Or has it already happened in anticipation?
Prices have already been dropping steadily lately. You might get a bit more out of them as they try to clear remaining Ampere stock come Black Friday/Cyber Monday time.
There's a competition running, where you can win an RTX 4090 with Cyberpunk 2077 backplate.
https://twitter.com/CyberpunkGame
DYOR, check T&Cs, etc.
GN had multiple games where the 4090 was twice the performance of the 3090 master
What where the other games? It had F1 2022 a few times, as it was used for RT, etc testing as well, where yeah, it was 100% again, but since it already was, that's hardly surprising.
GN also highlighted a number of CPU limited results, just like HuB, but almost none of that applied at 4K and as such none where 2x performance in HuB testing.
But I guess if all it takes is just one game to crack 2x in order to say that the 4090 is twice as fast as a 3090, then sure, I guess it is. I'm just saying that for most, it is less, around 50-80% faster, which while impressive, given the node shrink, and the fact the 4090 has around 2.5x as many transistors as the 3090 Ti, then the 4090 darn well should be nothing less then 50% faster.
Port Royal leaderboard:
https://www.3dmark.com/hall-
The top 3 are now RTX 4090 cards.
And are the top 5 in Time Spy graphics:
https://www.3dmark.com/hall-
And are the top 6 in Time Spy Extreme graphics:
https://www.3dmark.com/hall-
People have posted RTX 4090 scores in some of the others as well, such as Fire Strike Ultra, etc.
Might have been mentioned, but when does the 4090 go on sale in Australia?
Is it going to be like the 3080 release where all websites list their products at a specific time?
Assuming no founders again for us aussies?
What where the other games?
GN's 4k results were:
Game 3090 4090 Increase %
Total war Warhammer 68.9 124.9 81.27721335
F1 2022 115.5 230 99.13419913
Shadow of the tomb raider 105 201.3 91.71428571
Rainbow six siege 203.2 355.4 74.9015748
Horizon zero dawn 91.8 164.4 79.08496732
Final fantasy xiv 106.8 214 100.3745318
Gta V 120 185.5 Engine capped
Control RT 41 76.1 85.6097561
F1 22 RT 40.6 83.6 105.91133
Cyberpunk 2077 rt DLSS2 38.2 79.1 107.0680628
Tomb raider rt 59.6 118.5 98.82550336
GN also highlighted a number of CPU limited results
GN were also CPU bottlenecked (to a lesser extent) because they used a 12700KF for testing.
one game to crack 2x in order to say that the 4090 is twice as fast as a 3090,
Nvidia didn't make any claims for 3090 vs 4090 IAFAIK. It's clearly more than 1 game, with even more close, even comparing a stock 4090 vs an 3090 OC model.
I'm just saying that for most, it is less, around 50-80% faster
It's 79%+ on every game bar 1, which was probably CPU bottlenecking. At resolutions below 4k, it was clearly CPU bottlenecking on most games, so you will see less, but that doesn't tell you much about the GPU performance.
the fact the 4090 has around 2.5x as many transistors as the 3090 Ti,
A significant chunk of that will be the cache increase.
And it's all packed into a smaller chip, thanks to the 4N process.
(AMD will be using a mixture of 5nm and 6nm, for chiplets and controller, respectively.)
A significant chunk of that will be the cache increase.
100% this, which I think a lot of people are overlooking.
A significant chunk of that will be the cache increase.
I think a fair bit of the rest is likely the RT cores, as shown by the results when RT is turned on in those games. The Blender results somewhat confirm that as well.
Still, once you get to the point of pushing 120+ fps at 4k in pure raster performance, one may was well put all the rest of the silicon into other stuff and since its Nvidia, then of course they double down on RT/AI, etc.
There is also a higher ratio of RT Cores and Tensor Cores, as well as larger caches.
Those Tensor Cores now handle many more modes than originally. Also, the RT Cores are running some extra functionality for this generation, to more quickly ditch work it doesn't need to complete.
P.S.
https://www.techpowerup
Shading Units
16384
TMUs
512
ROPs
176
SM Count
128
Tensor Cores
512
RT Cores
128
L1 Cache
128 KB (per SM)
L2 Cache
72 MB
Pixel Rate
443.5 GPixel/s
Texture Rate
1,290 GTexel/s
FP16 (half) performance
82.58 TFLOPS (1:1)
FP32 (float) performance
82.58 TFLOPS
608mm2, down from 628mm2 for 3090/Ti.
I think a fair bit of the rest is likely the RT cores
It's also going to be cache. Things like on-die cache don't scale down very well with node shrinks, which is why slapping large amounts of cache onto the die itself is ultimately a tradeoff against adding more functional transistors.
It's a driving factor behind why AMD are pushing stacked cache with their X3D CPUs. I mean, just look at how much space is occupied by cache on AD102. AMD noted the "huge area cost" of adding more cache back in the beginning of 2021. Even AD103 has a comparatively large (in terms of die space) amount of cache.
100% this, which I think a lot of people are overlooking.
cache is SRAM, which takes up a fair bit of area and transistor budget, however it's also easier to structure to be defect tolerant, so it doesn't affect yield as bad as logic structures.
cache is SRAM, which takes up a fair bit of area and transistor budget, however it's also easier to structure to be defect tolerant, so it doesn't affect yield as bad as logic structures.
Yep. I'd be interested to see some measurements regarding cache hits on Ada Lovelace GPUs, to get an idea of just how much heavy lifting the extra cache is doing for performance. I know with the "Infinity Cache" on RDNA3 it's something like 50%+ at 1440p and ~80%+ at 4k so I'm curious what Ada's would be.
You could do your own checking on your favourite game or benchmark, etc, with Nsight profiling software.
Update:
A video here to give you an idea of Nsight profiling:
Seen all benchmarking results!!! 4k gaming performance justifies the $4k cost. But due to positive reviews, i’m afraid that ordering battle will be brutal. Bots, scalpers, fomo guys, and for average joe bloe like me.
What time today can we order the bloody thing???
Tweet regarding OctaneBench being 2x a 3090:
https://twitter.com/OTOY/status/1572268322481405954
And benchmark graph:
https://twitter.com/OTOY/status/1579884699983306753
wow! Crazy to finally hit 1300 #OctaneBench on one GPU!! @nvidia #4090 FE … we can also expect 5-15% more speed with final optimizations for Ada/hopper, and even more RT perf with Brigade mode
The graph shows (in part):
RTX 4090 FE scored 1299.7
RTX 3090TI OC scored 706.1
RTX 3080 Gaming X Trio 10GB scored 541.7
RTX 2080 Super scored 265.6
rasterization increase of 40 – 70% @ 4K over a 3090TI
That is amazing and along with the massive increases in RT performance this is probably the biggest increase in performance of a flagship card since the 1080 ti. Pretty much this is one of first cards that can run mostly in 4k across all games
Interestingly GN said a couple of positive things about the Nvidia founders edition, thermals are very good and they don't know if AIBs can compete and power draw of about 400-500 watts is normal now but there is no major power spikes, over clocking pushes power draw to 666.66 watts :p
https://www.techpowerup.com
There is not much doubt the performance of the 4090 is impressive, beyond my expectations, but there is a degree of brute forcing it to get there which is a little less impressive,