Why did I do this project?
I was interested to work with ai on a local system instead of all these cloud solutions, where you often don’t know what happens with your data. My first contact with local ai solutions was pinokio, a system that automates localized ai installations. Pinokio runs lots of python scripts locally and I have lots of respects of these scripts, because they could do anything on your system. That’s one reason why I wanted to run them on a non productive computer of mine. The second reason is, that lots ai models rely on Nvidias Cuda cores and wont even run on a Mac. The third reason is that I thought that I can speed up my workflow with very large scans with a system that has more ram (I was wrong on that one).
Reasons for the hardware I chose
For the reasons mentioned above, I wanted a Nvidia card. Nvidia are producing cards that are made especially for ai work (Nvidia A series), but these cards are very expensive. The high end gaming cards like the RTX 4080 and 4090 are also pretty expensive and are as well very energy hungry. So my decision went to the RTX 4060ti with 16 gigabytes of ram. The internet is full of hate for this card, because it’s too expensive for what it delivers on gaming. But gaming is not an important part for me. The 16 gigabyte of ram are essential for ai models and the low power (compared to 4070 and above) consumption is another thing that made it more compelling for me. My CPU choice, the i7 13700kf was done because of a great offer I got. I hoped the new Intel 14th generation will consume less power, but that was not the case. Maybe AMD would have been a better choice for less power consumption…
Mainboard wise I went for ASUS, because I had good experiences with this brand many years ago. I think hardware wise, they are still great, but software wise I am not sure anymore (I dont want to go down in that rabbit hole today).
To make the system as silent as possible I went for an aio (al in one) water cooling from Artic / the Artic freezer 240 and the Be Quiet pure base 500 case. Both brands I know from back than and they are still great today. For the power, I went for the Corsair RM750x Shift Fully Modular ATX Power Supply, because the plugs are all on the side and much easier accessible than on usual power supplies. I calculated the needed power supply here. https://www.corsair.com/us/en/pc-builder/psu-finder/psu
Here is the full list with my Amazon affiliate links. I always think it’s great to have also a look at local stores, but as an overview it’s easier to link to Amazon.
Intel i7 13700kf
International: https://amzn.to/41iT0m5 Germany: https://amzn.to/46MAFPu
Iceberg Thermal FUZEIce Plus High Performance Thermal Paste
International: https://amzn.to/41iUL2F Germany: https://amzn.to/3t3fHOF
ASUS Prime B760-PLUS
International: https://amzn.to/48eovjF Germany: https://amzn.to/3uTHJws
GeForce RTX 4060 Ti Graphics Card 16GB
International: https://amzn.to/3NlgMYN Germany: https://amzn.to/46MBkAs
Crucial Pro RAM 64 GB Kit DDR5
International: https://amzn.to/3RBk4K0 Germany: https://amzn.to/3NnYdU2
ARCTIC Liquid Freezer II 240
International: https://amzn.to/3tlln6x Germany: https://amzn.to/3TiTS8s
WD_BLACK SN770 NVMe SSD 1 TB
International: https://amzn.to/48cLuLE Germany: https://amzn.to/41emhOR
Crucial P3 2TB PCIe Gen3 3D NAND NVMe M.2 SSD
International:https://amzn.to/489Hrjw Germany: https://amzn.to/48cLCe6
Corsair RM750x Shift Fully Modular ATX Power Supply
International: https://amzn.to/485SHxP Germany: https://amzn.to/3RE1z81
be quiet! Pure Base 500
International: https://amzn.to/3NijfU3 Germany: https://amzn.to/3NlVESy
Finding all components without RGB was more difficult as I thought.
How did I do the tests?
As mentioned in the video, all my tests are not scientific in any case. I just wanted to see for myself how much of a difference such a machine makes to my workflow.
All recordings were done remotely. That means, I connected with Microsoft Remote Desktop to the Windows machine and with the Apple Remote connections to the Mac Mini and MacBook.
With that I wanted to make sure, that I don’t use to much computing power for the recording on the tested device.
For the Lightroom tests, I used the images of my award ceremony where I saved all the changes to xml files, so all clients have the same work to do. I chose Lightroom for testing, because lots of analog photographers will do their scanning with a digital camera.
As an analog photographer, especially with the wet plate process I work often with huge scans that I use for my talks, like my last one at the 8K Deep Space experience in the Ars Electronica Center.
With that big scans I experienced sometimes some slowdowns on my MacBook (So I thought). To make it more fair, I tested it with a 180 megapixel scan and 1.12 gigapixel scan. This test was for sure the least accurate one of them all. Because I did every step manually. That’s why I said the test times will vary up to 10 Seconds because of my click and work speed. But you will see that does not change anything on the results.
Why did I measure power consumption?
I think this is a part that lots of tests are not taking care of. I understand that time is money, but energy is as well. With the energy costs going up every year, it is an important topic to me. And when a device needs 5 or 20 times the power than another device, it makes a big difference. I think saving energy is an important task for everyone. I understand compared to big corporations (see my article here) this is nothing, but I still think we all can make a difference.
And yes, my power consumption measurements aren’t scientific as well, but you get an idea about the difference. I used the same power plug on all computers, to avoid differences. And yes the power plug was called Windows computer, even a Mac was plugged in :). One more thing I closed the lid of the MacBook to only measure the consumption of the computer and not the display. With the display turned on at medium brightness it was only 2 to 4 watt more.
The results
I am still surprised how well the base M2 Mac Mini with 8gb of Ram and a 256gb ssd held up. For under 600 Euros this is a great deal (also power consumption wise). If you invest 100 Euros on an external ssd, you can easily work with bigger scans in photoshop and will avoid the scratch disk is full message.
The second competitor to the windows machine is a MacBook Pro with an M1 Pro CPU. It has 16GB of ram and a 1TB ssd.
Please be aware that the power consumption is only a snapshot from the time when the task was finished. But overall it reflects the consumption that was used.
Saving on my Mac took about 10 minutes. Thats why I was surprised that it took the same time on the Windows machine with 64gb of ram. SSD wise both machines perform similar. As mentioned in the video, I have an open case with Adobe because saving takes that long. And just in that moment when I am writing that article I heard back from them. They tested my file and confirmed that these long saving times are as expected with large files like that. You can see that the power consumption on the windows machine is pretty low in that case, but I think its only that low, because saving a file for 10 minutes is a task that does not need much power.
My thoughts
The new installed Windows machine was faster at every single task. On some tasks it just killed it.
I am not sure if it would have made a difference to use a new installed Mac, because I work with the same backup on my MacBook since 2016, but I don’t think so.
For a final conclusion the question should be, how much benefit would I get from a faster machine. How important is the time I saved with the faster machine. I think there are always tasks you could do while Lightroom is exporting or photoshop is saving. During my work on the huge scan in photoshop, only the Mac Mini kept me waiting for some seconds when I duplicated the layer or applied the filter. Also the loading time took much longer. But otherwise, I was surprised how well it held up against the two others.
I know I went a little overboard when I compared the Windows machine with my heat pump. But sometimes the computer needed 350 Watts and much more. And when it’s about 15 degrees outside, my heat pump has a similar power consumption to keep my house warm. Let this thought sink in for a moment. Luckily I am can create my own energy and will try to limit the use of the windows client to the sunny days.
I hope this comparison showed you a different view of how powerful a computer must be.
The question is now, was it a good idea to build this computer for ai and scans. Honestly I am not sure anymore.
It’s fun to play with ai models locally, but I will limit the use for Lightroom or Photoshop. I guess I will use it for huge scans where I need to work with more layers, but for my usual workloads, my 16gb MacBook is more than capable.
Also the noise it makes is something you should think about. Even this machine is pretty quiet (it just sounds loud in the video, because I went very close with my phone) I dont like listen to the fans all day long, that’s why I put it in a different room and only connect remotely to it.
Honestly it’s hard to justify buying a Mac Studio with similar specs like this machine. You easily spend 5000+ Euros on it and with that extra money, you can power up this one with energy for years. But if you don’t mind waiting a little longer, you got cheaper options that consumes much less of energy.
At the end it’s always a decision what operating system you prefer and how much money you want to spend.
