Is that a positive or negative thing?
Generally we prefer processing with larger tile sizes, since this gives the model more visual context when analyzing input frames.
We’ve talked about this before I think. That version of Iris was not running correctly and is not representative of theoretical performance on Apple Silicon.
Apple - inadequate cooling, not made for heavy workloads, made primarily for ‘casual’ users.
They have a good marketing department though!
4 Likes
So, since I got 128GB RAM and 4080 with 16GB, having GPU scheduling on would be better then? Since it accesses more ram?
Does the rtx 5090 have ecc memory?
Didn’t read that anywhere.
This is not correct. That version of Iris definitely was running correctly until the early builds of Sonoma broke it.
Then you introduced a quick fix that hurt performance by a big margin. I believe it was told at that time that the speed loss would be temporary.
But it persists until now, mainly crippling the (I think) most used 2x upscale.
Even worse:apparently Apple reversed or fixed what they broke in the early Sonoma builds later on as the old model when copied over from a Time Machine archive suddenly did work again.
And even at the time where the bug existed it could be cured by simply starting the app under Rosetta - and, surprisingly this didn’t even have an impact on the speed.
See e.g., but not only my old detailed report here:
Strange enough we still have two Mac minis Late 2012 still running and that 24/7.
While the Core i7 12700 with RTX 4060 (so really not so high-end) broke down in only 5 months.
But again: this is not a Mac vs. PC argument here, I don’t really know why some windows users are always triggered when someone speaks about the Mac.
1 Like
That explains your experience with the heating problems and consecutive defects as that was before the AppleSilicon times - those Intel chips, albeit being mobile/Notebook-chips used there did use quite a lot of power and got HOT. Together with Apples chronically too low set fan speed in favor of silent operation the chips got temperatures up to 100 degrees.
This is not a problem with the Mx chips anymore, at least not those that I know.there you seldom get temperatures over 70 degrees.
(With the older intel models I always used custom Fan controls to defeat that issue)
1 Like
I was curious about this so ran benchmarks with and without GPU scheduling enabled. Results below for the curious. Most results are within the margin of error but APFast and Hyperion HDR saw the most notable increases by disabling GPU scheduling (on my particular hardware anyway).
GPU SCHEDULING **DISABLED**
Topaz Video AI v6.0.2
System Information
OS: Windows v10.22
CPU: Intel(R) Core(TM) i9-14900K 127.8 GB
GPU: NVIDIA GeForce RTX 4090 23.576 GB
GPU: Intel(R) UHD Graphics 770 0.125 GB
Processing Settings
device: 0 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 42.86 fps 2X: 18.81 fps 4X: 04.92 fps
Iris 1X: 41.51 fps 2X: 20.88 fps 4X: 04.96 fps
Proteus 1X: 40.20 fps 2X: 18.53 fps 4X: 05.22 fps
Gaia 1X: 15.70 fps 2X: 11.21 fps 4X: 04.68 fps
Nyx 1X: 17.94 fps 2X: 14.92 fps
Nyx Fast 1X: 34.52 fps
Rhea 4X: 04.42 fps
RXL 4X: 04.50 fps
Hyperion HDR 1X: 32.21 fps
4X Slowmo Apollo: 38.38 fps APFast: 77.11 fps Chronos: 33.24 fps CHFast: 34.78 fps
16X Slowmo Aion: 35.28 fps
GPU SCHEDULING **ENABLED**
Artemis 1X: 40.67 fps 2X: 17.36 fps 4X: 04.89 fps
Iris 1X: 37.65 fps 2X: 18.66 fps 4X: 04.82 fps
Proteus 1X: 40.09 fps 2X: 18.46 fps 4X: 05.13 fps
Gaia 1X: 15.69 fps 2X: 10.91 fps 4X: 04.73 fps
Nyx 1X: 17.40 fps 2X: 14.16 fps
Nyx Fast 1X: 31.13 fps
Rhea 4X: 04.48 fps
RXL 4X: 04.24 fps
Hyperion HDR 1X: 26.33 fps
4X Slowmo Apollo: 35.90 fps APFast: 67.79 fps Chronos: 32.01 fps CHFast: 31.33 fps
16X Slowmo Aion: 32.23 fps
1 Like
Now that seems like a good idea.
But why don’t you just do it yourself?
(I would do it but am on Mac where I didn’t notice the problem - and am not on a current subscription since everything after 5.3.x didn’t really bring something with it that justifies a new payment 
)
This is not user-friendly solution like in older version.
Additionally, in the previous version you could go back to the first preview frame with one click - now it requires a lot of gymnastics.
1 Like
Dear TopazLabs,
You seriously need to address the stuttering / jittery output of models such as Proteus.
This is a serious issue and makes these "ehancements " meaningless.
I have reverted to 5.3.6 to assure smooth video output.
4 Likes
It still happens that the first frame of the enhanced preview appears several dozen frames after the original frame. This is a very annoying error. How I can compare improvements?
When I set output resolution from 4:3 SD to 16:9 FHD I still don’t have equal zoomed view.
The program still appears lower and further to the right each time it is opened, without remembering the correct window position.
I’ve been using another ffmpeg build for software H.265 encoding. It was built like 17 days ago. No issues with frame rates in it.
3 Likes
Thanks for trying that out!
1 Like
Yes, I have. The situation this always happens in is when doing split-and-stitch carelessly. By carelessly I mean that the splits are made into chunks that aren’t framerate divisible. The ffmpeg calculation then “freaks” out in the concat filter and marks the clip as VFR, even though the frames are evenly spaced temporally in the final clip (i.e. CFR).
There’s a lot of talk about the VFR-misclassification on doom9 regarding Mediainfo in particular, with the repeated statement of “Don’t use mediainfo to check for CFR/VFR”.
I ended up just running anything that fiddles with frame adjustments via avisynth to get frame accurate management (ffmpeg libs suck at this).
Sorry, didn’t dive into the libav code regarding this, so don’t have more technical insight to share. But asking someone on the team to browse through those discussions at doom9 may be worthwhile.
1 Like
ffmpeg is used in TVAI which does correctly report correct VFR or CFR
I’ve remained on version 5.4 because the new UI is so pathetically bad. Confusing, not intuitive and slows workflow down dramatically to the point where I simply cannot use it. Reversion to a less insane UI is absolutely necessary. Why are you guys frustrating the heck out of paying customers?
4 Likes
Because whatever combination of settings/models I tend to use doesn’t present the problem. Doesn’t mean I can’t offer suggestions for others.
