Great to hear you are happy with it and it helps you in your work. The documentation is definitely on my list. I already provided the patch, the builds, and some docs before reaching the goal, so I hope you understand I wait for the community to do it’s part now.
I have other patches already used in production. I use this pack of patches as a test. If it works out, I’ll slowly make the other ones public.
Just sent you a little donation. I’d be interested in adaptive sampling development, since i often have problem with specific shaders needing far more samples to clean up.
Just a note, it took me some time to find the donation link. Don’t be shy and put it to the beginning of the post or I’m your forum signature
Thank you 
Regarding adaptive sampling, I’m also definitely interested in it. 2 peoples are already working on it, so I will concentrate first on getting my work reviewed and then I’ll discuss how to best organize the future development.
Thanks for the tip regarding donations, will do
donated aswell for code review 
Thanks a lot. Please post your result or question if any
Was it with CUDA or OpenCL? For OpenCL, 128x128 is the best tile size most of the time with this patch. I would be happy to see the results for both CUDA and OpenCL
By the way, it would be interesting to see if the cuda optimizations also stack up with your stripped kernel
It was CUDA. I use RX480 as my display card (except of nights, when I could leave the PC alone).
4hours45 in 6400 x 4400 and 4096spp with 800€ of cards.
Just to put it in perspective:
That’s 100 tiles of 640px*440px less than 3 min per tile.
Full HD tile would take about 20m.
4K tile would take about 01h 20m.
This with 4096 samples. That’s really nice.
I just rendered last night with AMD; is it possible to let the frame buffer to show the actual tile during the render? Thanks
While I appreciate crowdfunding efforts in blender I am quite confused by this approach. In recent days, competition in the archviz field is so fierce that quality really matters, and I don’t believe there is space for low quality methods like faking GI with ambient occlusion anymore. Full global illumination is more than standard these days, and while I do understand that GI caching solutions are extremely difficult to implement on GPU (Hence so few GPU renderers that can do cached GI), I really do not think that resources spent on fake AO based indirect illumination solutions are spent well.
Image above is an ironic example, as such kind of interior would be rendered most likely at the same (or even better) noise level on a 800EUR CPU (Ryzen 1920x) in something like V-Ray or Corona in 6400*4400 resolution, with the major difference of having true, realistic global illumination instead of fake colorless dirty shadows in corners without any direction. And since GPUs generally have a little bit more performance potential relative to the hardware price, it leaves me really skeptical about this direction.
An ironic example… I have colleagues using V-Ray, that would amuse me to know which of this picture you think has full GI or is using tricks like invisible lights, photoshop retouching, etc. http://www.moka-studio.com/images/works/lava_terminal_3_01.jpg http://www.moka-studio.com/images/works/behnisch_museum_20th_century.jpg http://www.moka-studio.com/images/works/lava_berlin2.jpg . And those guys are working for top architects around the world.
My boss’s clients were always happy with the images provided and I couldn’t a single client that could say the difference. Of course professional may be able to make the difference from a technical point of view, but you don’t sell pictures to other visualizers do you?
If you think and your clients also do, that photorealisme is the grail, that’s ok and then use what you think is the best appropriate for this task. But don’t come like the most smartest guy on earth telling which is the only truth, because their is still a good place for other types of drawings. Even photoshoping and hand drawing are coming back pretty strong. http://www.moka-studio.com/images/works/lava_ara.jpg
I completely agree with bliblubli. Archviz today on the surface is about unnecessary detail-porn, filmic effects without common sense and impressing other archviz guys instead of serving the client’s needs. 20 years ago we optimized render times for 1-2 hour renders for a still and clients were happy. For today output res became bigger, engines are more photo-realistic and render times became just longer and longer; 10 years ago I optimized Vray renders for 40 min/still with IR/LC. For now I read 8-9 hour renders on CPU, 6-7 GPU rigs and the client cannot tell the difference as human vision is not a precise tool; it goes for the whole image, the ‘mood’ and not for artifacts, etc.
Archviz for now is like pixel-peeping when the digital revolution arrived to photography. Most of the people thinks that the magic is in the barrel distorsion or the amount of CA, etc. , so they go for the most expensive gear. Brute force against learning a renderer, new hardware against optimizing, etc. Very bad way IMHO.
Another thing: one of the best feature a render engine could offer is the scalability for the task and the time. That is why I have a Vray license: it could be used flexible. I really appreciate Bliblubli’s efforts and I really hope that there are other Blender users who will donate for him (I already did it twice). This kind of approach is very similar to what I suggested may years ago when I became Blender user: pay for pros for the pro features, giving directions for development by pro needs. I will donate him next month, too.
That, exactly that. I sold images that were a no go from a technical point of view, because of time constraints, when the boss/client changes his mind just before deadline and you have to get a picture out in really 15 minutes (including modeling, rendering, retouching). In this case, such options will just save your day. But really, in fact seeing the reaction of the client with that picture I thought would never go through, I learned they didn’t care about the details, but the mood. I already knew it, but how important it was first became clear at that point.
Nonetheless, @rawavalanche some client want photorealism. So I’m interested in an updated comparison between V-Ray, Corona and Cycles (GPU) on a single scene, with comparable noise level if you have any (it’s subjective, but if pictures are attached, everyone make it’s own opinion) and with timings on same price hardware.
I would love to make that comparison and I will make it in the future, but cannot tell, when (unfortunately). I’m still in the modeling phase for my current project, although I made a quick comparison between Vray/Cycles some time before (of course Vray won with something like half of the rendering time with non-optimized settings vs Cycles optimized settings). I’m oretty interested in Corona; I helped its development for Blender some years ago with making its GUI more compact and I had pretty nice rendering times vs Cycles.
@Almatalp I updated the build to latest master it’s in PM. It actually updates tiles, but at good tile setting of 128x128 (the fastest one with this build), tile is finished in less than 1 second in many cases, looking as it would not update. It’s just that the first update is the final one. You can see it by setting tile size of 512x512 on purpose or by using very high spp count.
Thank you; I will test it soon.
New donations are coming. You can all send me a PM for the windows builds.
