Apple Mac mini and Mac Studio Aim to be AI Powerhouses
How do the new systems and the M5 Ultra and M6 chips stack up?Buying Guides
Photo: Jacob Bobo
On August 25 Apple announced two new desktop Macs with two new Apple M-series chips: a Mac mini with Apple M6 and a Mac Studio with M5 Ultra. They both arrived under a headline that promises a “big leap in performance and Ai compute,” and both are impressive in their own way.
Both provide performance gains, and one is especially capable of running the best local models at high speeds with a large context window.
The specs say a lot
Decoding the specifications of a modern Mac can be a little tricky at times, as Apple's chips include various types of CPU cores, varying GPU options, and have varying amounts of supported memory and memory bandwidth. But this, thankfully, is one case where the spec sheet is illuminating at a glance.
| CPU Super Cores | CPU Performance cores | CPU Efficient Cores | GPU Cores | Neural Engine | Max. Memory | Memory Bandwidth | |
|---|---|---|---|---|---|---|---|
| Mac mini M6 | 2 | 4 | 6 | 12 | Dual 16-core | 32GB | 153GB/s to 170GB/s |
| Mac Studio M5 Ultra | 10/12 | 20/24 | 0 | 64/80 | 32-core | 512GB | 1.2TB/s |
Clearly, these machines are in different leagues. The Mac Studio M5 Ultra has up to 3x more CPU cores, over 6x more GPU cores, up to 16x more memory, and up to 7x more memory bandwidth.
The only area where they sorta seem to trade blows is in the Neural Engine, where the two have the same number of Neural Engine cores, but in a different configuration. However, at this point it's not really clear how the Neural Engine will impact overall performance. Prior Apple chips also had NPUs, and while they're used for some AI features across different software, generative AI (like LLMs and image diffusion models) generally run on the GPU.
You'd be right to think this means the Mac Studio M5 Ultra is just strictly better at everything, from AI to games to creative workloads. But that's also obviously not a fair comparison, because the pricing gap between the two Macs is just as large as the gap in their specifications. The Mac mini M6 starts at $899 and goes up to $2,299. The Mac Studio M5 Ultra starts at $5,499 and goes up from there.
Apple's Mac mini M6: a good daily driver
The Apple Mac mini remains one of the best all-around budget desktops that money can buy. It provides great performance in a compact footprint, and it's still a good deal for most people. My daily driver laptop is a MacBook Air M4, and I have used it not just for writing and editing but also for 3D modelling and 4K content creation. The Mac mini M6 has two additional CPU cores and two additional GPU cores compared to the M4, so it's likely at least 20% quicker in multi-threaded workloads.
However, the Mac mini M6 isn't going to provide an "extraordinarily leap" in AI performance, at least, not in the entry-level launch configuration. This is mostly because of the launch model's memory configuration. At launch, the Mac mini M6 is only available with up to 32GB of RAM and provides a maximum memory bandwidth of 170GB/s.
While 32GB of unified memory is enough to load a decent quantized version of a leading large language model, like Qwen3.8 27B, it's going to be a little tight, and you're not going to be able to go any bigger than that. And once you load it, you'll no doubt see that performance is slow, as there's not enough memory bandwidth to efficiently move model weights in and out of the M6's compute hardware.
The Mac Studio M5 Ultra looks like the real deal
Apple's Mac Studio M5 Ultra is the real deal for AI. Some of that has to do with its CPU and GPU configuration, but its memory capacity and bandwidth are the heart of the M5 Ultra's improvements, at least when it comes to AI workloads.
The Mac Studio M5 Ultra's memory bandwidth of 1.2/TBs is truly incredible. That's more memory bandwidth than an NVIDIA RTX 5080 (960GB/s) and nearly twice the memory bandwidth of AMD's Radeon 9070 XT (644GB/s). The Mac Studio M5 Ultra has roughly 5x more memory bandwidth than the NVIDIA DGX Spark (273GB/s) and AMD Ryzen AI Halo (256GB/s). The Mac Studio M5 can also be configured with up to 512GB of total unified memory, which is a lot more than any of the alternatives I just mentioned.
Apple states that the M5 Ultra's GPU cores include a Neural Accelerator in each core which provides up to a 4.5x increase in AI performance over the M3 Ultra. The exact improvement Apple's new GPU cores provide is going to be critical, because this was the M3 Ultra's weakness. While it was a good choice for local generative AI thanks to its large amounts of memory and memory bandwidth (819GB/s), the M3 Ultra's GPU compute performance was a bottleneck.
You could load a model with hundreds of billions of parameters on an M3 Ultra with 256GB or 512GB of memory, but the tokens-per-second output would be in the single digits or the teens. That's ok, but I'd like to see the M5 Ultra roughly double that, bringing the tokens-per-second output up to 30 or 40.
If the M5 Ultra can help the Mac Studio achieve that, well, it'll be a huge improvement to an already impressive machine. And, personally, I suspect it will in fact deliver a nice uptick to performance, as both the Apple M4 and M5 chip generations delivered significant gains to GPU and NPU performance. The M5 Ultra will provide the same benefits in a larger chip, and I suspect it will ultimately make the M5 Ultra the best consumer PC for AI.
AI workloads aside, it's certainly true that the Mac Studio will be a powerhouse, and creative work remains its strength. This is due to the Mac Studio's combination of powerful CPU and GPU cores alongside lots of memory capacity. This is not merely the Mac that you buy for editing 4K video or editing photos from common digital cameras. This is the Mac you buy if you shoot in 6K or 8K, or you're stitching together multiple photos that add up to hundreds of megapixels. For creative pros, its ability to handle generative AI locally will just be the icing on the cake.
More from MC News
- Run AI Locally: The Best LLMs for 8GB, 16GB, 32GB Memory and Beyond
- How to Use Local AI With Your Own Files: A Beginner's Guide to RAG
- Reading vs. Answering: The Two Hardware Bottlenecks Behind Local AI Performance
- How to Use a Local LLM in Visual Studio Code
- Hands-On with the AMD Ryzen AI Halo
- Quantization Explained: Why the Same LLM Gives Better Results on High-End Hardware
- Why Context Length, Not AI Model Size, Is Your Biggest VRAM Problem
- How to Use LM Studio to turn your PC into an AI Server
- Why VRAM and Memory Bandwidth are Key for Powering Local AI
- Keyboard 101: Intro to Computer Keyboards
- Fix It Yourself: Talking to iFixit on Why Repairable Tech Matters
- The Local LLM Your Laptop's Been Waiting For
- Gemma vs Qwen: Which Local AI is Best?
Matthew S. Smith is a prolific tech journalist, critic, product reviewer, and influencer from Portland, Oregon. Over 16 years covering tech he has reviewed thousands of PC laptops, desktops, monitors, and other consumer gadgets for publications including PC World, Reviewed, PC Mag, MIT Tech Review, The Verge, Ars Technica, Wired, and Business Insider. Matthew also hosts Computer Gaming Yesterday, a YouTube channel dedicated to retro PC gaming, and covers the latest artificial intelligence research for IEEE Spectrum.
