Uh, the whole point of Directx 12 is to be a low-level API that puts the developer in the driver seat. Eurogamer even address that in the article along with this:
You didn't read the article I take it. Direct quote (bolding done by me):
They tell us clear as day that Directx 12 Hitman causes VRAM issues, and that developers are in charge of memory management.
The fact that DX12 is a low level API doesnt mean that it somehow becomes impossible for AMD or Nvidia to do driver based optimizations, this is of course also evidenced by all of the improvements we have seen in DX12/Vulkan from new drivers.
Hitman is ~12% slower at 1080P on the 3GB 1060 compared to the 6GB 1060, and when you move up to 1440P the gap only grows to ~15%. This is nowhere near the 30% slower we saw in scene 2 of the Tomb Raider benchmark, and quite frankly quite close to how much slower we would expect the 3GB 1060 to be simply based on shader performance (the 1060 3GB has 10% fewer shaders).
So yes I did read the article, but I also looked at the actual numbers instead of just blindly accepting Eurogamers conclusions.
Averages don't tell the story on a card like this, if you get spikes of low frame rates due to VRAM in intense scenes it doesn't make you feel any better that the average is high because the game runs at 80 fps (on your 60hz screen) in a basic landscape scene.
You are absolutely right that averages don't tell the whole story and the 1060 3GB obviously has spikes down below it's 51.2 FPS average, but apart from the initial scene change these spikes are actually very mild (only about 10% lower than the average) and the time frame curve is almost perfectly smooth, so it doesnt really look like spikes is an issue here.
Plus we have had evidence that Tomb Raider in Directx 12 mode is a VRAM hog for a while now:
http://techbuyersguru.com/first-look-dx12-performance-rise-tomb-raider?page=1
The test wasn't at 1080p though so it's hard to tell if that is what we are seeing here with the 3GB 1060. A good review with minimums listed would go a long way.
Tech Buyer's Guru makes the classic mistake of confusing allocated VRAM with actual VRAM usage. You really can't use VRAM measurements like this for much.
On another amusing point this link also happens to invalidate your previous claim of Nvidia not being able to do driver based VRAM optimizations on DX12:
"But look more closely: the R9 290 doesn't use all of its RAM, while all the Nvidia cards are pegged at their maximum,
except, interestingly, the GTX 970, which dropped to 3.8GB in the DX12 run. We bet this is Nvidia's driver working hard to reallocate (or more likely remove) data from the slow 0.5GB of VRAM portion of that card's 4GB of memory"
VRAM is simply about turning up the texture settings. Sure with Hairworks or Godrays on high you can kill a 6GB 1060, but it would also kill a 980 or a 480 or any card from that class. The issue is that in the same class of cards (ie below the 1070) this 3GB model might not be able to use the same quality of textures as the other cards in this class. We saw something like that last generation where the 390 could run Mirror's Edge with Hyper textures (at playable framerates) while the 970 couldn't.
Yes texture settings is the main problem with VRAM limits, whilst other settings tend to hit other parts of the GPU. The question then becomes what kind of settings will games offer in the future and even more importantly how big of an impact do they have on IQ. The image quality improvement from very high textures in Tomb Raider is marginal at best, and in Mirror's Edge with Hyper textures it's all but invisible.
Now agree that the 1060 3GB not being capable of running at the same settings as the 470 4GB is a problem, but this issue goes both ways. There are quite likely going to be games where the 1060 3GB cant run the same texture quality as the 470 4GB, and there are going to be other games where it's the 470 4GB that has to use lower settings.
Plus quite frankly I hate the implied concept (not by you but in general) of "well that is a budget card anyway, those peons should have to accept they have to turn the settings down. If they didn't want to turn down settings they should buy a high-end Nvidia card every two years like I do." In my opinion small differences matter at this price level, because by definition people have (or are willing to spend) fewer resources and so they need to get more for their money to maximize the experience.
In that light the 3GB 1060 is a clear clear winner for a "budget gamer" who gets games at Steam sales and doesn't touch AAA games until a year or more after their release. For those people by the time they have more Directx 12 only games than not the 1060 will be a dinosaur. But the 470 might be a better choice for someone who wants to buy and play new AAA games as they come out, as they will maximize the Directx 12 titles that are being increasingly released. At this level your use case matters, but when you buy a 1070 or a 1080 or Titan or whatever you are basically paying extra to not have to consider the use case.
I wouldn't really say the 1060 3GB is a clear winner in anything, I just don't think it's the turd others make it out to be either. The same goes for the 470 4GB imho.
And yes the 470 4GB may very well end up being the better GPU on aggregate, but at this point I just don't think there is anywhere near enough evidence yet to conclusively say so one way or the other.
Doesn't handle Mankind Divided well either:
Slower in Doom Vulkan as well compared to OpenGL vs 6gb:
But if you want to see how the memory will come into play... look at Mirror's Edge:
Any card without the VRAM drops like a rock
It seems to do just fine in Mankind divided as far as I can tell (8% slower than the 1060 6GB, which is certainly within expectations). DOOM Vulkan is also just fine on the 1060 3GB, it's 4.2% slower than the 1060 6GB with OpenGL and 6.5% slower with Vulkan, which is a completely negligible difference.
Mirror's edge on the other hand is definitely an issue for the 1060 3GB, but the 470 4GB would almost certainly do even worse here seeing as the 480 4GB performs worse than the 1060 3GB.
It's also worth noting that the 1060 3GB did better than the 480 4GB in Tomb Raider in this review.
So we now have 1 review (eurogamer) indicating that the 1060 3GB has VRAM issues in Tomb Raider compared to the RX 470, and 3 reviews (ABT, Techspot and hardwareunboxed) indicating that this is not the case.
Makes me think maybe the VRAM thing isn't the issue I thought it might be. Then we move to Mirror's Edge and holy crap:
The 6GB model is playable while the 3GB model slums it with a 380. And that is a game that likes Nvidia cards (notice the 970 even with the 390x, 3.5 GB is enough).
True, but the 470 4GB would most likely also be down at the same level in performance (quite likely even worse based on Hardwareunboxed numbers).
Also notice it loses more performance than the 970 does when Directx 12 mode is turned on in War Hammer:
It also completely falls apart in Vulkan:
Also falls behind the 970 in that game too.
Comparing performance loses in DX12 and Vulkan is silly, when no one in their right mind is going to be using those APIs anyway as long as they are slower than DX11/OpenGL.
There is no dual fan $200 3GB 1060 so the cheapest one with a decent cooler is $210-220. I have seen clearance 970s go for $200 and less sometimes recently, so I think for people wanting to play yesteryears's games (aka the 3GB 1060 forte) they get a better deal with a cheap 970. That is unless the VRAM black hole theory on the 970 is legit.
If you can get a 970 at roughly the same price as the 1060 3GB, then I agree that it might be a viable alternative, I would however be a bit worried about the level of driver support we're going to see for the 970 in the future.