My rig is a little different. I’ve been doing animation work for a while and I’ve been building my home setup for that time. I currently run 2 GTX cards, my main is a GTX 1070 with 1920 CUDA cores running at 1632 MHz, just replaced this last October, and my other card is a GT 730 with 384 CUDA cores running at 901 MHz. Much like the multi threaded process for the CPU when you render 8 tiles at a time, with the 2 GPU setup, you can render 2 tiles at a time. The trick is balancing the two. If you set the tile size too large, the faster core will work through it’s section and you’ll be sitting waiting on the slower GPU to finish up it’s tile. On the other hand if you set the tiles size too small, the system slows down, though admittedly, I’m not sure why the smaller tile size would or should do that. However that being the case, finding the optimum balance I reasoned must be in proportion to the rendering power of each GPU set.
I came out with the following thinking:
GTX 1080 w 1920 CUDA @ 1632 MHz = 3,133,440 (rendering power aprox.)
GT 730 w 384 CUDA @ 901 MHz = 345,984 (rendering power aprox.)
Ratio from 1080 to 730 is about 9.06 (or 9) to 1. I.E. 9 tiles for every 1 tile.
Cutting the render frame up into proportionate strips seems like it would keep things in check without getting to small.
Using this ratio, I chose a tile size of 192x108y. This gave me a result of 1m 13s, one of the fastest times on my machine. I had found a few better running at a square tile size but I still wanted to try tweaking it a bit.
I ran 2 more tests using this ratio and came up with the following results:
Tilesize:96x108y for 1m 28s
Tilesize 192x108y for 1m 13s
Tilesize: 192x216y for 1m 03s (Fastest time found on my rig.)
If anyone is interested in the other data, leave me a message and I’ll be happy to share.