Mac Studio CLUSTER vs M3 Ultra 🤯

The M3 Ultra and M4 Max Mac Studios working together as a cluster.

Use COUPON: ZISKIND10

🛒 Gear Links 🛒
⌨️🏢 The rack in this video:

🎥 Related Videos 🎥

— — — — — — — — —

❤️ SUBSCRIBE TO MY YOUTUBE CHANNEL 📺

— — — — — — — — —

Join this channel to get access to perks:

— — — — — — — — —

#macstudio #m4max #m3ultra

[jeevans]

Apple : who are buying most number of macs?
Alex : heheheheheheheheehe😊

[Singulainthony]

Your hard work on this one is evident, and it's really well done. This is exactly the kind of content that keeps bringing me back to your channel... Keep up the great work!

[banananar]

Greatest video ever! Cannot find better explanation and passion on the topic anywhere else! This guy's settup costs like 25-30 000 dollars just for this youtube video and even more to manage it. True legend

[SomeshDiwan]

I've been watching your videos for months now—they're consistently amazing! 🙌 I really appreciated how clearly you broke down the performance and value differences between the Mac Studio cluster and the M3 Ultra. Your insights make it so much easier to understand which setup suits different workflows. Keep up the outstanding work, Alex—your content is incredibly helpful and always a pleasure to watch!

[SunSin]

You have outdone yourself Alex. This is prime time quality content !

[sultanalsharfi]

I really admire the effort you put into your videos beyond just demoing happy day scenarios and flashy production tricks.. and that you don’t hesitate to pull up your sleeves and code when needed. Thanks for the great content

[pratyushkumar]

hello brother.. i love your content... i am a high school student who got his interest in AI and LLMS all thanks to you brotherr

[williamwilliam]

As a student learning ML/AI, I find your videos incredibly interesting and inspiring. Thank you for creating them, please keep them coming!

[KG_BM]

im too slow to understand any of this but I still enjoy showing up and pretending to :) super cool stuff

[Fawzi.Mahmoud]

Total respect that you're putting the time money and effort and you're sharing that with all of us completely free definitely deserve like and subscribe

[thomasmozdzen]

Finally the video I have been waiting for weeks / months / years! Thank you and please upload more for us mere peasants

[JoseAngelAcostaEngineering]

You should try thunderbolt networking "mesh" instead hub, more complex and requires a tb5 cable among each pair of Mac but for LLM means much faster non blocking dedicated link it will deliver a performance ever closer to M3 ultra 512, even you can grow the cluster to 5 machines without losing direct link among mesh thunderbolt networking

[PetriKrohn]

Pipelining a model between 4 nodes (or GPUs) only uses each node 25% of the time. For optimal performance and full utilization, you need to do tensor parallelism. LLM inference is bandwidth limited, A cluster of four M4 Max (4 x 526 GB/s) should be two to three times faster than a single M3 Ultra (819 GB/s). Distributed-llama can do tensor parallelism, but only works on the CPU. It might still work faster than the pipelined setup.

[alexexoxoxo]

Awesome video! We're going to make these heterogeneous setups super easy and stable with EXO in the upcoming version (soonTM)

[pandasambit]

Really appreciate the hard work, dedication and honesty you put into all of these videos. I never find such detailed walkthroughs anywhere else. Really appreciate this. Thanks.

[benoitferland]

Great video Alex! I’m just a part time geek and I love your channel. Keep up the great work! ❤

[marqgb]

Just ordered my M3U 512GB, 4TB SSD. I’m all in on running Local LLMs, I’m a serial entrepreneur, not a programmer or engineer, but want to explore AI before everyone else… have lots of plans, including making my own AI server and iPhone app and calling back to machine instead of cloud… don’t edit video or pictures, strictly LLM, was $11k…. Very few have this machine… I just subscribed, hopefully we can collaborate!

[MarcSpctr]

now get 4 MacStudio M3 Ultra with 512GB eaxh and try to run full DeepSeek r1 non-quantized.

[redthunder]

I just clicked on the video, but PLEASE run llama 4 on the mac m3 ultra.

It is a perfect fit for the m3, 109 billion parameters leaves tons of memory for the crazy context length it has, and with only 17 billion active parameters, it should be blazingly fast for its size.

[kirillkasyanov]

Alex is like crazy professor for me - “what if we take most expensive apple devices and make a cluster from them and ask them to tell a story!!!”😂