I guess the sudden demand is due to OpenClaw? But most people will still use cloud LLMs, right? Anything particular with the Mac Mini that non-Mac lack?
back
6 comments
Not just OpenClaw. The Mac mini is just stupidly good value for a desktop computer, and the RAM prices have only enhanced its appeal.
Apple doesn't make much of a fuss about it but their chip performance is laughably ahead of the other chipmakers.
The Mac Mini M4 gets a score of 3788 in Geekbench[0]. The top of the PC processor chart is 3395[1]. It's not even Apple's latest chip!
PC processors can only keep up by adding more cores, but real world performance in many workloads is enhanced by having a smaller number of higher performance cores.
Geekbench is basically trash. People keep using it for comparing Mac performance because many of the things people usually benchmark don't run on Macs.
But single-number outputs like that are useless. Is the number ~10% higher because it's consistently ~10% faster at everything, or because it's 100% faster on a minority of things and slower at everything else? The first one is pretty unlikely when comparing processors with different designs, and indeed that isn't it:
https://www.phoronix.com/review/apple-m4-intel-amd-linux/4
https://www.phoronix.com/review/apple-m4-intel-amd-linux/5
https://www.phoronix.com/review/apple-m4-intel-amd-linux/6
https://www.phoronix.com/review/apple-m4-intel-amd-linux/7
The CPU in those charts with a similar TDP to the M4 is the Ryzen HX 370. You can see that the M4 is ahead of it in a few of the tests (C-Ray, DuckDB, PyBench, FLAC) but in even more of them the M4 is at the bottom of the stack. (Only a third of those charts are actually performance; each performance chart is followed by two power consumption charts.)
And the ~20W TDP is a nice parlor trick (the HX 370 is the only one on the list that competes with it there) but in a desktop CPU that's pretty irrelevant. Whereas if you compare it to the CPUs that can be had for a similar price (e.g. Ryzen 9700X, 65W), it's only ahead in C-Ray and FLAC while losing quite badly in most of the others and subjecting you to unupgradable soldered memory that the PC hardware doesn't.
Meanwhile doing ray tracing on a CPU instead of a GPU isn't much fun, and FLAC is an audio codec so a ~10% improvement there is probably not going to be a big part of your day if you're not a full-time sound engineer. So does averaging those kinds of things in to make a single benchmark number make sense? Or should you be looking at the results on applications you actually use?
How are we supposed to trust these charts when it can't even be bothered to specify which Apple Silicon chip it's testing? The Mac Mini comes in two versions.
The different versions have different names. One is called M4, the other is called M4 Pro. The name tells you they tested the former and it has only the one CPU configuration.
The Pro has more P-cores and fewer E-cores (8/4 or 10/4 instead of 4/6), but even the 10/4 configuration starts at $1399 instead of $799 and for the extra $600 you can move from the 9700X to the 9950X3D (16 P-cores) and have $200 left over.
If you remove the Mac filter, its performance is not even in the top ten
Which is obvious if you spent more then half a microsecond thinking about it, because apple silicone barely draws any power - it's performance is fantastic in it's niche, which is squarely within what a home user cares about - but it's not leading on benchmark performance, because that's not what apple designed it for
The reason its coincidentally good for local ai inference is also just down to the fact the embedded GPU has shared memory access to the system VRAM. That means low performance/throughput but large memory.
Which is great for home use, but once again not gonna top charts.
Which top 10 are you talking about? If you mean the top absolute geekbench scores, those are always with the assistance of cryogenic cooling.
Sure, the top ten may be using highly advanced cooling.
And the fact that's possible should've already proven that Apple's decided on trade-offs that did not enable bleeding Edge performance, hence not going to top benchmarks.
But aside from that amount of transfer ability you should've been able to manage, you're ignoring that apple silicone is still being beat on all performance benchmarks even with stock settings.
Apple chose a performance profile for their chips, and it's not "highest performance while sacrificing cooling and energy usage". Others did. And apple did well not chasing benchmarks, as that'd be the epitome of idiocity for their target market. They're not targeting high performance servers with massive cooling setups. They're targeting mobile workstation and entertainment devices.
They do not have any need for bleeding Edge performance trade-offs. They need power efficiency and enough performance to feel snappy on all workloads people will run on these devices - which isn't benchmarks. because none of their users _need_ highly sustained processing power. its just not something they'd ever target.
and im not even adressing the fact that geekbench is notorius for being absolutely shit at showing actual processing power.
Mac mini has first-class access to iCloud, photos, iMessage etc. So if you are deep in the Apple ecosystem you might prefer it for that reason. I have a windows gaming desktop that I could use as a server for openclaw/cowork but I realized I simply don’t trust that system enough to give it access to all the personal stuff I’m giving to the AI. I trust Anthropic and Apple. I don’t trust whatever junk is running on my gaming desktop.
If you want to run local models, another advantage is Apple’s unified memory architecture. The biggest Mac mini has 64gb ram and Mac Studio has up to 512gb. Compare this little box to what monster Nvidia gpu system you would have to buy to get the same memory there. And how much your PG&E bill would go up. That doesn’t account for the shortage of basic $600 Mac minis though.
An M4 mini is overkill just to run OpenClaw. I'm running it on a Pentium J5005 and it's running 20 other services in Docker. I think the main thing was many wanted it to be able to access iMessage. I think people dream of also using the mac to run the LLM but the 16gb ones don't have enough ram.
You can run nullclaw etc on a Pi zero. People who are paying big $ are mostly trying to run local LLMs.
Personally, I would rather pay a few bucks for Qwen or just use gemma4 which runs on a potato. But I guess we are all different.
The shortage is for the 512, 256, and 128 models.
The basic 16GB Mac mini is also hard to buy. I bought one used not to save money but because I couldn’t find any store online with it in stock.
Those are the ones that can run the LLMs. Not a coincidence.
When they say 'due to openclaw' they refer to running AI models that openclaw uses, not to openclaw itself.
People are running openclown on microcontrollers.
My understanding is that openclaw is only a factor, and a relatively minor one.
Most likely the limiting factor is the crunch that chip companies are going through.
You can look up benchmarks. It's different depending on the model of Mac Mini and Model of LLM.
The take away is that some of the Apple hardware hits a sweet spot for performance and price which may change in the future but for now it's causing a lot of demand so people can run inference without GPUs.
Also Macs keep a lot of their resale value so you can use them for a while and then sell them for sometimes 80% of their original value.
Affordable ram!
I recently bought one for my k3s cluster, and it was the cheapest 16g ram I could get by a decent margin.