Bro. I’ve been looking for some reviews and bumped into your channel. You look so sincere and kind so it touches me somewhere inside) subscribed! Wish you more RUclipsr’s successes in future, you are definitely worth it)
Hats off for giving an amazing well rounded review, this is by far the best mac mini review i have seen so far on the internet, even when compared to big tech reviewers. The community contributions to making this video very relatable and resembling real life use cases was a chefs kiss, extraordinary efforts, i am subscribed to your channel.
Hey. Thank you so so much for including music production and logic pro. Everyone and their mother neglects stress testing in audio and music even though it is a special test that highlights single core performance above all. Thank you!
I think I found the definitive youtube channel for tech reviews. Thank you so much for considering what people say to you. Fo real, I'm a musician and the device-performances with Logic is really something to me and you got it. Cheers from Chile
🍄 Let me know if you see him too! 😅 This video was a labour of love! No sponsors again, but really hope you enjoy it. I took a lot of your feedback onboard for this one, but I am sure it can be improved even more! Have a lovely day 🙏🏼
Could you please give me some suggestions for the best docking stations for the mini M4? A docking station that has an SSD slot and some extra pins can help me pump up the memory of the 256gb mini..
Best one I found without spending a fortune is the Caldigit TS04 - I shared a link down below, it’s great even though it’s TB4, but will also look into the Thunderbolt 5 stuff when they come out
You did great. I will note that some of your stress tests are really about memory, and the simple M4 with 32GB RAM is likely to do much better than with 16GB of RAM. With the base RAM level the Mini will do a lot of swapping with these big programs and that is going to slow you down.
@@AlexGTech Kontakt is a sampler and thus doesn't tax the CPU that much -- just a lot of disk i/o, streaming samples from the ssd. Using a synth like Alchemy (built-in, so free once you have Logic) would give a more interesting result imo. Interesting video nonetheless; thx for making it.
@@Mpthedawn I found that it was still running past 100 but 200 made it slow down and crackle a bit. I’m guessing the limit in these type of tracks is in that range 100/200 but I need more time to find out when exactly it start to slowdown
Excellent video, so great to see you taking on community questions. I have a 2009 Mac Pro, 2 x 6-core 3.46 Xeons, 96GB ECC Ram, 2 TB NVMe, USB 3.2 Gen 2. RX580 8GB. I was waiting for a Mac Studio but don't know if I should just get a M4 or M4 Pro Mac Mini. Looking forward to your next videos on the Mac Mini's. A.
Commenting just because of the efforts it must have taken to create this detailed video. I started RUclips channel, created just 4 videos and the comparison video between iPhone 14 and 15 took such extreme efforts, even though that was much simpler than this. Hats off. Keep doing the great work
Hey Alex! In settings on Logic the i/o buffer is what really matters for track stuttering! (The lower this number, the less latency there is between hitting play and the daw playing back). This gives the daw time to calculate what the plugins are doing. It's generally a great test to put it to the lowest buffer (32) and see how many tracks you can get. When producing you'll often record/make a basic track with the lowest i/o you can and then crank this number up for adding layers and mixing since you don't need realtime playback. At any rate thanks for all the tests this was great to see! You can watch how L.Dre does his tests in Ableton (or iamamusicmogul in logic pro) for a similar example :)
No matter on what computer you use the LL model. If it's the same models, it will give you the same results. Well, randomly it would be sometimes better, sometimes worse (like in your case). The only real difference is speed. And considering speed, you should've include token tps (tocken per second) metric, because results will vary in volume.
@@chidorirasenganz I've never had that case. However I would assume maybe a thermal issue might present itself if maybe on the MacBook Air. For example Monster demo file, 1:48 with CPU and GPU, with GPU only 1:58. I ran though all my files and at minimum I see at least a 3%. uplift in render times.I've tried to find the source of what you are bringing up and the only thing I could find is about render tiling, which you would tweak to get scenes to fit in RAM better. As soon as the scene renders and uses tilting then GPU/CPU seems to come into play from my testing.
@ I don’t know if something has changed but previously all the info I’ve seen on using Blender on Mac since the M1 days has been GPU only is faster. I think it was the GPU waiting on the slower CPU to finish rending but I can always try it again on 4.2 but since Eevee Next came out (tbh Eevee in general) I haven’t really care about render times since most of the time it’s 5-10 secs at most
Excellent testing video and analysis. The M4 Pro is clearly for those high-end professionals with no time to lose. High end pressure news reporting, movies/documentary sets with a tight filming budget, and Machine learning on a professional level. The regular M4 Mac Mini seems to best for those without tight timelines.
Thanks so much, I enjoyed learning some of these different apps and can see why this machine has such potential! Much more to come, but for now I need a rest 😂
Please test Minis in Adobe dtp triad: PS, AI and ID. I'm thinking about getting Mini for myself but I'm not sure if I should pickup base M4 or base Pro. I'd like solid machine to work on 2D product visualisations, design 2D graphics and put everything into 50-100 pages catalogs. This is what I do professionally.
Good stuff Alex…nice to see clear and humble - non shouty - info. As a heavy LogicPro/NI/Arturia user I’m keen to see the music tests of course. Keep it up. JPMusic.
Thank you so much! Man, I am HOOKED! I have no clue what I am doing in Logic, but used to use Ableton back in the day, so there was a little muscle memory in there. Really want to bring some keys into the studio and have a Mac mini setup to mess around with it. I won’t be hitting those track numbers, that’s for sure 😂
The thing about most music programs, from what I've seen on another RUclipsr's (James Zhan) tests along with my own tests too, is that they dont utilize the efficiency cores AT ALL. It kind of made me wonder when I saw Reaper utilizing all the cores, but Logic didn’t.
Yeah, I am glad I am not going crazy! I did wonder for a while, if I had made a mistake in the settings. Will check that James guy out, I am learning so much making these videos! Love it!
@AlexGTech 😂 you're not alone, I thought I was too! And do check him out. He's been doing comparison videos on all the Apple Silicon Macbook Pros since the M1, so his videos will make for great references. Keep up with the great work yourself, too, man! 💪
Very useful test overall, thanks for the video and all the effort! Two suggestions on the LLM part: - The power of the chip, like M4 vs M4 Pro won't affect the quality, any such result is coincidence :) - The chip's speed will affect how fast the LLM can generate the answer, the usual benchmark for this is in token output / seconds. There is a built in 'randomness'. Answers to the same prompt will be different each time, so measuring completion speed doesn't make sense. Keep it up, great stuff!
thank you for confirming that, I thought it was a bit counterintuitive, and perhaps I had a 'processor' bias while I was doing the tests and my brain played a trick on me! The stable diffusion test was really odd though, how out of 10 pictures prompting for a hot air balloon and birds, the Pro generated 4 including both and the standard M4 only generated 1 picture with a bird, but nothing with hot air balloons. I guess I need to run multiple tests though to really see if it was a 'one-off' or if having a beefier CPU/GPU makes a difference
Definitely! I tried to enable that via the command line using Apple’s own bloody Metal code, but only steam games worked and wasn’t even Apple’s thing, it was just Steams basic counter. Really annoying. And I ran out of time to get an FPS counter app, but will research into a good one 👊🏼
This was a great video, you're very humble man. I'd love to see a comparision between m4 mac mini vs a similarly priced windows laptop and how it performs in both gaming and editing
Really helpful so far. I have been convinced that I needed the Pro MAC Mini to produce, edit and upload professional RUclips videos. Plus podcasts. But you are making me think I may be able use the base model and save myself some serious bucks. I'm going to enjoy your docking station video too. Great job! I'll end up viewing these videos again. If I'm not able to decide based on videos alone, you've given me a lot to ask people about if I need to shop these in store in person. And that's really meaningful. I'm tired of buying memory I'll never use, or buying the wrong product for my needs. I appreciate this more than you can know. Thank you again for all of your hard work.
My pleasure and so glad to hear thikese videos help! I agree with your approach though, access to external storage is a lot faster than it used to be. I’ve edited nearly 400 videos from external SSDs in the last few years. Saved thousands on local storage
A lot of great dedicated work here, but I wish you'd bumped the RAM on the base model to 24GB to eliminate that difference. For example, the LLM output quality rather than speed difference is almost certainly due to having more RAM on the Pro and therefore being able to run a larger version of the model.
The 70B test was definitely not suitable for these machines, as it was slow due to SSD swapping. If you could follow up with the Qwen 2.5 Coder 14B (which should work on both M4 minis), I believe it would be very helpful for many people. 🧠🍄
As someone who uses a lot of After Effects, I definitely think 24 GB RAM is necessary. Question is, do you think M4 Pro is worth the extra $400 vs. M4 Base with 24 GB RAM, where the only difference would be the chip?
for testing LLMs I recommend Ollama, run it with the --verbose flag, and it will give you after each response the actual speed in tokens per second. Otherwise you are not comparing apples to apples if one response is larger.
It's so interesting how different reviewers are getting different results with different configurations. One shows how just bumping the storage to 512gb drastically increased the speed without upgrading the ram on the base model. I struggled with my decision but finally decided to try the same thing while using a fast 2TB external drive to migrate the home folder to so that the internal drive stays relatively open. I guess I'll know if my plan works in a couple weeks. I really wanted the Pro version but the cost was just a few hundred more than my budget allowed so hopefully this config will be second best.
@@jacksprat2333 I ended up going with a Minisforum UM890 Mini PC that I could set up the way I wanted. No regrets. I installed 64 GB of RAM and 6 TB of Nvme SSD storage plus I'm dual booting it with Linux Mint 22 and Windows 10. I mostly use Linux and moved my power PC to my office where I do all my photo and video editing so yes, my plan did work out even though I changed it. I'm pleasantly surprised at the power of the UM890 too. A little beast for sure. Not as good on the Cinebench tests as the Mac but then again I can do hardware changes that I couldn't with the Mac. As a retired IT tech, that's kind of important to me. Thanks for asking!
The quality of an LLM's output does NOT depend on the machine running it. Only the speed will differ. If you ran the exact same flavour of Llama (number of params and quantization) on both Macs, then the difference is only down to chance.
You're not only going to run out of storage space on that base model, ssds don't work well when more than half their space is used, you're not only going to run out of storage you're going to run out of performance! Add AppleCare memory and storage upgrades and this is really not such an amazing deal.
im also rocking a similar TB4 enclosure with a 2TB NVME SSD with my M4 Mac mini and before that with my M1 MacBook Air. absolutely good enough for all tasks I have ever been doing ;)
70B model will probably run on M4 Pro with the most powerful 20 cores GPU and 64 GB RAM should also help but I know in this case VRAM is doing the job. Maybe Thunderbolt 5 and RTX 5090 via eGPU will solve the problem, but the issue in Apple is that if you use external GPU you cannot connect it with internal GPU, so in this case base model wins if you are interested in eGPU.
Good you did a logic test my m2 pro is all I need for audio. It’s normal for logic to not use efficiency cores. If you use some of the alternative DAWs they perform better and use all cores.
The answer to ‘who’s using 300 tracks on logic’ is a British artist called Jacob Collier. He’s got some videos on his RUclips going through the tracks.
Shockingly, 300-track songs are not uncommon. I regularly get up to the 150-200 range on big rock mixes, but those cinema scoring guys go up to around 1000 tracks! Great video! Did you experience any jank with external USB devices? My M1 Max Studio has been really glitchy with external NVMe and monitors disconnecting sometimes.
Thanks Alex. This is the type of video that I have been looking for as I am considering buying a mac mini for my work to run llms locally. Would love to see the results for fine tuning as well. 🧠. Thank you. Also, would you be open to sharing details about your desktop PC build as well? Looks amazing in the video and I am looking to build one myself.
I don't know if I'll go for the pro, or even the base model. But I do know that I appreciate your hard work and "beginner's mind" approach. Kudos. Subscribed. Cheers from overcast Vienna, Scott
Thank you so much! Welcome aboard and I am so glad to hear that. I wanted to really push these bad boys and the help from the community was great! Nearly killed me trying to learn some of this stuff, but I enjoyed that! 😅
in logic, you can actually "freeze" the tracks and it bakes all the plugins into the flat track, and then you can pretty much do another 3x the tracks ;)
6:21 Not all DAWs, but some DAWs including Logic Pro are using only the performance cores for some reason. Reaper on the other hand uses all available cores which is very good and you can actually see a bump in the tracks count or adding more effects etc to the project.
Nicely done! The biggest problem with the base Mac mini isn't just running out of space it's that ssds really need to be limited to half their space used given the nature of how they allocate. So you won't just be running out of storage you'll be running out of performance. 🧠
on logic or ableton the m4 will really move more than 1000 tracks, in the video you are using only 128 samples (minute 06:03 ), anyone today for mixing or mastering (means load tons of plugins) put more than 2400 samples in order to reduce the cpu usage.
@@AlexGTechI would definitely test with both. Lower sample buffer is necessary for live performances! Larger sample buffer means more time between pressing a key on a piano and actually hearing it. Great video!
I was wondering the same thing, and it happens to be linked in the Google doc he linked in the description called "Stress test Scoring Sheet document".
@@rohitbrr I meant to but I stopped when I saw it was a paid project, not sure about paying just to test it for myself. Maybe we should chip in to get it for all of us haha
@@PatoCrioyo I'll do you one better. I got the laptop. A render that originally took me 1 hour on an i7 9700h CPU in an acer predator laptop now takes about 15 minutes now on the m4 pro 14c - 14 inch. laptop
I wanted to try the after effects project to compare it to what my current system does to see before buying my mac mini m4 but in order to get that template you must pay for subscription. Any way you can use a free one in a video that anyone can try themselves?
Lots of those tests were made in applications that really take advantage of the GPU, so of course the M4 Pro outperforms the M4 by a large margin!, but costing more than double. So in the end of the day the base model for 599 is really a great deal!!!!
I greatly appreciate your effort in testing. For the CPU tests on DAW, perhaps you could try using algorithms like IRC4 from a mastering plugin like iZotope Ozone, or tools from UAD Spark such as the Manley Massive Passive EQ MST, or even dozens of API Vision channels.
For the 70B 3 bit quantized, you'll need 64GB RAM. The architecture allows 3/4 RAM usable by GPU, so the model size needs to be below this plus some margin.
The majority of DAWs don't make use of the efficiency cores. As far as I know, the only ones that use ALL cores are Reaper, Cubase/Nuendo, and the latest version of Pro Tools.
Great performance review. Would love to get the base model mini. But could you perhaps include a test video on machine learning tasks using at least YOLOv8 from Ultralytics?
it's a issue with how apple implements multi core audio (see my other comment). not well apparently; it causes dropouts if you use both. Ableton; Logic and some other don't use E-cores at all for audio as a result. Hopefully it gets solved at some point but the point is it seems Apple did an horrible job on their implementation (Ableton now uses a dummy load to force the use of E cores because the performances were WORSE for lighter load is it worked on E-cores; which tells me something is seriously f'ed up). Also ableton developper explained that the priority for them is live application (obviously) so you can't have dropouts.
Can you test a 32gb 512 mac studio as they arebgoing for just $1299 with full warranty faster media engines sd card more gpu cores ports. Comoared to these with the same specs
Thank you! I’ve been learning a lot making these videos, and fascinated by LLMs! Having worked with computers since the 1990s I am shocked by how much things evolved!
Nice video! BTW, apparently (and ironically), Apple's own Logic does NOT take advantage of the economy cores! Maybe on one of their next updates. Pro Tools, Cubase, Reaper and some of the others use them much more. Presonus Studio One not so much.
Crazy right? I wonder why they do that. Even Adobe seems to use all efficiency and performance cores evenly. That was quite noticeable on After Effects…
Great video. Thank you for putting this together. It seems to me that you are using an ultrawide screen with one of Mac Mini you testing. Do you mind sharing a bit about the model and how things are connected? As I am planning to get a Mac Mini to use with my uw monitor. Cheers. Over the course of the video, it looks like an Alienware. :)
could you please do an intensive after effects test and premiere pro test with the Pro :). I was actually intending to buy an MacBook Pro max but for the price I can buy the mini and also a camera and Im trying to figure out whether to go for the mini or not :). But it seems actually pretty powerful !
Thank you for running logic with Kontakt! How did the fan sound? Would super appreciate hearing (pun intended:) if the fan turned on in the M4 base / M4 Pro. p.s. logic only utilizes the performance cores
haha yes it did on both! I actually managed to record the audio here, there is a chapter just before After Effects testing I think. Essentially, it is a lot quieter than my M1 Max MBP which is already pretty quiet. I just answered a similar question, but here is what I experienced: When I was going crazy with it (like rendering and trying to play a game at the same time) the fans did start and I noted the temperature get close to 90 Celsius. The fans span at 3,000 RPM for a few minutes bringing the temp down to about 65C which I thought was pretty impressive as a short burst activity. I am not sure yet how that behaves on a sustained load, say for over 10-15 minutes but the cooling at least at this sort of realistic usage level, is working really well!
Can you please compare the base mini m4 (with pro chipset) with the base MacBook Pro m4 (with pro chipset)…? They both start with the exact same specs.
Only Cubase and Reaper use the efficiency cores. Also when testing tracks in Logic or other you must have something playing on all the tracks. A lot of your tracks had big spaces so only half the tracks were playing.
More to come! Glad you appreciated that! It did feel a little laggy when trying to play back the clip on both machines, but once the pre-render was done it was pretty smooth.
DAW - Only few DAW's use efficiency cores. Maybe 3 of them. Reaper, FL studio and something else as well. know Ableton not use them. Ableton said they create problems, and they do not have plans to use them in the future.
Thanks for making this video. You mention in the video that you are running Stable Diffusion (Im guessing that you had speeded up the video at that point, it cant be that quick right?). You said there was a downloadable document of the timings,? it is of huge interest to me, would you kindly link to where this can be found?
Yes the stable diffusion test was at 13:45, I found the Pro to generate slightly more consistent and accurate results. I forgot to screen record the Pro but the M4 base model example is there 👊🏼
Bro. I’ve been looking for some reviews and bumped into your channel. You look so sincere and kind so it touches me somewhere inside) subscribed! Wish you more RUclipsr’s successes in future, you are definitely worth it)
Thank you so much! I’ve just spent the last 3 days barely getting any sleep, editing videos so this comment put a huge smile on my face, thank you 🥰
I love the humility and effort put into the LLM topic!
🧠
The amount of work you put to make this video happen is much appreciated. 👊🏻
Thank you so much!! ☺️ much more to come
Hats off for giving an amazing well rounded review, this is by far the best mac mini review i have seen so far on the internet, even when compared to big tech reviewers. The community contributions to making this video very relatable and resembling real life use cases was a chefs kiss, extraordinary efforts, i am subscribed to your channel.
What a lovely comment, thank you so much and welcome to the channel!
As a youtuber just starting out I can really really understand and appreciate the amount of effort that you put into these videos. Thank you.
I appreciate that! Yes, it was pretty gruelling this one!
Hey. Thank you so so much for including music production and logic pro. Everyone and their mother neglects stress testing in audio and music even though it is a special test that highlights single core performance above all. Thank you!
Wow, twice the performance with the Pro chip! That’s a massive difference! Great job Alex. You’re the only one I know doing these stress tests.
@@MrOrtmeier thank you, yeah the Pro is crazy good!!
Yes because it has double the performance cores lol
@@AlexGTech what ram did you have in the pro? I've got the base pro with 24gb Ram and i'm bit worried whether it's enough for logic.
@@gt4032 no, the binned has only 2 more cores (6 total)
@@wentdont9 it's not enough for virtualization
I think I found the definitive youtube channel for tech reviews. Thank you so much for considering what people say to you. Fo real, I'm a musician and the device-performances with Logic is really something to me and you got it. Cheers from Chile
🍄 Let me know if you see him too! 😅
This video was a labour of love! No sponsors again, but really hope you enjoy it. I took a lot of your feedback onboard for this one, but I am sure it can be improved even more!
Have a lovely day 🙏🏼
Could you please give me some suggestions for the best docking stations for the mini M4? A docking station that has an SSD slot and some extra pins can help me pump up the memory of the 256gb mini..
Best one I found without spending a fortune is the Caldigit TS04 - I shared a link down below, it’s great even though it’s TB4, but will also look into the Thunderbolt 5 stuff when they come out
@@AlexGTech which mac mini version are you keeping?
You did great. I will note that some of your stress tests are really about memory, and the simple M4 with 32GB RAM is likely to do much better than with 16GB of RAM. With the base RAM level the Mini will do a lot of swapping with these big programs and that is going to slow you down.
@@TheDanEdwards
swapping in RAM, i wonder if onboard storage at 512G or more would make it less dependent on the amount of memory?
Thanks brother this the first video of yours that i watched you are great and i can feel the amount of work you have put in this video❤
100+ tracks with VST’s on the base model is impressive af. Thanks for this
Right?! Especially with all the 3rd party plug ins (about 60 in that example as I duplicated the original project with 20 Kontakt 8 3 times)
@@AlexGTech Kontakt is a sampler and thus doesn't tax the CPU that much -- just a lot of disk i/o, streaming samples from the ssd. Using a synth like Alchemy (built-in, so free once you have Logic) would give a more interesting result imo. Interesting video nonetheless; thx for making it.
when does he mention that 100 tracks is the limit of M4 base model ?
@@chaosme1ster there is a lot of Alchemy in those tracks too!
@@Mpthedawn I found that it was still running past 100 but 200 made it slow down and crackle a bit. I’m guessing the limit in these type of tracks is in that range 100/200 but I need more time to find out when exactly it start to slowdown
Now this is the tests all people want to see. Great work.
So happy that it was useful, I enjoyed putting these videos together despite the hard work, so hearing this makes it all worthwhile 👊🏻
Yet again, fantastic video dude, love watching your videos. Keep up the fantastic work my friend!
Cheers mate, I need some rest now 😂
Excellent video, so great to see you taking on community questions. I have a 2009 Mac Pro, 2 x 6-core 3.46 Xeons, 96GB ECC Ram, 2 TB NVMe, USB 3.2 Gen 2. RX580 8GB. I was waiting for a Mac Studio but don't know if I should just get a M4 or M4 Pro Mac Mini. Looking forward to your next videos on the Mac Mini's. A.
Commenting just because of the efforts it must have taken to create this detailed video. I started RUclips channel, created just 4 videos and the comparison video between iPhone 14 and 15 took such extreme efforts, even though that was much simpler than this. Hats off. Keep doing the great work
Thank you so much!! Yes, this one nearly killed me 😅 I was editing non-stop for like 4 days
Hey Alex! In settings on Logic the i/o buffer is what really matters for track stuttering! (The lower this number, the less latency there is between hitting play and the daw playing back). This gives the daw time to calculate what the plugins are doing. It's generally a great test to put it to the lowest buffer (32) and see how many tracks you can get. When producing you'll often record/make a basic track with the lowest i/o you can and then crank this number up for adding layers and mixing since you don't need realtime playback. At any rate thanks for all the tests this was great to see! You can watch how L.Dre does his tests in Ableton (or iamamusicmogul in logic pro) for a similar example :)
Oh wow, thanks so much! I am learning a lot making these videos! Have a great week ahead 🙏🏼
No matter on what computer you use the LL model. If it's the same models, it will give you the same results. Well, randomly it would be sometimes better, sometimes worse (like in your case). The only real difference is speed. And considering speed, you should've include token tps (tocken per second) metric, because results will vary in volume.
For Blender 3D you should enable both CPU and GPU it uses both and get even faster results.
Not for Mac, it known to slow render times
@@chidorirasenganz I've never had that case. However I would assume maybe a thermal issue might present itself if maybe on the MacBook Air. For example Monster demo file, 1:48 with CPU and GPU, with GPU only 1:58. I ran though all my files and at minimum I see at least a 3%. uplift in render times.I've tried to find the source of what you are bringing up and the only thing I could find is about render tiling, which you would tweak to get scenes to fit in RAM better. As soon as the scene renders and uses tilting then GPU/CPU seems to come into play from my testing.
@ I don’t know if something has changed but previously all the info I’ve seen on using Blender on Mac since the M1 days has been GPU only is faster. I think it was the GPU waiting on the slower CPU to finish rending but I can always try it again on 4.2 but since Eevee Next came out (tbh Eevee in general) I haven’t really care about render times since most of the time it’s 5-10 secs at most
I salute you for all the effort you put into these videos. Outstanding work. Keep it up. Definitely a subscription for me.
Welcome aboard! And thanks so much, means a lot hearing that 🥰
Excellent testing video and analysis. The M4 Pro is clearly for those high-end professionals with no time to lose. High end pressure news reporting, movies/documentary sets with a tight filming budget, and Machine learning on a professional level.
The regular M4 Mac Mini seems to best for those without tight timelines.
Very well put! I am still amazed at how much can be done on the base model M4
Wow this is nice! As a developer mostly working with LLMs I needed to see the LLM usability for this machine. Thank you so much!
My absolute pleasure! I am enjoying learning this stuff! Much more to come, for sure 🙏🏼
Hey, looking to start playing with local Llms too, did you go for the pro or still looking around?
@@DrEmmePhD No. I've decided to get the Macbook M4 Pro when stocks will be available in my country. All due to portability.
You really tried with this video Alex and wow! It payed off! A fantastic effort! Well done!
Thanks so much, I enjoyed learning some of these different apps and can see why this machine has such potential! Much more to come, but for now I need a rest 😂
Incredible video great content Alex well done on your hard work 😊😊
knocked out of the park with your last two vids. Congrats, man!
Thank you so much! Loved putting these together! More to come for sure, but I need to make some easier videos for a little while 😂
The effortss 🤯 Thanks a lot sir!
Please test Minis in Adobe dtp triad: PS, AI and ID. I'm thinking about getting Mini for myself but I'm not sure if I should pickup base M4 or base Pro.
I'd like solid machine to work on 2D product visualisations, design 2D graphics and put everything into 50-100 pages catalogs. This is what I do professionally.
Good stuff Alex…nice to see clear and humble - non shouty - info. As a heavy LogicPro/NI/Arturia user I’m keen to see the music tests of course. Keep it up. JPMusic.
Thank you so much! Man, I am HOOKED! I have no clue what I am doing in Logic, but used to use Ableton back in the day, so there was a little muscle memory in there. Really want to bring some keys into the studio and have a Mac mini setup to mess around with it. I won’t be hitting those track numbers, that’s for sure 😂
Wow! That was a very good and informative review. Thank you very much.
Excellent testing. Well done and thanks for sharing
Thank you very much! I need some sleep now 😂
The thing about most music programs, from what I've seen on another RUclipsr's (James Zhan) tests along with my own tests too, is that they dont utilize the efficiency cores AT ALL. It kind of made me wonder when I saw Reaper utilizing all the cores, but Logic didn’t.
Yeah, I am glad I am not going crazy! I did wonder for a while, if I had made a mistake in the settings. Will check that James guy out, I am learning so much making these videos! Love it!
@AlexGTech 😂 you're not alone, I thought I was too! And do check him out. He's been doing comparison videos on all the Apple Silicon Macbook Pros since the M1, so his videos will make for great references. Keep up with the great work yourself, too, man! 💪
Very useful test overall, thanks for the video and all the effort!
Two suggestions on the LLM part:
- The power of the chip, like M4 vs M4 Pro won't affect the quality, any such result is coincidence :)
- The chip's speed will affect how fast the LLM can generate the answer, the usual benchmark for this is in token output / seconds. There is a built in 'randomness'. Answers to the same prompt will be different each time, so measuring completion speed doesn't make sense.
Keep it up, great stuff!
thank you for confirming that, I thought it was a bit counterintuitive, and perhaps I had a 'processor' bias while I was doing the tests and my brain played a trick on me! The stable diffusion test was really odd though, how out of 10 pictures prompting for a hot air balloon and birds, the Pro generated 4 including both and the standard M4 only generated 1 picture with a bird, but nothing with hot air balloons. I guess I need to run multiple tests though to really see if it was a 'one-off' or if having a beefier CPU/GPU makes a difference
How about displaying frame rates on your gaming tests? Also would love to see a max RAM configuration LLM comparison (32gb base/64gb pro) 🧠
Definitely! I tried to enable that via the command line using Apple’s own bloody Metal code, but only steam games worked and wasn’t even Apple’s thing, it was just Steams basic counter. Really annoying. And I ran out of time to get an FPS counter app, but will research into a good one 👊🏼
Unparalleled test! Awesome job.
Appreciate the effort. Impressed that you’ve included ollama 👏
thank you! Not gonna lie, it's been a journey learning some of this stuff, but I am so glad I did - and all from the help of the community here 🥰
🧠 An LLM test would be great! For people trying to use the mac mini for more academic work :)
This was a great video, you're very humble man. I'd love to see a comparision between m4 mac mini vs a similarly priced windows laptop and how it performs in both gaming and editing
Thanks so much, glad you enjoyed it! I will be doing a comparison with another Windows mini PC very soon, but I like the laptop idea too!
Really helpful so far. I have been convinced that I needed the Pro MAC Mini to produce, edit and upload professional RUclips videos. Plus podcasts. But you are making me think I may be able use the base model and save myself some serious bucks. I'm going to enjoy your docking station video too. Great job!
I'll end up viewing these videos again. If I'm not able to decide based on videos alone, you've given me a lot to ask people about if I need to shop these in store in person. And that's really meaningful. I'm tired of buying memory I'll never use, or buying the wrong product for my needs. I appreciate this more than you can know. Thank you again for all of your hard work.
My pleasure and so glad to hear thikese videos help! I agree with your approach though, access to external storage is a lot faster than it used to be. I’ve edited nearly 400 videos from external SSDs in the last few years. Saved thousands on local storage
excellent content, this is great information for so many pros!
A lot of great dedicated work here, but I wish you'd bumped the RAM on the base model to 24GB to eliminate that difference. For example, the LLM output quality rather than speed difference is almost certainly due to having more RAM on the Pro and therefore being able to run a larger version of the model.
The 70B test was definitely not suitable for these machines, as it was slow due to SSD swapping. If you could follow up with the Qwen 2.5 Coder 14B (which should work on both M4 minis), I believe it would be very helpful for many people. 🧠🍄
Thanks, great suggestion! I did wonder about a 13B or similar size so thanks for that! And noted on the fine tuning, will give this a proper go soon!
As someone who uses a lot of After Effects, I definitely think 24 GB RAM is necessary. Question is, do you think M4 Pro is worth the extra $400 vs. M4 Base with 24 GB RAM, where the only difference would be the chip?
for testing LLMs I recommend Ollama, run it with the --verbose flag, and it will give you after each response the actual speed in tokens per second. Otherwise you are not comparing apples to apples if one response is larger.
It's so interesting how different reviewers are getting different results with different configurations. One shows how just bumping the storage to 512gb drastically increased the speed without upgrading the ram on the base model. I struggled with my decision but finally decided to try the same thing while using a fast 2TB external drive to migrate the home folder to so that the internal drive stays relatively open. I guess I'll know if my plan works in a couple weeks. I really wanted the Pro version but the cost was just a few hundred more than my budget allowed so hopefully this config will be second best.
Did your plan work?
@@jacksprat2333 I ended up going with a Minisforum UM890 Mini PC that I could set up the way I wanted. No regrets. I installed 64 GB of RAM and 6 TB of Nvme SSD storage plus I'm dual booting it with Linux Mint 22 and Windows 10. I mostly use Linux and moved my power PC to my office where I do all my photo and video editing so yes, my plan did work out even though I changed it. I'm pleasantly surprised at the power of the UM890 too. A little beast for sure. Not as good on the Cinebench tests as the Mac but then again I can do hardware changes that I couldn't with the Mac. As a retired IT tech, that's kind of important to me. Thanks for asking!
Superb. Thank you for the insights.
The quality of an LLM's output does NOT depend on the machine running it. Only the speed will differ. If you ran the exact same flavour of Llama (number of params and quantization) on both Macs, then the difference is only down to chance.
Was looking for this comment.
🧠🧠🧠🧠🧠🧠🧠🧠
THANK YOU!!!
first video ive seen Alex is a top G
You're not only going to run out of storage space on that base model, ssds don't work well when more than half their space is used, you're not only going to run out of storage you're going to run out of performance! Add AppleCare memory and storage upgrades and this is really not such an amazing deal.
So what would you advise to avoid that issue?
Thanks for your incredible job Alex..Really appreciated and made things clear 😍😍
So happy to hear that! More to come 👊🏼 have a lovely week ahead
❤ really love your work
im also rocking a similar TB4 enclosure with a 2TB NVME SSD with my M4 Mac mini and before that with my M1 MacBook Air. absolutely good enough for all tasks I have ever been doing ;)
70B model will probably run on M4 Pro with the most powerful 20 cores GPU and 64 GB RAM should also help but I know in this case VRAM is doing the job. Maybe Thunderbolt 5 and RTX 5090 via eGPU will solve the problem, but the issue in Apple is that if you use external GPU you cannot connect it with internal GPU, so in this case base model wins if you are interested in eGPU.
I rarely comment any videos, but I am just sincerely impressed by the amount of effort you put into these videos! Hats off to you sir!
Thank you very much! That really means a lot. These were brutal to create in terms of effort but I’m glad it’s been useful. Have a lovely day
Good you did a logic test my m2 pro is all I need for audio. It’s normal for logic to not use efficiency cores. If you use some of the alternative DAWs they perform better and use all cores.
This is exactly what I needed to see!
Thanks mate! These are powerful little beasts!
The answer to ‘who’s using 300 tracks on logic’ is a British artist called Jacob Collier. He’s got some videos on his RUclips going through the tracks.
Shockingly, 300-track songs are not uncommon. I regularly get up to the 150-200 range on big rock mixes, but those cinema scoring guys go up to around 1000 tracks! Great video! Did you experience any jank with external USB devices? My M1 Max Studio has been really glitchy with external NVMe and monitors disconnecting sometimes.
Thanks Alex. This is the type of video that I have been looking for as I am considering buying a mac mini for my work to run llms locally. Would love to see the results for fine tuning as well. 🧠. Thank you. Also, would you be open to sharing details about your desktop PC build as well? Looks amazing in the video and I am looking to build one myself.
ngl one of the few creators that produces the best video in youtube
Thank you so much, what a lovely comment!
I don't know if I'll go for the pro, or even the base model. But I do know that I appreciate your hard work and "beginner's mind" approach.
Kudos. Subscribed. Cheers from overcast Vienna, Scott
Thank you so much! Welcome aboard and I am so glad to hear that. I wanted to really push these bad boys and the help from the community was great! Nearly killed me trying to learn some of this stuff, but I enjoyed that! 😅
in logic, you can actually "freeze" the tracks and it bakes all the plugins into the flat track, and then you can pretty much do another 3x the tracks ;)
Oh nice! I’m learning so much doing this! Thanks for sharing that
6:21 Not all DAWs, but some DAWs including Logic Pro are using only the performance cores for some reason. Reaper on the other hand uses all available cores which is very good and you can actually see a bump in the tracks count or adding more effects etc to the project.
@@TYLERBALLADSI think protools also use all cores on their latest versions. Not sure about cubase
I like your passion and humility.
Nicely done! The biggest problem with the base Mac mini isn't just running out of space it's that ssds really need to be limited to half their space used given the nature of how they allocate. So you won't just be running out of storage you'll be running out of performance. 🧠
on logic or ableton the m4 will really move more than 1000 tracks, in the video you are using only 128 samples (minute 06:03 ), anyone today for mixing or mastering (means load tons of plugins) put more than 2400 samples in order to reduce the cpu usage.
Thank you, I left that setting alone as I wasn’t sure what it would do, will give that a go for the next test! 🙏
@@AlexGTechI would definitely test with both. Lower sample buffer is necessary for live performances! Larger sample buffer means more time between pressing a key on a piano and actually hearing it. Great video!
Also, can one use an existing old iMac as a monitor for the mini?
I would like to try rendering that After Effects project and compare it with my old setup. Can you please put the link to the project file, please?
I was wondering the same thing, and it happens to be linked in the Google doc he linked in the description called "Stress test Scoring Sheet document".
@@PatoCrioyo I’m not sure if I didn't saw that earlier ago, thanks for pointing out!
Did you guys test it out and get any results?
@@rohitbrr I meant to but I stopped when I saw it was a paid project, not sure about paying just to test it for myself. Maybe we should chip in to get it for all of us haha
@@PatoCrioyo I'll do you one better. I got the laptop. A render that originally took me 1 hour on an i7 9700h CPU in an acer predator laptop now takes about 15 minutes now on the m4 pro 14c - 14 inch. laptop
Chatgpt suggested me your video as a source and reference when I questioned it about m4 vs m4pro
I wanted to try the after effects project to compare it to what my current system does to see before buying my mac mini m4 but in order to get that template you must pay for subscription. Any way you can use a free one in a video that anyone can try themselves?
Lots of those tests were made in applications that really take advantage of the GPU, so of course the M4 Pro outperforms the M4 by a large margin!, but costing more than double. So in the end of the day the base model for 599 is really a great deal!!!!
Thank you so much for this video! It was really helpful as I've been considering buying a Mac Mini. I truly appreciate your detailed review.
You are so welcome! These were hard work to produce but I’m glad I helped 😊
Thanks for the noise update, really missed that in the previous vid.
Cool, thanks for noticing it - very rare to hear the fans but when they come on it’s very gentle
I greatly appreciate your effort in testing. For the CPU tests on DAW, perhaps you could try using algorithms like IRC4 from a mastering plugin like iZotope Ozone, or tools from UAD Spark such as the Manley Massive Passive EQ MST, or even dozens of API Vision channels.
Thank you, great suggestions! I am hooked into this now 😂
Awesome Video, much appreciated for the LLM benchmarks. Would like to see the spreadsheet containing the comparisons
My pleasure! I’ve put a link in the description to it. Still updating some results but the LLM ones are there 👊🏼
For the 70B 3 bit quantized, you'll need 64GB RAM. The architecture allows 3/4 RAM usable by GPU, so the model size needs to be below this plus some margin.
Thank you! Makes sense, I did realise it was probably RAM as 19GB swap is crazy 😂
The majority of DAWs don't make use of the efficiency cores. As far as I know, the only ones that use ALL cores are Reaper, Cubase/Nuendo, and the latest version of Pro Tools.
I beg to differ - the majority of DAWs seem to DO use the e-cores, while Logic pretty much is the odd one out...
Reaper Cubendo PT are the best anyway 😅
@@nukenoly8751 At least, that's true for one of them. 😉
Great performance review. Would love to get the base model mini. But could you perhaps include a test video on machine learning tasks using at least YOLOv8 from Ultralytics?
Great suggestion! Thank you, will do!
Great video Alex
Thank you! Enjoyed making this one!
Logic was made. Not to use effeciency cores on tracks.
Cubase Pro tools and Reaper do.
Studio one and albelton need more optimisation
interesting! I did wonder about other DAWs. Cubase sounds interesting, hope I can give that a whirl soon
it's a issue with how apple implements multi core audio (see my other comment). not well apparently; it causes dropouts if you use both. Ableton; Logic and some other don't use E-cores at all for audio as a result. Hopefully it gets solved at some point but the point is it seems Apple did an horrible job on their implementation (Ableton now uses a dummy load to force the use of E cores because the performances were WORSE for lighter load is it worked on E-cores; which tells me something is seriously f'ed up). Also ableton developper explained that the priority for them is live application (obviously) so you can't have dropouts.
Can you test a 32gb 512 mac studio as they arebgoing for just $1299 with full warranty faster media engines sd card more gpu cores ports.
Comoared to these with the same specs
🧠Looking forward to use your videos to get started with LLm's
Thank you! I’ve been learning a lot making these videos, and fascinated by LLMs! Having worked with computers since the 1990s I am shocked by how much things evolved!
Nice video! BTW, apparently (and ironically), Apple's own Logic does NOT take advantage of the economy cores! Maybe on one of their next updates. Pro Tools, Cubase, Reaper and some of the others use them much more. Presonus Studio One not so much.
Crazy right? I wonder why they do that. Even Adobe seems to use all efficiency and performance cores evenly. That was quite noticeable on After Effects…
Great video. Thank you for putting this together. It seems to me that you are using an ultrawide screen with one of Mac Mini you testing. Do you mind sharing a bit about the model and how things are connected? As I am planning to get a Mac Mini to use with my uw monitor. Cheers.
Over the course of the video, it looks like an Alienware. :)
Yeah it’s the 34” Alienware, I left a link in the description- I also used an LG at times, both linked in there for you.
@@AlexGTech Thanks. Are they connected just via HDMI or a dock?
🧠. VERY well done!
Leaving the like before watch! GREAT CONTENT!!!
You're the best, thank you!!
Best review 🎉
Thank you!!
could you please do an intensive after effects test and premiere pro test with the Pro :). I was actually intending to buy an MacBook Pro max but for the price I can buy the mini and also a camera and Im trying to figure out whether to go for the mini or not :). But it seems actually pretty powerful !
James Zhan has a great video for Audio Production 🔊
Thank you for running logic with Kontakt! How did the fan sound? Would super appreciate hearing (pun intended:) if the fan turned on in the M4 base / M4 Pro.
p.s. logic only utilizes the performance cores
haha yes it did on both! I actually managed to record the audio here, there is a chapter just before After Effects testing I think. Essentially, it is a lot quieter than my M1 Max MBP which is already pretty quiet. I just answered a similar question, but here is what I experienced:
When I was going crazy with it (like rendering and trying to play a game at the same time) the fans did start and I noted the temperature get close to 90 Celsius. The fans span at 3,000 RPM for a few minutes bringing the temp down to about 65C which I thought was pretty impressive as a short burst activity. I am not sure yet how that behaves on a sustained load, say for over 10-15 minutes but the cooling at least at this sort of realistic usage level, is working really well!
Hi Alex. It would be great to see Comfy UI running Flux on the Minis 🧠
Great suggestion! Will take a look at that for sure !
Can you please compare the base mini m4 (with pro chipset) with the base MacBook Pro m4 (with pro chipset)…? They both start with the exact same specs.
get the 4 bit MLX version for apple (smaller and almost lossless compression) , and tokens per second is a good metric , for llm performance
Only Cubase and Reaper use the efficiency cores. Also when testing tracks in Logic or other you must have something playing on all the tracks. A lot of your tracks had big spaces so only half the tracks were playing.
Crazy isn’t it, considering it’s Apple software? And thanks for the tip on the tracks, will fix that for the next one!
Thank you for this quality video, sir
Thank you for including After Effects! So rare to find these in other tests in RUclips.
More to come! Glad you appreciated that! It did feel a little laggy when trying to play back the clip on both machines, but once the pre-render was done it was pretty smooth.
oh woah, the thumbnail made me think they were like a deep-ish blue colour - looks really nice !
They’re very photogenic, lovely machines
Can you retest yhe render on m1 max without blender running
DAW - Only few DAW's use efficiency cores. Maybe 3 of them. Reaper, FL studio and something else as well. know Ableton not use them. Ableton said they create problems, and they do not have plans to use them in the future.
Thanks for telling us the average fan speed. I use iStatistica Pro and wanted to know the fan speed of the M4 Pro Mac mini.
Could you conduct tests for the proxy media renders in Final Cut Pro?
Thanks for making this video. You mention in the video that you are running Stable Diffusion (Im guessing that you had speeded up the video at that point, it cant be that quick right?). You said there was a downloadable document of the timings,? it is of huge interest to me, would you kindly link to where this can be found?
Yes the stable diffusion test was at 13:45, I found the Pro to generate slightly more consistent and accurate results. I forgot to screen record the Pro but the M4 base model example is there 👊🏼