On August 25, details from Apple's (AAPL.O) official website revealed that its newly introduced M5 Ultra chip boasts a high-capacity, high-bandwidth unified memory configuration of up to 512GB. The unified memory bandwidth on this chip soars to an impressive 1.2TB/s, marking a 50% enhancement over its predecessor, the M3 Ultra. This advancement allows users to effortlessly store vast datasets directly in local memory, boost the number of Tokens generated per second, and facilitate the on-device operation of ultra-large large language models (LLMs) with hundreds of billions of parameters.
Meanwhile, the M6 chip is designed to support up to 32GB of unified memory, empowering it to manage multiple high-demand applications concurrently and execute large language models on the device itself. This capability enables it to perform AI agent tasks with a strong emphasis on security and privacy. Furthermore, the unified memory bandwidth of the M6 chip can reach up to 170GB/s, signifying a 10% improvement over the M5 and a remarkable 2.5 times increase compared to the M1.
