DegustatoR
Legend
A new chip on a new process have a new data interconnect. News at 11.
- Status
Another simple explanation would be that they aren't gonna repeat the 3080 situation again, having an x80 product that is a cut down top end GPU, and pricing it at a 'reasonable' $700, which was overall pretty decent value for the performance at the time.
So even if AD102 is a bit smaller than GA102, they still might not want to offer a cut down version for $700 as a 4080 or something, given the higher costs of 5nm and inflation and whatnot. I'd guess they dont want to increase the x80 tier product pricing above $750 or so either, for perceptual reasons. So they could sell AD102 flagship at like $2000, cut down version at maybe $1400, then AD103 comes in to fill the 4080Ti at $1000 and 4080 tier at like $700 again. This might feel cynical to some on-paper, but if the performance improvements are still big with AD103 versus GA102(30%+), then I think they'll get away with it. AD102 will truly be like all new tier of GPU as a result.
Not a great situation for us, but it makes plenty of sense from Nvidia's perspective.
DegustatoR
Legend
I'd argue that we're firmly NOT in such an era, with most gains coming from improving efficiency of the h/w we have and adding new h/w into new GPUs. "Adding more units" was somewhat true around Maxwell days, not anymore.
Rootax
Veteran
Ampere feel like a turing with more fp32 to me. The new rt units or whatever doesn't seems quicker in the end for exemple. Main gains => more fp32...
Now, rdna1 to 2, I agree, and my bad I had only nVidia in mind when I posted
TopSpoiler
Regular
Looks like Kopite got some interesting information from the source.
Very curious what is the "simple" solution to achieve 2x performance of 3090 with maintaining VRAM bandwidth.
arandomguy
Veteran
One of the "simple" things is going to be cache.
Let's avoid arguing the fine details of whether they match or not but you can say the 6900 XT has roughly the same performance as the RTX 3090. So how much VRAM bandwidth do you need to double 6900 XT performance? 384 bits at 21 Gbps is 83% more bandwidth than 256 bits at 16 Gbps.
TopSpoiler
Regular
Yeah. Perhaps he said "simple" to mean "easy".
DegustatoR
Legend
Remains to be seen 2X performance in what exactly.
TopSpoiler
Regular
TopSpoiler
Regular
Easy 2x perf of the 3090 with no additional power.
DegustatoR
Legend
Well it is a 2-3 node generations change.
Easy 2x perf of the 3090 with no additional power.
His series of tweets implies though that it’s the 4080 that is being tested and it’s the 4080 that is 2x the performance of the 3090 at the same power. So power consumption still creeping up…
DegustatoR
Legend
AFAIR his tweets it's 3070 which is supposedly at 300W so it being 2X of 3090 would mean that it's "easy" to do without consuming more power.
He's since said this specifically is 'rasterized'(non-RT) performance.
Still find it hard to believe. Maybe that they could do 2x performance of a 3090, but not 'easily'.
It's about a 1.5x node jump, much like Maxwell to Pascal was, which went from 28nm planar to 16nm FinFET. That was about a 75% lift in performance, including the fact that Pascal tended to boost stronger out-the-box and thus had less overhead stock. Though I guess you could say GP102 was a somewhat smaller flagship GPU, too.
I dont know. I'm just starting to lend less credence to these guys' constant rumors anyways. For all that they may have gotten right the last generation, they've been absolutely all over the place the last year about the new generation and the claims seem to keep changing and contradicting previous claims and whatnot. Getting hard to take it seriously.
TopSpoiler
Regular
Seems the "Easy 2x" is 4090.
- Status
Similar threads
