You empuzzled me. How does it even work? Game engines differ, various resolutions need different and not linearly changing amounts of VRAM bandwidth and on-die calculating power, plus God knows what else; all this makes 20 Gbps a massive overkill in some scenarios and not even half enough in others. I do agree it's not completely tragic to only have G6@20 with N48 in mind but there definitely will be scenarios when it's anemic. 9070 non-XT feels a lot more balanced (yet a lot worse per $).
I assume using G7 is way too expensive even if we're talking chips that can't even reasonably achieve 28 Gbps and this is why we don't see 9070 XT with, say, G7@26 coming.
Unfortunately no 160+ CU beast to witness. I'm sick of top tier GPU only being green and of AMD never even trying.
It's about saturation points. Limitations in what respect or another. I don't get what you don't understand?
GDDR7 truly isn't that expensive, but the reality is it's just not needed yet. It will be next gen. Right now it is largely used to sell as an advantage, more than actually being one.
As shown with N48, there are ways to optimize designs (towards current gaming trends) without it, as I explained. This comes with matching compute ability, cache, and memory bandwidth accordingly.
Buffer size is important to an extent because of current trends in relation to according compute capability and standardized resolution/settings (especially as new ones emerge, such as up-scaling/RT/FG).
It changes, and is fluid, but there are general guidelines that correspond to products. I have outlined this before. It is why a 12GB nVIDA card does not hit >45TF, and why 7800xt was limited this way, etc.
It is also how a product like 9070 xt, which is just over 45TF, and 9070 vanilla likely limited in this regard. Above allows a similar market until it generally limited by buffer, as in the case of 5080.
This is why you could literally see a 3500mhz N48 competing with 5080, even though a 5080 is capable of much higher compute throughput.
5080 can run at faster clock, but general playable settings (ie 60fps mins) are unattainable do to inadequate buffer. Why do you think nVIDIA clocked it at 2640mhz, and not closer to 3154 (like AMD)?
The amusing part is for 5080 to make sense it needs 24GB of memory (really 18-20, but 24GB is the only option) at 3.23ghz, which is the *exact* top of the 5nm dense voltage curve (seen on
A15/M1).
If nVIDIA productizes this I will be amazed, because that is what they will sell next-gen and claim it is faster than a 5080...100% without doubt. They *could* do this, but likely won't unless they HAVE TO.
This is how you make safeguards within your product stack/chip designs, but don't push things forward too quickly if you don't have to (and hence save money on each inch of progress).
You will also notice 3 out of 10 5080 16GB skus reviewed on this website are capable of that for any prolonged period of time. This is by design.
The reason the RTX 4090 outperforms the 5080 is because its core is (sometimes significantly) more powerful, not because the 5080 is memory capacity starved. To run into the limitations of 16 GB, you currently have to go all-out, with the most extreme scenarios (and it would still fit into memory by a hair) - not to mention W1zz tested this on the 5090, where this would be about 49.5% of its capacity. On a 16 GB card it would preallocate less, and use a bit less as a result.
We'll see how it bears out. 4090 outperforms it for these reasons, yes, but minimums (and often benchmarks) are largely correlated to limitations.
I'm just sayin' a 9216sp part, with 3780mhz core and 36gbps ram is a much more rational pairing of resources. Again, we shall see!
As mentioned about a trillion times, I do not agree with how W1zard perceives memory usage, not how he tests it's limitations. There are many variables in this beyond purely benchmarks, such as swapping.
What happens when a game is targeting 16GB, as you say, and you turn on FG? nVIDIA doesn't want you to know and hence hides native FR when you do this. This what I'm talking about; many variables.
