E-Cycles - Faster CUDA rendering and better AI denoising

commercial
(mathieu) #849

Welcome on board :slight_smile: There is a nano turorial for the new presets until the new real one is adapted to the april update:

I recommend using the beta to directly get used to the new UI and it’s also much faster when using the new parameter :wink:

1 Like
(mib2berlin) #850

Hi, work fine so far and thank you for the builds.
I have some problems with own files to get more than 30% speed and if I cant fix it I will you send PM with file(s).
Can you add your logo in the builds?
Sometimes I have own build with OIDN, Master and E-Cycles running and during comparing I have problems to see on wich build I am.

Cheers, mib

(mathieu) #851

I will add a custom logo. The linux builds have the windows cubins, so that may explain the smaller win on Linux as the buildbots are already 10% faster on Linux. However, 30% sounds very low. As you have both, can you please test on windows first to see if the problem comes from the Linux builds? I will try to build the cubins on Linux to see if it helps.

(mathieu) #853

It’s a beta, I’m open to discussion about design. I can let the spp outside of the presets. But all my beta testers were happy to quickly have all parameters set for speed or quality without understanding and tweaking 20 sliders. You can also add your own presets and modify existing ones.

1 Like
(mathieu) #854

Regarding the reported bug, I’m still investigating, but at least using an HDRi from HDRi haven (as I don’t have pro lightning skies) on RTX cards and even using the 2.8 april beta with scrambling distance, it looks the about the same in buildbots and E-Cycles:
master:


E-cycles:

E-cycles even has a firefly less.

3 Likes
#855

the only thing i dont like about E-cycles is that i haven’t used it earlier

5 Likes
(DDB) #856
Classroom

Official Master 2.8 01-04-19 16x16
1x 2080ti 1:34.82 289M = 1x
2x 2080ti 0:49.80 289M = 1x
= 1,90 Dual Gpu Ratio

E_cycles_2.8_v20190402_win_multi_gpu
1x 2080ti 1:00.11 298M = 1.57x faster than 2.8 Master
2x 2080ti 0:33.70 298M = 1,47x faster than 2.8 Master
= 1,81 Dual Gpu Ratio

2.79 27_beta_april_feature_update
1x 2080ti 0:44.88 288M = 2,11x faster than 2.8 master
2x 2080ti 0:24.98 288M = 1,99x faster than 2.8 master
= 1,79 Dual Gpu Ratio

good work, can’t wait to see thoses result in 2.8
and with RTX its going to be a massacre

3 Likes
(rbx775) #857

Thx for the times!
Since you have 2 2080Ti’s: Have you tried connecting them via NVLink to join the Vram ?
(2x 11GB => 1x 22GB) Is this even possible from the get go?

(DDB) #858

its possible but it need a programmer to code that.
i dont think that this functionnality will ever be implemented because its too niche.

(TiborNyers_BC) #859
2 Likes
(mathieu) #860

Transforming a 980Ti into a 2080Ti for 42€ is possible :smiley:

Thanks a lot for the very serious test and the full logs. Very precious for in depth analysis on different architectures. It definitely helped to make E-Cycles even better :slight_smile:

2 Likes
(mathieu) #861

Actually, to be correct, the 980Ti is even 16% faster than the 2080Ti.

Some of you asked for a video of a render session. As there was a lot of post today, I add it here again. Real render time without screen recording and compression is around 8sec per frame with one GPU only. It uses the preprocessing option of 2.79x.

2.8x with 50% off is still available for some hours with this coupon https://gum.co/tsTuV/gumroadday, 2.79x with this coupon https://gum.co/SeYhy/gumroadday

and if you want the 2x speedup and the AI denoising node on top of a branch of your choice like Mantaflow or Fracture modifier branch, you can learn to make your own Blender here https://gum.co/ecycle/gumroadday (with 30% off today)

1 Like
(DDB) #862
EVERMOTION 09 BENCHES

Official Master 2.8 01-04-19 16x16
2x 2080ti 3:35.47 = 1x

E_cycles_2.8_v20190402_win_multi_gpu
2x 2080ti 2:08.78 = 1,67x faster than 2.8 master

2.79 27_beta_april_feature_update
2x 2080ti 1:56.36 = 1,85x faster than 2.8 master

carreful with interpreting this benchmark, e-cycles scale a bit less than 2.8 master, two gpu for my case in master is +1.9x power, e-cycles is 10% less, so for a single gpu the result would be better up there

the april update is still really good !
still in beta and 2.79 only tho

(mathieu) #863

The beta is also available for 2.8x.

(DDB) #864

oooooooooooooooooh didnt see, i will update every benchmark with the added result of 2.8 april beta

(Frank) #865

Bought a copy of E-Cycles this morning to show some support. :slight_smile:

2 Likes
(mathieu) #866

Very appreciated :slight_smile:

(TiborNyers_BC) #867

Glad you find it useful! The big benchmark using the Evermotion scenes is still coming! :slight_smile:

1 Like
(rafbr) #868

Hum I am not seen any speed up here. All I see is the great Intel denoiser in action.With exactly same settings it is slower than Cycles. And to me Cycles have less firefly. Or am I missing something?

(mathieu) #869

I would need the number and model of GPUs, and CPU if you activated one as well as the exact name of the build you used. I already hear problems of people having D-Noise activated in parallel. To avoid conflicts, it’s always good to try with factory default, then only activate the new AI denoiser addon and the GPUs without CPU.