- DeepSeek R1’s 671 billion parameters run smoothly on the M3 Ultra’s unified memory
- Apple’s Mac Studio proves AI workloads don’t require expensive, power-hungry GPU clusters
- M3 Ultra consumes under 200W, far less than traditional multi-GPU AI setups
Apple’s Mac Studio with the M3 Ultra chip has demonstrated a capability that no other personal computer can match, running the DeepSeek R1 AI tool with 671 billion parameters entirely in memory.
A test by YouTube reviewer Dave2D showed despite using a 4-bit quantized version of the model, it retained its full parameter count and performed smoothly.
The DeepSeek R1 model, a hefty 404GB of storage and high-bandwidth memory typically found in GPU VRAM, is usually run on multi-GPU setups that distribute processing across several high-end graphics cards.
A unique feat: running DeepSeek R1 in memory
However, the M3 Ultra’s unified memory system, instead of relying on external GPUs, uses its 512GB of unified memory to store and process the AI model in a way that no other personal computer can.
Although MacOS imposes a default VRAM limit, Dave Lee manually increased it through the Terminal to allocate up to 448GB for AI processing, eliminating memory bottlenecks and reducing the need for multiple components to streamline AI performance on a single system.
One of the most striking aspects of this test was the M3 Ultra’s power efficiency, as it consumed less than 200W while running DeepSeek R1.
The ability to run such a demanding AI model without a multi-GPU setup challenges the industry standard, which relies on high-end Nvidia and AMD graphics cards, as the best workstations and server farms typically use GPU clusters that consume vast amounts of electricity.
Apple’s unified memory architecture enables significant power savings by sharing the M3 Ultra’s memory pool across CPU and GPU workloads, unlike conventional PC setups where VRAM is separate from system memory, maximizing bandwidth while minimizing energy use.
Apple’s Mac Studio, launched with the M3 Ultra chip, features up to a 32-core CPU and an 80-core GPU, making it one of the best LLM workstations and one of the best video editing computers.
Via Wccftech
You may also like
https://cdn.mos.cms.futurecdn.net/Rw3hh4WnYgWwq3Vwt4jScC-1200-80.jpg
Source link
