I know it’s too soon to talk about this since that Mac isn’t even out yet. I know sometimes reviewers do check that out before the Mac is out. Has there been anything about M5 Ultra and Topaz Video?
You can look at past experience with M3 Ultra, some have published Benchmarks… It seems to be scaling pretty well, especially on the GPU side
I’m holding off on ordering a new M5 Ultra until I see some performance results. Starlight produces remarkable results on my M2 Ultra, but takes days to process a 1 hour video. I assume that the M5 Ultra with the most GPU cores will be the best performer with Topaz?
I preordered the M5 Ultra (96GB, 80 CPU core), will post benchmarks when I have it (October). Didn’t go for the 512GB RAM ($$$), doesn’t look like Video AI makes any use of a lot of RAM, even when available. GPU is the most critical for Starlight models
What are your current speeds on the M2 Ultra and Starlight precise? I currently have a 5090 that I only bought for Starlight Precise. I’m not much of a Windows user and I’m almost done with the music videos that I wanted. lol. I’ve upscaled around 300 music videos and I have 600 on my list, so I’m halfway!! I could live without the other ones and thought about selling the 5090 for a M5 Ultra… Still undecided.
I have M2 Ultra 60 cores GPU (I believe top of the line is 76 cores) RTX 4090 and 5090
To upscale a 5s (25fps) clip of 540p resolution to 1080p using Starlight Precise 2.6 took 10min 44s on my M2 Ultra
2 min 29s on 4090 and 1 min 44s on 5090
It sounds unlikely that M5 even with 80 cores could be more than 6 times faster than M2 60 cores running Starlight Precise
I have a M2 Ultra 76 core and just ordered a M5 Ultra 80 core. (January delivery). I am hoping to see about a 4 times improvement over the M2 based on some reported M5 max results. Right now, I have been using starlight mini followed by a second enhancement with Rhea. When we finally start seeing some results in a few weeks with the M5 Ultra, I will have to make the decision to keep or cancel the order.
The first Geekbench scores on the new M5 Ultra are out. Not sure how Geekbench scores scale to Topaz performance.
Can anyone tell me how many FPS running SLP on M3 Ultra outputting at 1920x1080? If the speeds are decent enough, I might consider pre-ordering an M5 Ultra. With the crazy high unified memory it has (up to 512gb), I think there is huge potential to tune SLP for this system. I was able to see about 2x speed gain by utilizing the sysram and vram on my 96gb vram RTX card using my SLP Tuning Launcher so there could be even more potential with the mac.
I just processed a 1 hour 43 minute video using Starlight mini on my M2 Ultra with 96GB. The video was at 1920X1080 and not upscaled during the processing. I left it running and went out of town. It took almost 10 days to complete. The processing information varied between 0.5 and 0.6 FPS. at 24FPS, the actual performance was closer to .2 fps. I am suspicious that my Mac was going to sleep at times as I had “put hard disks to sleep if possible” checked in the energy preferences.
Wow that is ULTRA slow. I can process 1 hour 43 minutes in maybe 18 hours. Maybe Mac is not suitable after all. Btw, Topaz’s speed counter is completely broken. In my SLP Tuning Launcher, I show the accurate fps and Topaz’s FPS is never anywhere even close to the accurate fps.
Note sure Video AI will make use of 512GB of RAM (maybe with 4K videos?), I currently have a M4 Max with 36GB of RAM but neuroserver doesn’t even use half of it. The videos I enhance are old, mostly 360p or 480p.
Macs are very slow in comparison with a 5080 or 5090. My hope is the M5 Ultra with 80 cores would get closer to a Windows/NVIDIA machine, maybe a 5070? but aside from raw performance, I also wonder how efficient is Topaz’s use of Mac GPU/Neural accelerators vs. CUDA.
This is why I ordered the 96GB only.
Topaz speed counters are completely broken, this is why I have developed scripts to do the math (it’s not very complex math, it baffles me this is still not resolved!).
To give you some idea of the actual speed, here are some image sequence timestamps for my M3 Ultra 96GB:
v1.7.1 MacOS 27
SLP 1x 1920x1080
Process started 23:46:20
Frame 0 23:59:57
Frame 178 00:11:34
Frame 278 00:23:01
Frame 378 00:34:26
Frame 478 00:45:56
Frame 578 00:57:24
So once it gets going after the initial setup, it’s around 6.9 spf.
About M5 vs. M3 comparison, Got a quick summary online (AI)
- Memory Bandwidth: The M5 Ultra offers 1.2 TB/s, a massive jump from the M3 Ultra’s 819 GB/s, significantly improving local LLM and large dataset handling.
- Graphics Performance: Despite identical core counts, the M5 Ultra’s redesigned GPU is 40% faster overall, with a 59% higher Geekbench 7 Metal score (366,744 vs. 230,420).
- AI Acceleration: The M5 Ultra features per-core Neural Accelerators, making GPU compute 4.5x faster for AI workloads compared to the M3 Ultra.
Without knowing how Topaz will support the new ship, we should start with an expectation of a 40% increase over the M3 Ultra, maybe more?
Thanks
Is Starlight Precise coming to Mac? The current release only allows local processing for Mini. Everything else is cloud only.
For what it’s worth, PC can run Starlight Precise and all the other Starlight models locally and it’s substantially faster. My PC ran the same clip in about a quarter of the time. (rtx4090 vs MacStudio M3 ultra)
It’s been working on Mac for a while now… There are RAM requirements, 16GB I think, 24GB recommended.
A Mac version of your SLP Tuner Launcher would be much appreciated! Topaz are notorious for not bothering to attempt any form of optimization for Mac.
It’s as though Topaz thinks Mac users should just be so grateful to them for producing a Mac version at all, and we should be satisfied with that.
I can almost here my M3 Ultra yawning out of boredom when running any model - even SLP. At least half the VRAM sits there unused and total power consumption seldom goes above 100W when running SLP. The fan speed only increases to around 1300 rpm out of a range of around 980 to 3650 rpm.
Thanks.
Andy
What is your GPU utilization? My M2 Ultra always runs at almost 100% when running a Starlight model
On 3 different machines (M4, M4 Max, M5) I am almost always at 100% GPU for Starlight Precise. However, RAM appears to be underutilized.
The issue is the speed is just too slow for M5 Ultra, even if I can squeeze a 3x out of it with a similar Tuning app, it will still be much slower than my Nvidia PC that does SLP at 2.0 fps processing 1920x1080 with my Tuning App. I’ll wait for actual benchmarks to decide if I should consider getting a Mac. It is not a small investment.
