches in that vram range could provide some useful info for us all.
Cheers! see u around!

Not only the VRam but also other graphic unites are cut and marked incorrectly.
Nvidia claims it is a misunderstanding of the technical marketing department.

I hope enough users might come together to actually press charges.

The 970 looks like a recycled model of failed 980 chips, which is fine, but then declare it also correctly.

Is there an app than can measure VRAm usage for CUDA with Cycles?

The web is basically full of people mentioned game performance hits significantly once they hit the 3.5 GB VRam mark.

As usual, people could file a class action lawsuit and be refunded 15$ or something (after a couple of years of legal proceedings, of course).

Is there an app than can measure VRAm usage for CUDA with Cycles?

As I wrote in the opening post, you can use a tool like GPU-Z to monitor the VRAM usage of the system (you don’t want to just monitor the usage of the Blender process in this case).

The web is basically full of people mentioned game performance hits significantly once they hit the 3.5 GB VRam mark.

I’d rather people performed some actual tests. The games benchmarks I’ve seen did not show a big hit after 3.5GB. I have yet to see any (non-synthetic) compute benchmarks.

I’ve been trying to explain to lots of folks that in-game fps ain’t not the only concern that comes with this fiasco, sadly to no end…
nwm I think it’s really important to do some benches and help other on future decisions on buying this card

I think any monitoring tool should do, although there were some notions of some apps not correctly displaying VRAM loads, just use techpowerup GPU-Z or MSI afterburner and give cycles a try on a scene that takes 3.5~4GB vram

Why did you link to the old article From AnadTech, rather than the one from the 26th that gives Nividas corrected specs and explains in detail how the memory systems differ between the 970 and 980?

This whole controversy seems very meh. Although the new info does explain the weirdness people have been observing it doesn’t change any of the existing benchmarking that has been done. The real world performance of the 970 is baked in to those benchmarks regardless of what the spec sheet says. The 980 is still about 15% faster than the 970, while costing about 65% more.

If I had purchased a 970 I certainly wouldn’t be regretting it at this point.

The best part of those class action suits is how little money the class is entitled to, while the lawyers rack in huge fees. That $15 for the P4 settlement is hilarious compared to the $16,400,000 the lawyers are getting paid.

The older article actually contains the statement from NVIDIA (along with some numbers on performance), the newer article is much longer and doesn’t add any relevant new information (unless you care about technical minutia). As far as I can tell, those “corrected” specs were made up by the author and NVIDIA did not correct their specs.

The real world performance of the 970 is baked in to those benchmarks regardless of what the spec sheet says. The 980 is still about 15% faster than the 970, while costing about 65% more.

If the benchmarks don’t use >3.5GB they tell you nothing about the performance characteristic of this issue. For a Cycles user, it is very relevant to know if their GPU takes a hit beyond 3.5GB and none of the benchmarks here reflect that.

I want to ask, how much of a real world difference does this make? I’m only a novice at blender and I was wondering how cycles allocates VRAM and what would take for it to go over 3.5GB. Is it the complexity of the scene, tile size, detail of textures or some other factor what I am unaware of?

Tech Report and PCPer are reporting the same thing.

http://techreport.com/review/27724/nvidia-the-geforce-gtx-970-works-exactly-as-intended

but wouldn’t using the full ram allocation per GPU unit allready cause instability problems and other isues.

I allways had the impresion you should NOT push a product to its limits just because you CAN.

Or is that only valid for Overclocking and temperature?

I’m not saying it’s untrue, but it is not the case that NVIDIA officially declared these the new specifications. The specifications on their website remain unchanged.

It might cause performance issues, it shouldn’t cause stability issues. Now it causes performance issues 500MB earlier.