YouTuber Dave Lee from the Dave2D channel showed how the new Mac Studio with chip M3 Ultra can effectively run the DeepSeek R1 AI model if the device is equipped with maximal capacity of 512 GB of memory. According to his tests, it is possible to run a version of the model with 671 billion parameters directly on this workstation, but it requires significant memory resources, including 404 GB of storage and manual allocation of 448 GB of video RAM using commands in the terminal.
You might be interested in
A key element of successful operation is the chip's unified memory architecture M3 Ultra, which allows for efficient processing of the quantized 4-bit version of the DeepSeek R1 model. This quantization slightly reduces accuracy, but at the same time preserves all parameters and provides a performance of approximately 17-18 tokens per second, which is sufficient for most practical applications. The energy efficiency is also impressive. Mac Studio manages to operate the model with less than 200 watts of power, which is a stark contrast to traditional PC systems, where similar performance would require multiple graphics cards with up to ten times the power consumption. The ability to run advanced AI models locally brings significant advantages, especially in areas where data security is important, such as healthcare or enterprise applications.
Gallery
However, such performance costs a lot of money. Mac Studio s M3 Ultra and 512GB of RAM starts at around $10. A fully-featured configuration with a 16TB SSD and a chip M3 Ultra with a 32-core CPU, 80-core GPU, and 32-core Neural Engine, it will cost $14.Apple says that M3 Ultra is the most powerful chip in Mac history to date, with its performance achieved by combining two M3 chips Max through technology UltraFusion, which doubles their specifications.
