Apple's M5 Chip Rumored to Be 4x Faster on LLMs
Speculation suggests Apple's upcoming M5 chip will be four times faster than the M4 on large language model tasks. The M5 Max is rumored to feature over 500 GB/s of memory bandwidth for local inference, potentially making cloud-based AI processing optional for many workloads. This follows reports of a 3.3x speedup for distributed inference on current Apple Silicon, reducing time-to-first-token.
- The M4 chip's Neural Engine, which the M5 will succeed, is capable of 38 trillion operations per second (TOPS), a performance metric that has become a key battleground for on-device AI. Competitors like Intel with its Lunar Lake architecture and Qualcomm's Snapdragon X Elite are also heavily focused on NPU performance for their Windows-based AI PCs. - Apple's primary chip manufacturer, TSMC, has begun mass production of its 2nm process, which is the anticipated node for the M5 chip. This transition from the M4's 3nm process is expected to increase transistor density and improve power efficiency, which are critical for more complex AI models. - The significant jump in rumored memory bandwidth to over 500 GB/s for the M5 Max is a direct enabler for running larger language models locally. For comparison, the M4 Max chip tops out at 546 GB/s of memory bandwidth, while the base M4 is limited to 120 GB/s. - This focus on on-device processing aligns with Apple's broader AI strategy, which emphasizes user privacy by minimizing reliance on cloud servers. This approach contrasts with some competitors and is a key part of Apple's talent acquisition and retention narrative in the competitive Silicon Valley market. - The push for more powerful domestic AI hardware is set against a backdrop of evolving U.S. export controls on advanced semiconductors and manufacturing equipment to China. These regulations are designed to limit China's AI capabilities and influence global supply chains, affecting long-term manufacturing and resource planning. - Following the 2nm node, TSMC is accelerating its timeline for a 1.4nm process, aiming for mass production around 2028, which will be crucial for future Apple silicon generations beyond the M5. This roadmap is vital for Apple's long-term product engineering and maintaining a competitive edge.
