Am I correct in saying that Nvidia GPUs are bandwidth starved but not AMD GPUs due to infinity cache?
No, you are not right, not in general.
Every GPU benefits from more memory bandwidth.
Only to a certain point. A very low tier GPU will be saturated with very little bandwith.
One thing to note: looking at
hwBot, an overclocked 4080 can get over 10000 points, which means most of the preformance increase came from the core oc and not from the memory replacement.
Exactly.
No, I mean literally all GPU architectures, memory bandwidth is always a constraint on performance, some architectures are more efficient at utilizing available memory bandwidth than others.
Again, only to a certain point, depending on GPU configuration. But yes, some architectures are more efficient at utilizing memory bandwith than others, and this has interchanged between both manufacturers in the past.
Is it? The 3GHz on the screenshot should be ~10% faster clock than a 4080 which would make up a very large part of the shader deficit. This at the same time with the main headline of memory bandwidth increase of 15-25% and I would be rather disappointed if it did not beat a stock 4080
But that show that the 4070TiS benefits from more bandwith to a certain point, otherwise it couldn't beat the stock 4080.
The whole comparison is dumb of course, because any 4080 will reach the same or better clocks and still have a lot more shaders. It only shows the potential of the 4070TiS.
The use of increased memory bandwith coresponds to fill rates etc. of the particular SKU. They always have to be in the right balance. Some SKUs are bandwith-starved and every percent of increased bandwith will lead to increases performance 1:1, but increased core clock will be barely notable. Others are oversaturated with bandwith an will instead produce better performance equal to increased core clock.
I got the impression that the original 4070Ti was slightly bandwith starved, but the 4070TiS with only 10% more cores, but 33% more bandwith should be oversaturated, which explains the performance gain only slightly above core clock increase even though the bandwith is massively increased.
Mods like these are very interesting from a technical perspective, but mostly useless in real world use. If you could double the memory of 4070/4070S/4070Ti for a sensible amount of money and use faster memory while you're at it, that would be useful. But Since they already use the biggest memory-ICs available, that is not possible.