- The 671 billion parameters of Deepseek R1 take place gently on the unified memory of the M3 Ultra
- Apple’s Mac studio proves that AI workloads do not require costly and powerful GPU clusters in power
- M3 Ultra consumes less than 200 W, much less than the traditional configurations of Multi-GPU AI
The Apple Mac Studio with the M3 Ultra chip has demonstrated a capacity that no other personal computer can correspond, running the R1 Deepseek R1 tool with 671 billion parameters entirely in memory.
A test of the Dave2D Youtube examiner has shown despite the use of a 4 -bit quantified version of the model, it has retained its full number of parameters and worked smoothly.
The R1 Deepseek model, a heavy 404 GB of storage and a large-band memory generally found in GPU VRAM, is generally executed on multi-GPU configurations that distribute treatment on several high-end graphics cards.
A unique feat: Running Deepseek R1 in memory
However, M3 Ultra’s unified memory memory system, instead of counting on external GPUs, uses its 512 GB of unified memory to store and treat the AI model in a way that no other personal computer can.
Although MacOS imposes a default VRAM limit, Dave Lee increased it manually via the terminal to allocate up to 448 GB for AI treatment, eliminating the bottlenecks of memory and reducing the need of several components to rationalize the performance of AI on a single system.
One of the most striking aspects of this test was the energy efficiency of the M3 Ultra, because it consumed less than 200 W while running Deepseek R1.
The possibility of executing an AI model also demanding without a multi-GPU configuration calls into question the standard of the industry, which relies on high-end Nvidia and AMD graphics cards, because the best workstations and servers farms generally use GPU clusters that consume large amounts of electricity.
Apple’s unified memory architecture allows significant food savings by sharing the memory pool of the M3 Ultra on CPU and GPU workloads, unlike conventional PC configurations where VRAM is separated from system memory, maximizing bandwidth while minimizing energy consumption.
Apple’s Mac studio, launched with the M3 Ultra chip, present up to a 32 core CPU and a GPU of 80 cores, making it one of the best LLM workstations and one of the best video editing computers.
Via WCCFTECH
