We're taking, again, a very flawed instrument to have this discussion.
Well, yes. Predicting the future will always be flawed. That's why discussions are productive, though! We can at least try to get down to
why predicting this is difficult, if nothing else.
Far Cry 6 is a very recent example.
That was the point - about as up to date as you're going to get in terms of game development and thus VRAM usage in AAA titles.
What it shows, is that on cards at 4-5 years of age, the combination of lower VRAM and tighter bus is killing the otherwise still 'decent enough' core performance. Far Cry 6 is a mainstream title. What's more interesting is what type of games you'll be playing at a 4-5 year age on your GPU.
I've lived the practice of that, as I'm doing right now. And its very easy to distill what happens. The fact that I'm still fine with 8GB, is what's saving my 1080's performance and doesn't cause it to fall off sharply like the 1660ti does today (and note, has a mere 2,5 years in age!). I can play FC6 just fine and Cyberpunk happily takes 8GB and produces 50 FPS nonetheless, without meaningful IQ sacrifice. But in a relative sense - the 1660ti at 6GB and the 1080 at 8GB are quite well balanced in terms of core power. Now, fast forward to today, because in the absence of data of the future 4-5 years from now, we need to extrapolate what we have. We're looking at much faster GPUs endowed with the same 6GB (2060, on par with 1080 core perf) and with the same 8GB on a 3060 that is way faster on the core.
IMO you're still mixing a few factors here that make this comparison problematic though: That a 2.5-year-old midrange card (1660Ti) struggles with a brand new AAA title at 2160p? That's expected. Heck, that was expected when it was new. The 1660Ti has never been able to handle 2160p Ultra in any reasonable way. The 1080 supposedly was at the time it was new, but that's due to games then being less demanding and no faster cards existing - and most reviews back then still pointed out that even the 1080 Ti wasn't
really fast enough for 2160p60 Ultra.
So, while you're right that the 1660Ti underperforms, it underperforms in a scenario it has never been suited for, and where it would perform poorly no matter what. Sure, it would likely be around 30fps instead of 20. But at that point, the
only sensible approach (unless you're playing a game where framerate
really doesn't matter) is to turn down your settings or resolution. Thus, the point is moot. It was unsuited to the task before the VRAM bottleneck ever came into question.
I'm not sure what's in the way of logical thought processes here with people, but we're specifically talking about what GPU X or Y is going to be worth in terms of future proofing 4-5 years down the line. Its the exact question the OP is asking. And in thát situation, 8GB on a card with 3060ti performance is just bad balance. No matter the cache it has and how things changed in its architecture. No matter what arcane technology gets stacked on top in certain titles. The bottom line, stands. We had an 8GB card in 2016 and now we have an 8GB card in 2021-2022 with 40-50% higher core perf. It won't last, and it will be capped at the exact same IQ level my 2016 1080 is going to be, while it has lots of core oomph to spare, in any similar use case. At the same time, I'm seeing a very nice, rather well balanced dropoff on the older 8GB card, where as you say, most of the time, most games are forcing you to dial back regardless and stay within your VRAM cap. But I can still max textures at nearly no perf hit. All I need to do is kill some overexpensive post processing (Ultra > High) to keep my perf. A 3060ti, 4-5 years down the line, will be making far greater sacrifices on IQ, and likely not the ones you'd prefer.
There's a flaw in your logic here: just because the older 8GB cards aren't bottlenecked doesn't mean they're actively using 8GB - the next step down is 6GB, so all we know is >6GB <8GB. And VRAM usage across games on average creeps up relatively slowly. Over the past decade we've gone from 2GB at the high end to 8GB in the mid-range and 16GB at the high end (and from 1080p at the high end to 2160p, though lower resolution were still
very common in 2011-ish), with visible bottlenecks at 2160p showing up first at 4GB a couple of years back, then 6GB more recently. Expecting 8GB to be performance-breakingly insufficient within 4-5 years, at resolutions that these GPUs are reasonably equipped to handle? Yeah, no, I don't believe that. At 2160p Ultra, sure, but these GPUs can't handle that
now, and
definitely won't handle that in 4-5 years, however much VRAM they might have. At 1440p or 1080p? Even with high quality textures, 8GB is likely to be plenty for many years to come. In the FC6 test, the 4GB 5500 XT
might seem to be toeing the line at 1440p, as there's a noticeable relative performance loss compared to 1080p - but it's overall still small, likely indicative of occasional stuttering. And if 4GB is borderline enough for 1440p Ultra today, there's no way whatsoever 8GB will be insufficient in 4-5 years. VRAM needs have
never increased that rapidly, and there's no reason why they would start doing so now.
This mismatch between VRAM cap and much higher core perf serves an obvious purpose, one Nvidia is well known for. It hard caps VRAM so that the urge to upgrade will arise even for people who are just fine with somewhat lower FPS and good IQ alongside it. Its a form of planned obscolesence and while Turing pushed that button slightly, Ampere is ramming it like no other. The market is clearly asking for higher VRAM cap GPUs, consoles are clearly going for over 8GB, AMD has a stack with more across the board per tier and games are already looking for higher allocation than 8GB as it is. How many writings on the wall do you really need? Don't even believe my word for it. Believe Nvidia itself, when it produces double VRAM, same SKU products in the same Ampere stack - and note, even in the absence of working SLI... the last GPUs they doubled VRAM on were all with SLI fingers, where the double VRAM serves an obvious purpose.
Double VRAM SKUs is a marketing tactic, nothing more, nothing less. The 3060 has 12GB because 6GB would make it look bad compared to AMD's 8GB cards for spec-illiterate buyers. The 2060 12GB is a cash grab, period - if they cared about supply at all, they would just stuff the market with standard 2060s - the 6GB isn't holding it back at resolutions it's suited for (1440p and 1080p). But precisely because there's this prevalent idea that you
need 8GB or even more because a lot of games opportunistically allocate tons of VRAM if it exists, people get roped into these misconceptions about what makes a SKU worth buying. "The market" is not rational. And it is certainly not well informed. Thus, what the "market wants" is often a poor choice of guideline for making a good product.
I mean, look at the FC6 numbers for the 2060. The VRAM "use" reading is >7GB for every single resolution on the 3090, yet the 2060 shows no sign of underperforming whatsoever - it's exactly where you'd expect it. Heck, except for 2160p the 8GB 2070 has
worse performance regressions compared to the TPU test suite mean, even if those differences are marginal. It is of course possible that these average numbers are hiding stuttering and poor 1%/.1% performance, and I'd kind of expect that at 2160p given results with other 6GB cards in that test), but at 1440p and 1080p? Not a chance.
And, again, even if you have a 2160p monitor and "only" a 2060: insisting on playing on ultra is
dumb. Heck, most 2160p panels have sufficient pixel density to look decently sharp even at non-native resolutions, so 1440p (or 4:1 scaled 1080p) would be a much better choice on that monitor for most games if you insist on Ultra, though 2106p mid-high would likely work decently too. And you can of course again complain that this is a new and expensive card and so on, but again: it isn't a 2160p Ultra card. It never has been. Not when new, not now, not 4-5 years after launch. And that's fine.
IMO, a
huge part of what has changed here is not the imbalance between compute and VRAM, even if that has been on a gradual increase since the birth of GPUs more or less, but the fact that reviewers have started focusing on 2160p and that games all stream their assets aggressively and opportunistically. When we're talking about GPUs barely capable of 2160p Ultra today, the question of longevity at that setting level is already answered. That isn't what these GPUs are suited for, so the answer to the question of which would handle that best in the future would be like asking which human is best at breathing underwater - there might be a difference, but it definitely doesn't matter. Thus, the question
must be moved to lower settings levels and/or resolutions - that's a necessity when thinking of GPU longevity. And VRAM needs drop off dramatically then, with even 6GB clearly being sufficient today. Thus, it's reasonably safe to assume that the 8GB 3060 Ti will perform the best at reasonable settings and resolutions in 4-5 years, just as it does today. There will no doubt be outliers then too, of games with astronomical VRAM needs and dumb bugs, but ultimately, none of these GPUs are likely to be meaningfully VRAM bottlenecked before they are bottlenecked by their compute capabilities. Which was exactly what I said in my first post here, just without the 2000 words explaining the reasoning behind it.
