Plus, like others mentioned, the power use is quite a bit higher.
I posted above a chart that shows the performance per watt. It is a Nvidia chart even if half true it is significant. If true it means you have 2x performance at 150w, and more above.
what about when running in idle?
does running a beefy bpu increase the bill when doing mundane task? 
anyone have an idea?
2 Likes
up to %70 faster in blender 2.93, they say in here https://blogs.nvidia.com/blog/2022/09/20/nvidia-studio-geforce-rtx-40-series/
1 Like
So you prefer to spend more energy?
The 4090 is rated at 450w over 350w of 3090, but the increase in speed is proportionally better than that.
Iām not a programmer, but I believe the shading is done on the CUDA core and other aspect take advantage of the RT core if Cycles support it, Nvidia support has been great so far.
I am also kind of hype about Omniverse + 4000 series support, I wonder if you can generate a .EXE from it, being able to create a racing game is kind of appealing.
If you want to create a 3d app you would use Unreal, Unity or another game engine. Omniverse will be a Nvidia āaquariumā. By that I mean it will be itās own world where Nvidia sets the rules and reaps the profits. A bit like a VR version of AOL. A āwalled gardenā.
Valid point, Bullit. The new cards will be faster and more energy efficient. More rays cast per watt.
The hardware will be pushed to the edge of breaking though, that is unless Nvidia gives you a way to put the card into a lower gear so you know for sure it will keep working years from now (a ādonāt break my hardwareā mode that does not involve booting into the UEFI).
In fact, Intel and AMD are really pushing their products to the edge too (meaning you pay more for chips with an increased fail rate). I wonder how long until the Geico Gecko starts yapping about insurance for computer hardware?
1 Like
You already get a ārecommendedā 50 a month Allstate insurance plan when you buy a GPU that you have to opt out of, I canāt believe thereās people out there who would actually buy that
As igorās Lab proven you could potentially unvervolt it to 350W and have only minor performance hit. Like 40% less power with 10% less perf.
Fastest GPU never had good perf/wat, cause every % of perfomance cost you few % of power.
Anyway, iāve composed table with TGP of RTX GPUs for ease comparsion:
| GPU | RTX 4090 | RTX 4080 16 GB | RTX 4080 12 GB | RTX 3090 Ti | RTX 3090 | RTX 3080 Ti | RTX 3080 | RTX 3070 | RTX 2080 Ti | RTX 2080 | GTX 1080 | GTX 1080 Ti |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Node | TSMC 4N | TSMC 4N | TSMC 4N | Samsung 8 nm | Samsung 8 nm | Samsung 8 nm | Samsung 8 nm | Samsung 8 nm | TSMC 12 nm | TSMC 12 nm | TSMC 16 nm | TSMC 16 nm |
| Base freq | 2230 MHz | 2210 MHz | 2310 MHz | 1560 MHz | 1395 MHz | 1365 MHz | 1440 MHz | 1500 MHz | 1350 MHz | 1515 MHz | 1607 MHz | 1481 MHz |
| GPU Boost | 2520 MHz | 2505 MHz | 2610 MHz | 1860 MHz | 1650 MHz | 1665 MHz | 1710 MHz | 1725 MHz | 1545 MHz | ? | ? | ? |
| VRAM transfer | 21 Gbps | 22.5 Gbps | 21 Gbps | 1000 GB/s | 937 GB/s | 912 GB/s | 760 GB/s | 448 GB/s | 616 GB/s | 448 GB/s | 320 GB/s | 484 GB/s |
| VRAM | 24 GB GDDR6X | 16 GB GDDR6X | 12 GB GDDR6X | 24 GB GDDR6X | 24 GB GDDR6X | 12 GB GDDR6X | 10 GB GDDR6X | 8 GB GDDR6 | 11 GB GDDR6 | 8 GB GDDR6 | 8 GB GDDR5X | 11 GB GDDR5X |
| TGP (base) | 450 W | 320 W | 285 W | 450 W | 350 W | 350 W | 320 W | 220 W | 250 W | 215 W | 180 W | 250 W |
| TGP (Max) | 660 W | 516 W | 366 W | ? | ? | ? | ? | ? | ? | ? | ? | ? |
Its hard to compare just TGP cause now they introduced Max TGP, but besite that in this generation unlike in previous in same segment (so 1080ā2080ā3080ā4080) base TGP remained the same.
1 Like
Thanks we will get more with benchmarks hopefully including power consumption
I prefer to spend less, I have a 3070 that Iām perfectly happy with. I was hoping the 4000 series would have more VRAM, so in 10 years when I upgrade, I could get one of those, but I guess not hopefully at that point thereās a good professional GPU with a ton of VRAM that doesnāt require itās own generator. Even the 3090 takes more power than Iād like
3 Likes
You will not spend less watts. Since 3000 series appear to be as not as efficient it will take more time working to achieve same results spending more energy overall.
Now if you are happy with what you have and the kind of work you do donāt warrant another investment is your right to do the correct call for your situation.
I agree that VRAM situation with Nvidia continues to be disappointing, and i bet that laptops will continue to be a even bigger problem. That is one of the reasons to AMD, Intel. Apple, Qualcomm(?) need step up their game.
I am keeping an eye for HEVC 4:2:2 if there is now hardware decoding in 4xxx, Da Vinci Resolve seems to have got a boost with 4xxx too.
2 Likes
Interesting technology for the 4000 series: https://developer.nvidia.com/rtx/ray-tracing/micro-mesh
3 Likes
I believe the 4X increase was mostly achieved through the use of DLSS3 as it is listed at the bottom of all the performance slides.
Idle draw tends to be quite good. As for running mundane tasks, it will depend on how efficiently it manages its boost clocks.
Oh ! The " Opacity Micro-Map SDK" looks like it solves the problem of transparency cost in pathtracing ! 
Despite using specialized hardware such as the RT cores, the ray tracing pipeline still relies on CUDA cores and the CPU for a handful tasks, and here NVIDIA claims that SER contributes to a 3X ray tracing performance uplift (the performance contribution of CUDA cores). With traditional raster graphics, SER contributes a meaty 25% performance uplift. (ā¦) The Tensor cores deployed on Ada are functionally identical to the ones on the Hopper H100 Tensor Core HPC processor, featuring the new FP8 Transformer Engine, which delivers up to 5X the AI inference performance over the previous generation Ampere Tensor Core.
The third-generation RT Core being introduced with Ada offers twice the ray-triangle intersection performance over the āAmpereā RT core, and introduces two new hardware componentsāOpacity Micromap (OMM) Engine, and Displaced Micro-Mesh (DMM) Engine. OMM accelerates alpha textures often used for elements such as foliage, particles, and fences; while the DMM accelerates BVH build times by a stunning 10X.
Letās see how many of these performance uplifts will sneak into blender and which ones will be gaming only.
2 Likes
From what this test is showing the 3090 is idling at ~17-20w and the A6000 at ~11-13w, itās also showing GPU utilization/RAM usage/power consumption for things like viewport navigation and final rendering.
So you should expect the same for the 4000 series with their 2x power efficiency.
2 Likes
My laptop RTX3060 in performance mode idles at 13w. Changing tabs here in Blender Artists which is not heavy donāt change that value it is still 13.xxx watt changes only the .xxx - i have 3 browsers opened and about 100 pages overall. Changing to a digital store site tab it goes to 24w. Playing VLC HEVC 1440x1080 23.97fps do not add much a large part is still in 13-14w. A 4K60FPS video in Youtube goes up to 26w at start and reduces to 14-17w after.
It is actually overall less than in hybrid mode since the APU+RTX goes to 18w or so. That was surprise i discovered recently.
1 Like