I did this for a book cover. Because I hope to put photographed people in this scene, this may not be the final render, but everything is there that shows.
Because it’s so render unfriendly, I sent it to McNeel for examination.
This was rendered at 4k, with Cycles, using 15,000 samples, and 32 bounces. Some of the materials such as the asphalt and the bricks have 4096 sized textures. Others like the yellow barrier are smaller, given their distance from the camera. The image has 10% contrast enhancement, for mood. The background will need to be replaced. The black paint gloss took quite a few tries to get it as close as this. The bushed stainless has been and always will be a work in progress.
Some 19 years ago, I used this same kind of diner in a free video game level pack.
My computer is a 7 year-old 3.4Ghz i7 2600k Quadcore, with 16GB RAM, and a GTX 1080.
Render time? Are you holding onto a table or something?
About 16-18 hours, for 15,000 samples. The online image is smaller than the 4k version I rendered and tried to upload.
At 10,000 samples, only fine noise remained on the green chargers, Now I don’t notice any.
The upcoming AMD 3800x is calling me. : ) A second video card would help a lot, but then it will be bottlenecked by the CPU, anyway.
Yes, for a presentation-style render, I wanted to drive the noise below a noticeable level, and that took some time.
In watching Windows System Monitor, the CPU has small 100% per core spikes while feeding the GPU, so there’s little point in getting a used GPU from ebay, when the processor, is now some 30% lower than modern single-core speed. More cores would help other things, too. I suspect that the new processor can turbo or micro power manage better than what I have now, so the 2-core speed might be a lot better than what I have–even if the single-core turbo speed is not anywhere near twice as fast.
The memory with my current processor was never the fastest, either, but it’s been checked with Mem86 several times with no errors.
Except for a power glitch compromising the UEFI/CMOS, this MSI motherboard has been reliable. It has spent plenty of nights exporting videos for my youtube channel. It has spent others running off animations and rendering from Rhino. The computer has an extra-thick motherboard, a 1050watt power supply. The old video card had a lifetime warranty. It has an old 1TB SSD. It’s in a rackmount case, that’s inside a roto-moulded portable case. It has bluetooth and Wireless Wifi. The inside of the rackmount case is lined with indoor/outdoor carpeting to make it quiet.
Still the Ryzen 7/9 is calling. The memory costs as much as the processor. I need 64GB, for my data visualizations in Paraview, though I guess I could get 32GB–if I could buy a brand I would be sure to match. Money is tight right now. Real life happens.
Sorry, this is a windows-only tool at the moment. When I’m done with clipping volumes in Raytraced for v7 I’ll be looking into adding denoising in different ways, also non-CUDA.
Not in this render. I wanted this baseline before I ran it. I wanted to see what puts out the noise first, which seems to largely be 10,000-15,000. Now to pick out a number to start with, like 5,000 or 7,500, and run the test.
I also have an issue, where I wanted to get this thing largely done. My computer has been running most of last week, and I don’t mean that figuratively. I only have one Rhino license, which means that I cannot do anything else in Rhino while it’s rendering.
I will run a denoise test in the morning.
[I finally tried Quake II RTX, this evening. The denoising can be toggled in it.]
There isn’t that much geometry in this drawing, BUT because the diner is full of reflective metal and glass, it makes quite a rendering challenge. Because the diner is almost fullbright near the ceiling, and quite dim outside the diner, it also complicates things for any raytracer as far as noise is concerned.
Because I am interested in learning Cycles’s capability, I wanted to understand what could be done before any image filtering, which would likely give a comparable image with less rendering time.
Also, as stated, my computer is not only old, but my GPU is likely processor-bound.
That shouldn’t be much of a problem. Before Raytraced does its actual tracing it uploads everything into GPU RAM (assuming it all fits). From there on the CPU only manages jobs and reads results - no further scene data updating from there on. It could be a bit of a factor once you’re doing huge resolution renders, but I haven’t profiled such a case.
Windows happily reports 20%-40% overall utilization while rendering, but I’ve noticed very short multiple 100% per thread peaks on Windows System Monitor. The speed has to be turned up to see them.
A little faster per-core speed would help. More cores might even help.
I imagine reading and even sending the data to the GPU to be a mixture of 32 and 64 bit instructions. It seems reasonable that whenever there’s a 64-bit instruction, the hyperthreading is going to hiccup.
Because all the cores are working, I have almost no turbo going on; the stock CPU clock is 3.4Ghz. A newer processor would also help because they can better micromanage the clocks for the DTP.
If the processor had just a few hundred MHz and perhaps one more physical core, it could be enough GPU render free from bottlenecks for a single GTX 1080.
The kernel times are putting a nice peak on the usage : P Could be slow memory, I suppose. It’s not a VM write; VM is completely off on my machines.
(GPU0 Reads ziltch because it can only see certain GPU Processes with their crappy monitor. Yes, I did check this with GPUz.)
I’m no expert in this but I’d have thought the performance hit from your GPU waiting for a momentary instant waiting for the CPU is unlikely to affect overall performance dramatically? If your CPU is running 100% for long periods of time that might have more of a noticeable impact on render times.
My hunch would be that adding a second GPU would have more impact on render times assuming Cycles can use them both? You might then see more 100% CPU utilisation but I’d imagine that the overall reduction in render times would be greater.
