Apple Unveils M5 Ultra and 2nm M6 Enabling 100B+ Parameter AI Models on Desktop.

Apple Unveils Next-Gen Silicon: The 2nm M6 Chip and Heavyweight M5 Ultra

Apple has officially expanded its custom silicon portfolio with the debut of two groundbreaking processors: the baseline M6 and the flagship M5 Ultra. Debuting alongside refreshed hardware configurations for the Mac mini and Mac Studio, these chips mark major technological milestones in semiconductor fabrication, power efficiency, and on-device artificial intelligence processing.

The M6 introduces Apple's first-ever 2-nanometer (2nm) manufacturing process node. It features an innovative core layout comprising a 12-core CPU divided into 2 Super Cores, 4 Performance Cores, and 6 Efficiency Cores. Paired with a 12-core GPU, it supports up to 32GB of unified memory with a peak memory bandwidth of 170GB/s. Crucially, the M6 incorporates a dual 16-core Neural Engine architecture designed to drastically accelerate local machine learning and AI execution.

Targeting enterprise workstations, the M5 Ultra skips the M3 Ultra generation entirely to fuse two M5 Max dies using Apple's proprietary UltraFusion interconnect technology. The resulting titan features up to a 36-core CPU, an 80-core GPU, and a 32-core Neural Engine. Supporting up to 512GB of unified memory at a staggering 1.2TB/s memory bandwidth, the M5 Ultra enables developers and researchers to run large language models (LLMs) with hundreds of billions of parameters entirely on-device without relying on external cloud infrastructure.

The shift to 2-nanometer semiconductor manufacturing and the reduction in gate transistor size allowed Apple to significantly pack more logic processing units into a smaller silicon area while simultaneously reducing power consumption and heat generation. The introduction of dedicated "supercores" within the 12-core CPU architecture ensures extremely high single-threaded performance for active front-end tasks, while performance cores smoothly handle background system operations.

Training and running large AI models requires rapid data transfer between system RAM and processing units. Delivering a combined memory bandwidth of 1.2TB/s coupled with 512GB of RAM, the M5 Ultra bypasses standard memory bottlenecks plaguing traditional PC architectures, enabling the internal hardware to process neural networks with over 100 billion parameters with near-zero latency.

Previously, running high-end open-weighted AI models required expensive server clusters or subscription-based cloud APIs. Running billions of parameter models locally on the M5 Ultra's Mac Studio ensures complete data privacy, reduces ongoing cloud hosting costs, and allows developers to fully fine-tune specialized models offline.

 

Source: Apple 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Apple Unveils Next-Gen Silicon: The 2nm M6 Chip and Heavyweight M5 Ultra

Apple has officially expanded its custom silicon portfolio with the debut of two groundbreaking processors: the baseline M6 and the flagship M5 Ultra. Debuting alongside refreshed hardware configurations for the Mac mini and Mac Studio, these chips mark major technological milestones in semiconductor fabrication, power efficiency, and on-device artificial intelligence processing.

The M6 introduces Apple's first-ever 2-nanometer (2nm) manufacturing process node. It features an innovative core layout comprising a 12-core CPU divided into 2 Super Cores, 4 Performance Cores, and 6 Efficiency Cores. Paired with a 12-core GPU, it supports up to 32GB of unified memory with a peak memory bandwidth of 170GB/s. Crucially, the M6 incorporates a dual 16-core Neural Engine architecture designed to drastically accelerate local machine learning and AI execution.

Targeting enterprise workstations, the M5 Ultra skips the M3 Ultra generation entirely to fuse two M5 Max dies using Apple's proprietary UltraFusion interconnect technology. The resulting titan features up to a 36-core CPU, an 80-core GPU, and a 32-core Neural Engine. Supporting up to 512GB of unified memory at a staggering 1.2TB/s memory bandwidth, the M5 Ultra enables developers and researchers to run large language models (LLMs) with hundreds of billions of parameters entirely on-device without relying on external cloud infrastructure.

The shift to 2-nanometer semiconductor manufacturing and the reduction in gate transistor size allowed Apple to significantly pack more logic processing units into a smaller silicon area while simultaneously reducing power consumption and heat generation. The introduction of dedicated "supercores" within the 12-core CPU architecture ensures extremely high single-threaded performance for active front-end tasks, while performance cores smoothly handle background system operations.

Training and running large AI models requires rapid data transfer between system RAM and processing units. Delivering a combined memory bandwidth of 1.2TB/s coupled with 512GB of RAM, the M5 Ultra bypasses standard memory bottlenecks plaguing traditional PC architectures, enabling the internal hardware to process neural networks with over 100 billion parameters with near-zero latency.

Previously, running high-end open-weighted AI models required expensive server clusters or subscription-based cloud APIs. Running billions of parameter models locally on the M5 Ultra's Mac Studio ensures complete data privacy, reduces ongoing cloud hosting costs, and allows developers to fully fine-tune specialized models offline.

 

Source: Apple