I wanted to get a rough overview over the speed of the Redefine and Recover models in Topaz Gigapixel with various hardware. Since there is no benchmark built-in (as we have in TVAI), I quite randomly chose two images, a low-res and a higher-resolution 12MP one.
So here are the times:
Mac Studio M2 Ultra 24 CPU cores / 60 GPU cores, 64 GB RAM; Gigapixel 8.1.1
OK, so this really is a magnitude faster than on my rig!
P.S.: As TVAI, TPGP doesn’t seem to be anywhere optimized for AppleSilicon - it’s often only using about 60Watts in total for the whole system, not really using neither GPU cores, nor CPU and even only 2 Watts for Neural Engine (while I know from TVAI that Neural engine can consume over 20Watts in use)
Hi-Res Redefine: I stopped after an hour
Hi-res Recover: too long to bother!
So there is no performance advantage with an M4 over an M2 Ultra. I have no idea how to measure things like wattage etc. to know whether full advantage is being taken of the neural engine.
I am curious whether TPAI Super Focus makes better use of the M-series neural engine. Here are my results for jo.vo’s hi-res image (SF entire image, all settings minimum)
Calculating ETA: 34"
Predicted ETA: 12 min
Actual time: 9’30"
Intel Core i9 13900K limited to 5.5 Ghz, 300 A IccMax, 90 °C max. temp. + 32 GB DDR 5 @ 6 Ghz + ASUS ROG STRIX OC RTX 4090. GPU power consumption up to 500 W on full power during test omg.
As @lhkjacky said: this comes up after you use the preview entire image button.
Which is what I always use first anyways as the results of those generative models are quite unpredictable and so I only save the result when I like it ( thus not having to delete the output file if the result was crap).
No fear: after previewing the whole image you can save the result in seconds through the export dialog without it needing to be reencoded.
Recover 2x
low res: 10m 47s
high res: still to be evaluated
Redefine
low res: 1m 51s
high res: 30m 16s
I edited the first post for those recent numbers.
I use Gigapixel 8.0.2 now as 8.1.1 has the bug with sometimes broken redefine renderings - but I shortly tested this before and there was no real difference in speed between those two GP versions.
The only other change there was is that I’m now on Sonoma 15.3 Beta (and, of course needing a restart for that to install). So, did Apple change something that leads to those faster times in Sonoma 15.3 - or did something get stuck before and was cured by the reboot?
The numbers look a bit as if the computer was stuck in power saving mode - but then again, if I turn on power saving mode now there’s no real difference in rendering speed of Redefine/Recover
Maybe someone other with still MacOS 15.2 can test this before/after going to 15.3 (beta) to verify this. And: if you get slow rendering speeds it might be worthy trying a reboot (according to that old Chinese saying: “reboot is good” )