LW B747 Benchmark (LW 2020+)

[erikals]

I converted this scene for rendering in eCycles for fun. [GTX 1080]
50 seconds for a comparable result. Much better.

LightWave = much better reflections / realism
eCycles = way faster, less good reflections
noise, eCycles is better, but due to a denoise filter. hard to say exactly how much without more data.

further comments >
https://www.youtube.com/watch?v=ZyBfI_IeMgw

 

[Rayek]

LightWave = much better reflections / realism
eCycles = way faster, less good reflections
noise, eCycles is better, but due to a denoise filter. hard to say exactly how much without more data.

further comments >
https://www.youtube.com/watch?v=ZyBfI_IeMgw

Thanks for making that video! It is interesting!

I did not strive to replicate the materials exactly, of course. The floor material is different, for example, and reflections differ. The plane shader's reflections are different as well compared to the LW setup. I hardly spent any time on optimizing the materials for quality. It was merely a quick test to check the speed, not meant to reproduce the exact same quality, but rather to see how much faster GPU rendering would be in this particular case.

And don't forget that my GPU is a (almost) two generation old regular gtx1080 (not even a TI!). A RTX 2070 would probably render this same scene in less than half the 50 seconds in the RTX version of eCycles. Perhaps faster, even.

So please avoid making comparisons of quality: the materials could be improved much to resemble the LW version with a bit of extra effort: I am certain of that.

It's really late here now, and I should get a nap now ;-)
If you like, I can share the scene with you tomorrow. I'll see if I can up the quality of the materials and reflections, but I don't want to spend too much time on it.

If you render in vanilla Blender/Cycles it will be slower since eCycles is quite a bit faster. It's competitive with Octane in terms of speed. Interesting to mention as well: it seems that eCycles can make use of NVlinked GPU memory: if you install two GPUs each with 12GB memory, eCycles will make use of the full 24GB. Which means the memory bottleneck is no longer a real issue for most scenes, allowing for large scenes to be rendered on GPU.

Anyway, Lightwave's CPU renderer is really quite a zippy beast for a CPU based one. I think if the developers would implement a GPU mode that it would fly. Perhaps they will.

 

[erikals]

If you like, I can share the scene with you tomorrow. I'll see if I can up the quality of the materials and reflections, but I don't want to spend too much time on it.

Thanks, I know Bryphi would want to test it. maybe post link over at YT.
https://www.youtube.com/watch?v=ZyBfI_IeMgw

other that that I think we should end the discussion here, I guess we've made the point, and NT might be tired of the word in these threads.

never the less, i (we) thought it was important to address.

 

[Wireframex]

Hi LW Lovers, here are my benches :

Bench --> 3m 7s
Bench-Volum --> 10m 27s

CPU : 3990x
Mainboard : Gigabyte TRX40 DESIGNARE


https://ibb.co/LddHJg7

https://ibb.co/2nynFg2

 

[Rayek]

Posted the link of the updated improved version on YT for you and Bryphi. It now renders in 43 seconds (eCycles / 1min6secs in regular Cycles), and the reflections on the plane are similar to the LW version. Turns out I had way too much specular defined.

 

[Lewis]

Hey guys, can we keep it on topic ?

It is not cycles bench but LW engine bench. I have Octane version but i didn't post for a reason . GPUs are so much quicker even than threadrippers and even without Denoiser that is not comparable.

Thanks, so far i see 3990X TR is very close to 3 minutes, I wonder can it go below 3min with some overclock maybe.

 

[Rayek]

I'd be very interested in the Octane results as well. Perhaps a new thread? Anyway, chiming out. Thanks for the scene and play!

 

[erikals]

Hey guys, can we keep it on topic ?

yep.

sidenote to other users, LW2020.0 Trial renders this scene with lots (!) of noise, beware.

 

[Nicolas Jordan]

I ran this again today and got 6 min 53 sec at normal room temp was just slightly better vs my 7 min 6 sec I got on a hot day.

 

[gar26lw]

Threadripper 3970x | 256 GB Ram @ 3200 16 18 18 38 1T | 970 EVO Plus NVMe M.2 SSD

Bench : 4 min 47s

Volumetric : 16 min 9s

LightWave3D v2020.0.1

changed RAM config from 3200 to auto (2333) in bios and got

Bench : 4 min 34s

Used Ryzen Master in Auto OC and got :

Bench : 4 min 41s

any thoughts on that ?

 

[fishhead]

For what its worth:
LW_B747_Bench: 12 min. 11 Sec.
LW_B747_Bench_Volumetric: 42 min. 02 Sec.

my system:
Threadripper 1950x 3,4 GHz, 16 cores, 32GB RAM, MSI X399 SLI Plus, Win10 Enterprise (Homebuild system)

 

[Lewis]

Nice result fishead, also it clearly shows how good is AMD updates and IPC percentage that now-days TR 3xxx series is so much faster in only 3-4 years.

 

[ELinder]

2017 iMacPro 10 core 3 GHz Intel Xeon W, 64GB ram. Emailing and web forum reading while the volumetric version rendered was probably a bit of a performance hit.

LW_B747_Bench: 18min 39sec
LW_B747_Bench_Volumetric: 1hr 5min 53 sec.

 

[scallahan1]

My time is not bad but some of you dudes got some crazy fast systems...
18m 4s

Intel Core i7-6950X (10c/20t)
Evga GTX 1080 FTW
32gb G.Skill 3300 ram
Asus Strix X99 Gaming board
Cooler Master AIO liquid cooler

Steve
PS> Like Erik said 2020.0 is yucky but renders WAY faster. Weird.

 

[gar26lw]

yeah, I was thinking, no point posting unless faster

 

[scallahan1]

yeah, I was thinking, no point posting unless faster

LOL. I'm not even gonna attempt the Volumetric test cuz I don't want to see my electric bill.

It'd be cool for someone to tabulate the results into a bar graph or something. I know someone
already mentioned this idea. Love to see where we all stand. I'm cool not being at the top of the
chart.

Steve

 

[Lewis]

Yes, having a list with bench top results (2 columns one for regular one for volumetric) would be nice, who is volunteering to create google docs/table ?

 

[Cageman]

Tests done with LW2020.0.1...

LW_B747_Bench = 03m 07s (187.1s)
LW_B747_Bench_Volumetric = 10m 45s (645.1s)

Just for fun, I went through the different render tile-sizes for LW_B747_Bench

Render Tile Size 8 : 4m 36s (276.4s)
Render Tile Size 16 : 3m 43s (223.8s)
Render Tile Size 32 : 3m 21s (201.0s)
Render Tile Size 64 : 3m 09s (189.2s)
Render Tile Size 128 : 3m 07s (187.1s) [Benchmark Settings]
Render Tile Size 256 : 3m 07s (187.4s)
Render Tile Size 512 : 3m 07s (187.7s)
Render Tile Size 1024 : 3m 08s (188.2s)

As a last test, I will reset the Render Tile Size back to 128, but change Polygon Intersection Mode from Watertight to Fastest... then test both scenes again.

LW_B747_Bench = 02m 58s (178.0s)
LW_B747_Bench_Volumetric = 10m 17s (617.3s)

Both of them having no visual difference from Watertight vs Fastest setting. I would argue that Watertight as Default in LW is wrong. It seems to be a mode that you only need to use if your models aren't up to snuff, especially for interiors. In these two scenes, the setting adds a bit too much rendertime, especially if it would have been an animated sequence... saving 9s/frame for LW_B747_Bench scene or 28s/frame for the Volumetic version is important when rendering a sequence.

The computer I used to test this with, you can see in my signature.... however, that said... this computer is far from being finished in terms of memory and GFX card. I still use my "old" workstation for regular work. I use this computer to offload stuff towards (including screamernet for personal projects). That is why it has a very cheap Quadrocard... Remote Desktop and all that.

Having this fast of a CPU also means that I will get sloppy when trying to optimise things for rendering... that is why I still want to use my "old" workstation as well.

Thanks for the Benchmark scenes Lewis!

 

[Cageman]

Thanks, I know Bryphi would want to test it. maybe post link over at YT.
https://www.youtube.com/watch?v=ZyBfI_IeMgw

other that that I think we should end the discussion here, I guess we've made the point, and NT might be tired of the word in these threads.

never the less, i (we) thought it was important to address.

Sure... but since this is a benchmark scene, I am fairly sure that those using Blender, will try to optimise the **** out of it, while we, LW users are being handcuffed behind our backs becuase it is a "benchmark scene" and you are not allowed to touch it in any form, in order to make it render way faster.

What we will get is a banana vs apple comparsion in the end, unless we are allowed to optimize the scene, right?

 

[allabulle]

Very interesting, Cageman.

So I tried changing Polygon Intersection Mode from Watertight to Fastest to see the impact of the setting in a fairly slower machine.
Here it is:

PIM set to Fastest: LW_B747_Bench.lws 20m 36s
PIM set to Watertight: LW_B747_Bench.lws 22m 11s

So indeed that was quite substantial.

Renders done with LW 2020.0.1 on an i7-5960X 8C/16T OC to 4.2GHz