Qualcomm CEO Says Smartphones Will Run 100B AI Models by 2028

Qualcomm CEO Cristiano Amon has expressed his belief that smartphones will be capable of continuously running AI models with 100 billion parameters by 2028. While acknowledging the challenges related to memory capacity, cost, and heat, Amon stated that some AI companies are already developing smartphones capable of handling these large models.

However, memory remains the most significant obstacle, with current DRAM supply situations suggesting smartphones will likely remain in the 16GB RAM range for the foreseeable future. Despite this, Amon expects future devices to feature much higher memory capacities, even with the high costs of DRAM. Running a 100B model locally would require substantial memory, with even reduced quantization levels remaining higher than most current smartphones' memory capacities.

Additionally, the continuous operation of such a model would generate significant heat, potentially leading to performance throttling. While some hardware modifications by Apple and Qualcomm may help alleviate these issues, more aggressive solutions are likely necessary. Approaches such as aggressive model quantization, sparse MoE models, and smaller distilled models could help overcome these limitations, though they may sacrifice some model quality or increase complexity.

Amon has not provided a detailed timeline for achieving this ambitious goal, but the potential for smartphones to run ever-more-capable AI models remains a topic of interest in the industry.