Here is the review or 680 vs 7970 3 years after their released. It's quite clear that back in 2012, 680 is generally faster. And getting slower and slower if compare to the 7970. By considering both card's processing power won't change over time. It's not too hard to conclude that should be the VRAM size limitation make the difference. Therefore, I will consider a 4GB VRAM card is a better option.
http://www.babeltechreviews.com/hd-7970-vs-gtx-680-2013-revisited/
[Alien] Babeltech!? I'm sorry, but that is a terrible reference. For example, if we had this discussion on a reputable website -- such as Anandtech, The Tech Report, TechPowerUp, or arstechnica -- no one would ever bring up ABT in a GPU discussion. Everyone knows that they have a large Nvidia bias.
Your ABT review contradicts the data provided by most reliable websites.
Can you provide a link that from a reputable website to show that the 680 is not fast enough to utilise 4GB VRAM?
Sure:
https://www.techpowerup.com/reviews/Point_Of_View/GeForce_GTX_680_TGT_Ultra_4_GB/28.html
The performance gains of this 4GB 680 model (at most ~5% at 2560x1600 -- a resolution that the GTX 680 can't provide playable framerates at, even with 4GB of RAM) come from overclocking the GPU, not from the extra RAM. The same performance gains were simulated in the original reviews of the GTX 680 from reliable websites such as Anandtech and TR.
Again, buy considering the 7970 is fast enough to utilise 3GB or VRAM. and even though my 7950 can be 3GB VRAM limiting today. I don't think the 680 can never utilise that 4GB VRAM because of it's speed. And even 4GB is too much, at least it can utilise the first 3GB of VRAM like the 7970 does, and make it able to perform (like the comparison back in 2012).
Even the extra RAM on the 7970 is
generally unhelpful. For example, the Tonga GPU (which is fitted with either 2GB or 4GB) which was basically a replacement for the Tahiti GPU in the 7970/280X performs almost identically to Tahiti. Most benefits come from bandwidth increases (memory speed increases) or GPU clock speed increases.
I asked for a reliable website or a class action lawsuit -- That LTT post links data to a single French online PC component store. Hardly reliable. Furthermore, you didn't mention these points from the LTT post:
"
We have to add that these statistics are limited to products sold by this e-vendor, and returns done specifically to said vendor, which is not always the case because people will sometimes return the product to the manufacturer, however this is a minority of the cases."
"
The reported failure rates concern products sold between April 1st, 2012 and October 1st 2012, for returns created before April 2013"
"
*Please note that obviously not all brands of particular components are noted either because of retailer availability, regional availability or sample sizes that are too small for this large French e-vendor*"
So we basically have a limited timeframe, with limited sample sizes, and limited model selection too. Hm...
Furthermore:
"
Certain numbers are very strongly impacted by certain models, which is the case with the 7870s by Sapphire for example. With the 7970, if we exclude the problematic Sapphire model, we get 5.47%[...].
the rate of failure for 7870 lowers considerably, although it's still abnormally high, with sapphire cards still having the problems. In general, we see that GeForce models are more reliable according to this data, notably with an excellent ROF for the GTX 660."
Reading the original post, most AMD cards sold were Sapphire custom board overclocked models. But what about the OEM reference board 7970s and 7870s? Where are other prominent AMD IHVs such as HIS? Were the GeForce models Nvidia reference board designs?
Therefore, the failure rate of 7970 is 172% higher than 680. It's no doubt "significantly higher".
The worst 7970 is the Sapphire Radeon HD 7970 OC Edition 3 GB, 14.29% fail.
The worst 680 is the ASUS GTX680-DC2O-2GD5 2 GB, 6.98% fail.
So, The average failure rate of the 7970 is higher than the worst 680.
And the worst 7970 has failure rate that 105% higher than the worst 680.
As discussed by the points above. This is dubious accounting, and a great example of the phrase: there are lies, damn lies, and then there are statistics.
The 7970 failure rate (once excluding Sapphire's questionable custom overclock model) falls to 5.47% -- and even that statistic is questionable because it may have a small sample size.
So, actually, it's what, 5.47% for the 7970 verses 6.98% for the Asus GTX 680? Also, if we ignore the ASUS GTX 680 model and look at other GTX 680s, the GTX 680 failure rate is just 2.66%.
5.47% (7970) - 2.66% (GTX 680) = 2.81% difference
Hardly statistically significant. Furthermore, we don't even know sample sizes.
Since the performance difference (between 680 2G and 7970 3G) getting bigger and bigger when resolution increase. Also occasionally a sharp drop. I will say it's more point to hardware limitation (e.g. VRAM), but not software maturity.
This has been proven incorrect many times by the aforementioned websites. Kepler is simply a rubbish GPU design for modern games. It was great for the DX10/DX11 generation, but it's horrible with DX12/Vulkan. For example, the GTX 780 Ti with 3GB has aged very poorly, and sometimes struggles to match Tahiti (with 3GB), a significantly smaller and cheaper GPU design.
Finally, as I said, the RAM myth was debunked when looking at Tonga card with only 2GB of RAM.
I'll leave with these performance summary charts from techpowerup:
https://www.techpowerup.com/reviews/AMD/RX_480/24.html
As you can see GK104 (GTX 680/GTX 770) and GK110 (GTX Titan/780 series) did not age well -- RAM had very little to do with it.
