NVIDIA’s GeForce GTX 970 Has A VRAM Flaw, But Should It Concern You?

It’s been rumored for a little while that NVIDIA’s GeForce GTX 970 has a memory issue, but as I could never find substantial evidence of it, I passed it over. Well, this past weekend, it was impossible to ignore. A couple of weeks ago, a reddit user posted some findings about how his GTX 970 exhibited severely throttled memory performance after a certain amount of its VRAM was utilized. That in turn led to countless numbers of GTX 970 owners to run the same benchmark and post their results.

Ultimately, it seems that after 3.5GB is filled up on the GTX 970, memory performance plummets. This was established with the help of a tool called Nai’s Benchmark, and should you want to learn more about running it yourself, you should refer to here.

There are a couple of things to bear in mind, though: it seems that in order to produce an accurate result, the benchmark needs to be run “headless”, which means hitting enter to run the benchmark when your monitor is completely unhooked from the GPU (or the monitor, if it’s convenient). Doing so will produce more accurate results. If Nai’s Benchmark could be run in full-screen mode, this work-around wouldn’t need to be done, but alas, it’s a Windows CLI tool.

Our friends at The Tech Report reached out to NVIDIA over the weekend to see if there was an explanation, and it confirmed a popular suspicion: Due to the SM arrangement on the card, there are fewer crossbar resources (lanes where memory can be transacted) than on the 980. That affects the final 500MB of VRAM.

NVIDIA doesn’t consider this to be as severe an issue as many lead on. It’s run some tests in its lab that show when more than 3.5GB is utilized, the FPS drop on the card is on par with what’s seen on the GTX 980, give or take an additional FPS or two (which could be variance more than anything else).

It goes without saying that this is an unfortunate design, and to some, it should have been one that NVIDIA told people up-front about. But given its performance findings, I am not sure that would have been entirely necessary. At the end of the day, it’s the final 500MB of VRAM that’s crippled; the rest is fine.

Credit: The Tech Report

One of the reasons I don’t consider this to be as huge an issue as some is because the vast majority of GTX 970 owners will not even surpass 3.5GB of VRAM to begin with. That takes special circumstances. If you’re running 4K, this is an issue you might want to be aware of, but for 1440p or lower, this bug is likely to be invisible. I’d say the same thing about 4K, but it seems like benchmarks could show a frame or two detriment.

But does that matter? Not likely – the GTX 970 still delivers a ton of performance at its price-point. It’s as simple as that. I don’t think this is an issue NVIDIA mistakenly let into the card; this is its design. To produce a model like this, in the way that NVIDIA wants it to perform, certain subtle bottlenecks might result. In this case, it would seem NVIDIA deemed this one to be a non-issue, and from what I’ve seen up to this point, I’d agree. But again, this still isn’t what I’d call ideal, it’s just subtle enough to not warrant a stink being made like one is.

For those concerned about the GTX 980, it’s unaffected due to the fact that it has a proper SM layout that allows all memory to move through the pipeline without a bottleneck.