Mac Studio M5 Ultra hits 4.8TB/s bandwidth when four units are clustered

Apple unveiled the new Mac Studio and Mac mini side by side on August 25. The Mac Studio comes with the M5 Max and the newly built M5 Ultra, while the Mac mini gets Apple's first 2-nanometer chip, the M6, along with the M5 Pro. Both products are open for preorder starting today, with actual shipping beginning September 22.

Mac Studio targets on-device LLMs with M5 Ultra

The Mac Studio is Apple's flagship desktop aimed at running large language models locally. The base configuration with the M5 Max packs an 18-core CPU, up to a 40-core GPU, and up to 128GB of unified memory. Apple says embedding a neural accelerator in each GPU core makes AI computation up to 3.9 times faster than the previous generation.

The higher-end M5 Ultra configuration is the first to use a quad-die design, fusing four dies together. It offers up to a 36-core CPU, up to an 80-core GPU, and unified memory scaled up to 512GB, with memory bandwidth reaching 1.2TB/s — 50% higher than the previous generation. Apple says this chip delivers 4.3 times the AI performance of the M3 Ultra, and 9.8 times that of the M1 Ultra from three generations back.

Mac mini M6 and M5 Pro: LM Studio benchmarks released

The changes on the Mac mini side come with concrete numbers, according to Apple's announcement as cited by Testing Catalog. The M6 is Apple's first chip built on a 2-nanometer process, featuring a 12-core CPU, a 12-core GPU, and a dual 16-core Neural Engine. In LLM prompt processing tests run on LM Studio, it reportedly reaches 13.5 times the speed of the M1-equipped model and 4.8 times that of the previous-generation M4.

The higher-tier M5 Pro offers up to an 18-core CPU, up to a 20-core GPU, and support for up to 64GB of unified memory. On the same LM Studio benchmark, it delivers prompt processing speeds 8.5 times faster than the M2 Pro and 4 times faster than the M4 Pro.

Category	Mac Studio M5 Max	Mac Studio M5 Ultra	Mac mini M6	Mac mini M5 Pro
CPU	18-core	up to 36-core	12-core	up to 18-core
GPU	up to 40-core	up to 80-core	12-core	up to 20-core
Unified Memory	up to 128GB	up to 512GB	16GB (up to 32GB)	up to 64GB
Memory Bandwidth	614GB/s	1.2TB/s	up to 170GB/s	307GB/s

Clustering four units reaches data-center-class bandwidth

Both products support Thunderbolt 5 and RDMA (Remote Direct Memory Access), letting multiple units be linked into what functions as a single memory pool. Apple says clustering four Mac Studios together yields AI inference up to three times faster than a single system.

EXO Labs is the company that worked with Apple to refine this clustering technology. According to a post EXO Labs shared on X, the two have spent the past year building low-latency RDMA networking on top of Thunderbolt 5, which now lets multiple Macs be linked to run large models like Kimi K3 or GLM-5.3 at API-level speed. Kimi K3 is an open model built by Moonshot AI, while the GLM-5 series comes from Zhipu AI.

Memory bandwidth across RDMA-linked Macs scales nearly linearly, and EXO Labs says clustering four M5 Ultra-equipped Mac Studios together produces combined bandwidth of roughly 4.8TB/s. The company notes that speeds like this have, until now, only been achievable with data-center GPUs.

Software ecosystem and macOS 27

Apple also introduced a new framework called Core AI alongside these announcements. It's a tool that optimizes unified memory, CPU, GPU, and Neural Engine performance for Apple silicon, letting developers deploy full large language models locally or embed their own custom models directly into apps. Combine that with MLX, Apple's open-source machine learning framework, and developers can handle training, fine-tuning, and deployment entirely on a single Mac, according to Apple.

Images released by Apple show examples of running local LLMs with LM Studio Bionic, generating images with Draw Things, and linking multiple workstations together using exo. Both products will ship alongside the upcoming macOS 27, codenamed "Golden Gate," which introduces a new version of Siri capable of understanding on-screen content and carrying out tasks across apps on a user's behalf.

Liquid AI's lightweight vision-language model LFM2.5-VL-3B, released August 12, reportedly achieved a decode speed of 228 tokens per second on the Apple M5 Max. This new Mac Studio and Mac mini announcement is focused on scaling that kind of on-device model performance up to larger models and across more machines.

Editor's view

What Apple sold here wasn't really a chip — it was a promise that you can run big models without the cloud. OpenAI and Anthropic sell their models through APIs, but Apple sells you the machine to run those models on. The emphasis on skipping per-token costs and keeping data from ever leaving the building fits squarely into that same pitch.

The leap from the previous generation is easy to feel. Back in the M3 Ultra era, just loading a large open model onto a single Mac was already a challenge. Now, with 512GB of unified memory per unit and four units clustered together, you're looking at 4.8TB/s of bandwidth. Anyone who's tried running a model above the 100B-parameter range locally at working speed will immediately understand what that number means — until now, reaching that kind of bandwidth required a rack full of server-grade GPUs.

For teams in Korea, the group worth watching here is anyone trying to run open models in-house. Deploying a large open model like Kimi K3 or GLM-5 internally, without relying on a cloud API, has typically meant building a server stacked with multiple NVIDIA GPUs. Apple's answer instead is to cable together a few Macs. That said, this is strictly about inference — running an already-trained model. It doesn't mean a Mac cluster can replace the training process itself, where a model is built from scratch.

In the coming weeks, expect local AI tools like exo and LM Studio to roll out versions optimized for the M5 Ultra and M6 chips. Real-world benchmark posts from users testing out cluster configurations are also likely to start showing up soon.