ghostship2 posted at 8:55AM Sun, 27 October 2019 - #4368229

here is a link to that standard test scene

http://www.sharecg.com/v/86389/browse/11/Poser/Superfly-render-test-scene

Thanks!

If you've got a 1080ti, you're hurting performance by using a bucket size of 256. Depending on the scene, 10XX cards work a lot better at 512 - 1024.

seachnasaigh posted at 2:14AM Tue, 29 October 2019 - #4368143

For CPU rendering, networked rendering would sure speed things up. Distributing samples for Superfly, and buckets for Firefly.

The Vue renderer and Lux distribute samples/tiles over a network.

Poser Pro has runtimes for additional hosts and shares processing across the runtimes.

I'm in pretty much the same situation as mr_phoenyxx myself...

I run two machines-

An FX8230 with a GTX1060 6gb and 32gb (4x8gb) main ram that I call Sarah

A Ryzen7 2700 with a GTX 1080ti 11gb and 32gb (2x16gb) main ram which I call Amber.

Amber was built to replace Sarah and her lengthy render times... (oh the irony)

Both machines are running Windows 7x64 Ultimate Sp1.

Sarah is running Poser 11.2.276. Amber is on 11.2.307

Both machines have the graphic card in the first PCIe x16 slot.

Just going by the numbers, Sarah's 1280 CUDA cores shouldnt be a challenge for Amber's 3584.

I've just downloaded the test scene that Ghostship2 has linked, here are my GPU superfly render times -

Sarah : 2131 seconds ( 35.5 minutes) using 1403mb rendering memory on device GeForce GTX 1060 6GB

Amber : 7140 seconds (119.0 minutes) using 1403mb rendering memory on device GeForce GTX 1080 Ti

There are so many free hosting sites that do not require user accounts, why post share files on a site only logged in users can access?

GWild99 posted at 2:37PM Sun, 03 November 2019 - #4369080

There are so many free hosting sites that do not require user accounts, why post share files on a site only logged in users can access?

Most Poser users probably have ShareCG accounts. Great place to get and distribute freebies. That test scene was posted three years ago.

FWIW, I tried the test scene back in June, and again today. The results were the same. 20 minutes.

12 core processor, 64 Gb of RAM, Nvidia GeForce GTX 1080, Windows 10.

Running Poser 11.2.272

@Graytail If you can, put both of your video cards in your new machine and you can run two render tiles at one time. That should cut render times for you. I have a Ryzen5 1600x and my cards are a GTX970 and GTX980. I unchecked progressive refinement and set my tile size to 256. My render time with both cards working was 976 seconds. With this setup I can also opt to just use the card that is not running my screen and I can do anything I want on my computer while it renders including watching a DVD or youtube videos or playing a video game.

Follow up, I did a little tinkering and slipped Sarah's gtx 1060 into Amber, then did two more test renders. Updating those performance numbers with the same test render-

Sarah : 2131 seconds ( 35.5 minutes) on GTX 1060 6GB

Amber : 7140 seconds (119.0 minutes) on GTX 1080 Ti

Amber rendering on 1060 only : 1893.41 seconds, 31 1/2 minutes

Amber rendering on both 1060 and 1080ti : 796.95 seconds, 13 1/4 minutes

I dont know what it is about the 1080ti on its own, but the 1060 seems to be fixing it. One thing to note is that using both cards did use double the ram, whether that is the ram on the cards, or system ram I dont know, but single card rendering used 1403 mb and using both went up to 2906mb

Graytail posted at 7:34AM Mon, 04 November 2019 - #4369133

Follow up, I did a little tinkering and slipped Sarah's gtx 1060 into Amber, then did two more test renders. Updating those performance numbers with the same test render-

Sarah : 2131 seconds ( 35.5 minutes) on GTX 1060 6GB

Amber : 7140 seconds (119.0 minutes) on GTX 1080 Ti

Amber rendering on 1060 only : 1893.41 seconds, 31 1/2 minutes

Amber rendering on both 1060 and 1080ti : 796.95 seconds, 13 1/4 minutes

I dont know what it is about the 1080ti on its own, but the 1060 seems to be fixing it. One thing to note is that using both cards did use double the ram, whether that is the ram on the cards, or system ram I dont know, but single card rendering used 1403 mb and using both went up to 2906mb

Yeah, my ram usage was around 2880mb with both cards running.

ghostship2 posted at 2:57PM Mon, 04 November 2019 - #4369144

Yeah, my ram usage was around 2880mb with both cards running.

That's why if you are using multiple cards you are limited to the minimum ram of either card rather than the combined.

@graytail, could you re-run the test just using the 1080 but with the other card still in situ?

Edit: Sorry, just re-read and realised you did exactly that!

@GWild99 - it was posted 3 years ago at a readily available free site. Anyway, I've now uploaded it to the Renderosity freestuff, pending approval.

GWild99 posted at 7:20PM Mon, 04 November 2019 - #4368580

seachnasaigh posted at 2:14AM Tue, 29 October 2019 - #4368143

For CPU rendering, networked rendering would sure speed things up. Distributing samples for Superfly, and buckets for Firefly.

The Vue renderer and Lux distribute samples/tiles over a network.

Poser Pro has runtimes for additional hosts and shares processing across the runtimes.

I'm not clear on what you mean; I do use Queue to distribute animation frames among my remotes (aka "render slaves") and to distribute whole renders from a batch list. What I'm yearning for is the ability to use my network to cooperate on a ~single~ render.

Imagine a wallpaper project of a complex scene; your low-quality quickie test renders took a long time, and so you know that a high quality render at full size is going to take several days. If you could spread that work across all of the machines on your network, it would finish much more quickly.

Just as Firefly/Superfly now distribute a bucket (or sample, for progressive) to each core of your machine's processor, imagine if Queue were able to distribute a bucket/sample to each core of each machine in your network.

This is what Lux and Vue's HyperVue do.

I've been trying out the 11.3 beta today and I've come across the same issue with the 1080ti that others have found. As per Gwild99's post - I just tried the default La Femme scene with the suggested settings.

1080ti: 20.12 sec (twice as long as Gwild99's 1080)

2080ti: 5.00 sec

2080ti+1080ti: 3.64 sec

With the test scene mentioned above (https://www.renderosity.com/mod/freestuff/benchmark-test-scene-for-superfly-/83640), using the settings as loaded, I got:

GTX 1080Ti: 6600.24 Seconds

RTX 2080Ti: 622.54 Seconds

GTX 1080Ti+RTX1080Ti: 373.8 Seconds

I also tried a render with one of the new included scenes - AJ Futuristic Base 5960X 151.41 sec 1080Ti 208.41 sec 2080Ti 14.70 sec 1080TI+2080Ti 10.73 sec

So as others have found, running a 1080ti with another card seems to get it back up to expected performance, but alone it is worse than a cpu render.

I have to agree, slowed down here as well, I actually avoid using Superfly because of it.

Dim_Reaper posted at 11:59AM Wed, 19 February 2020 - #4380811

I've been trying out the 11.3 beta today and I've come across the same issue with the 1080ti that others have found. As per Gwild99's post - I just tried the default La Femme scene with the suggested settings.

1080ti: 20.12 sec (twice as long as Gwild99's 1080)

2080ti: 5.00 sec

2080ti+1080ti: 3.64 sec

With the test scene mentioned above (https://www.renderosity.com/mod/freestuff/benchmark-test-scene-for-superfly-/83640), using the settings as loaded, I got:

GTX 1080Ti: 6600.24 Seconds

RTX 2080Ti: 622.54 Seconds

GTX 1080Ti+RTX1080Ti: 373.8 Seconds

I also tried a render with one of the new included scenes - AJ Futuristic Base 5960X 151.41 sec 1080Ti 208.41 sec 2080Ti 14.70 sec 1080TI+2080Ti 10.73 sec

So as others have found, running a 1080ti with another card seems to get it back up to expected performance, but alone it is worse than a cpu render.

Oh Hum, looks like the 11.3 is not going to help me, not with my 1080ti anyway.

Hi guys

Can you try install older Nvidia drivers like 388.13 or so(if you are running Turing RTX or GTX series you can't use these drivers) , with these drivers I have good experience in rendering with "older" GPUs like GTX1080Ti or GTX1080

Can't comment on 11.3, I'm still on 11.2 and didn't tried new update for RTX, which I will be installing later on and see what difference it will makes like on GTX GPUs and RTX

What I tested few months back, older drivers do work better for GTX cards in rendering like in Octane or Cycles or E-Cycles, with newer drivers not sure why performance is not really there

I have run for long time 388.13 drivers till I got RTX 2080Ti with which I can't use these drivers sadly and performance tanked for me in Cycles or E-Cycles and few other renderers with latest Nvidia drivers branch

For me personally 388.13 have been best drivers for GTX1080Ti and GTX1080's and lastly can you use SIV64 and check GPU usage and do you have enabled progressive and how many buckets are you using?

For GTX1080 or GTX1080Ti try 384 or 512 as for bucket size,you can try download some of my scenes and test them in Poser, usually I have included like render settings and light sets which should help a bit and can you post yours times

Hope this helps

Thanks, Jura

First, thank you all for this discussion.

Second, thank you #ghostship2 for link to test scene. As I've just said there, it may settle how many of my twin GTX 750 Ti cards' cores to allocate to bucket. Some of the available info says 'Bucket = GPU core count'. So, I merrily set 1200 of my 2 @ 640, leaving a few for Windows UI, and 'progressive' renders are interesting...

Some of the info suggests setting 2^n which, here, would be 1024 rather than my 1200. Then I can compare / contrast pixel count, too...

FWIW, I'm currently a dozen hours into 'crazy' test-render of a simple park scene lit only by 'super ambience' from the street lights' glazing. It's currently doing ~10k samples per hour, and the delicate shadows are impressive...

Bucket size is about memory. With a GPU use the largest bucket size possible without running out of memory as only one bucket will render at a time (IIRC). With CPU as there will be one bucket rendering for each thread it's better to use a smaller bucket size. CUDA cores are about speed - the more there are, the faster the card will render.

my experiments tell me that CPU rendering likes 16x16 pixel tiles and GPU likes 256x256 pixel tiles.

Unexpected problem: A lot of the files called by that PZZ/Pz3 superfly test scene are scattered among PPro 11.3's legacy content zips, often with slightly different names.

I have a fairly lean core runtime, prefer to pull in what I need from 'external' libraries. After several hours hunting and cross-referencing 'not found' Amy hair OBJ and textures, Pauline's tennis shoe textures etc, I've had enough for tonight.

Some good news is that I've just tried running the test scene with the demo of Poser 12 and the 1080ti slowdown seems to be fixed. Also, the Optix mode for RTX cards is a LOT faster. Speeds seems to be faster across the board. The bad news is that at the present time, you can't use more than one card to render. Here are the results:

Downloaded Test Scene Poser 11.3 RTX 2080ti 622.54 Seconds

GTX 1080Ti 6600.24 Seconds

RTX 2080ti+GTX 1080Ti 373.8 Seconds

Downloaded Test Scene Poser 12 RTX 2080ti 475.13 seconds.

GTX 1080Ti 531.63 seconds

RTX 2080ti Optix 266.89 seconds.