PI and Mac Studio M1 max - feedback?

rodolgo

rodolgo

Well-known member
Hello PI-ers,
I'm considering upgrading my processing power with a Mac Studio (M1 max).
Has anyone any REAL-WORLD experience to share regarding working with PixInsight on that system?
Looking particularly at overall speed, Starnet V1 or V2, and all those lenghty operations like WBPP, etc.
Thanks in advance for sharing,
Rodolphe
 
m1 ultra w/. 128GB. only problem so far is that there is apparently a rosetta bug that causes imageintegration to hang for 10-15m after integrating RGB images. i don't see the problem with mono images. this likely happens on all ARM macs.

i was expecting OSX to be pretty good with scheduling and thought that interactive performance would be good while PI had all the cores busy, but the machine gets a little choppy under load. so i think i backed off to either 18 or 16 threads.

WBPP speed with ramdisk as a target and internal SSD as a target is the same. the IO bandwidth on this SOC is pretty phenomenal.

as for actual numbers, i didn't write anything down so i'd have to repeat the tests.

rob
 
Thanks for the feedback Rob,
It seems I'll get a big boost, coming from a Macbook Pro I9 equiped with a Radeon Pro 560X 4 Go GPU!
Rodolphe
 
i think the biggest win is power consumption. with all 20 cores running, the m1 ultra consumes about 100W and around 20W when idle. my 16C/32T intel machine would pull 400W when loaded and almost 200W idle. the video card alone seemed to want almost 100W at idle. electricity is very expensive where i live so this has been a significant cost savings.
 
FWIW, I use a Studio Ultra 64GB and most everything is very fast with PI. My point of comparison is with the M1 mini I was using with a similar workload. The Ultra is about 2.5 times faster for what I do measured by time taken. I fully expect that when the native version of PI is available it will gain an additional 25-30% performance. If only the GPU cores could be used…

As a side note, I do manual steps and not WBPP and I haven’t run into issues with ImageIntegration and all my images are OSC. Although now I have adopted a bayer drizzle workflow so it doesn’t matter anymore. I have noticed that PI never seems to see all the available memory on the Mac. It usually reports about half or a third of what is actually available. A few others have noted this as well on the forum.
 
FWIW the settings I use in ImageIntegration:

  • Average
  • Additive with scaling
  • FITS keyword
  • Generate Integrated image
  • Generate drizzle data
  • Subtract pedestals
  • Close previous images
  • Automatic buffer sizes
  • Use file cache
  • Generalized ESD
  • Scale + zero offset
  • Clip low, high, low range
  • ESD Outliers 0.3
  • ESD Significance 0.05
  • Range Low 0
  • Reject low and high structures set 2 for each
  • N* robust noise estimator
  • 5 scale
  • PSF Auto
Since I'm always using drizzle now I don't have my setup for how I was using ImageIntegration before but I believe I tried both linear fit and ESD with defaults for pixel rejection 1 for both. I don't recall extended delay times.
 
pfile said:
m1 ultra w/. 128GB. only problem so far is that there is apparently a rosetta bug that causes imageintegration to hang for 10-15m after integrating RGB images. i don't see the problem with mono images. this likely happens on all ARM macs.

i was expecting OSX to be pretty good with scheduling and thought that interactive performance would be good while PI had all the cores busy, but the machine gets a little choppy under load. so i think i backed off to either 18 or 16 threads.

WBPP speed with ramdisk as a target and internal SSD as a target is the same. the IO bandwidth on this SOC is pretty phenomenal.

as for actual numbers, i didn't write anything down so i'd have to repeat the tests.

rob
Click to expand...

Hi Rob,

I have also recently moved to an M1 Ultra 64GB, can you give me some pointers on the best settings you have found to work on the M1? eg. Parallel Processing settings and swap folder location and number of swap folders? Any info would be much appreciated!

Thank you.

Reggie
 
not sure that it's tuned properly but i've assigned 16 CPU threads and 4 reader/4 writer threads.

i put 4 swap file directories on the internal SSD but i think this can be increased. you can change the values and then run the benchmark to see if disk performance increases as you add swap directories.

as for the rosetta problem, i've noticed that for some reason the OS is not offering me the 12.5 upgrade. i probably have to reboot the machine. there's a chance they fixed something in rosetta - on another M1 computer i have a security camera application running (under rosetta) and it would kernel panic the machine about every 5 days. i'm not seeing that anymore after updating to 12.5 over there.

rob
 
i haven't tried yet but i wanted to upgrade the OS first. maybe i'll try it. it would be easier for me if you could save a process icon, zip it, and attach it here.

rob
 
pfile said:
i haven't tried yet but i wanted to upgrade the OS first. maybe i'll try it. it would be easier for me if you could save a process icon, zip it, and attach it here.

rob
Click to expand...
I've attached it.
 

Attachments

  • ImageIntegration.xpsm.zip
    1.8 KB · Views: 49
i missed it but having the drizzle integration icon doesn't help; my point was i could just load your icon and double click it to try (well, and load my own images). it's actually easier to configure II from your forum post than try to dig it out of the drizzle icon.

anyway at this point it could be a "me problem" so i don't know if i'm super motivated to set it all up again.

rob
 
I bought a Mac Studio Ultra for other reasons a few weeks ago, but haven't been able to try it for use with PI until this weekend.

I have to say its a great experience - its much faster (~ 4 times) than my previous i7 Mac mini in pre-processing, and noticeably faster when I do something computationally intensive such as deconvolution.

It's also very quiet.

One key thing is to use the Amphetamine app to stop the studio from sleeping during pre-processing if you leave the computer unattended. I've set up a rule to prevent sleep if more than 40% of the CPU is being used, which works very well.

I had thought about buying a Threadripper based machine and running that remotely to avoid the fan noise - but this box is such a pleasure to use, especially paired with the studio display.

Colin
 
You must log in or register to reply here.
Back
Top