[Maui, United States = DigitalToday reporter Dae-geon Seok] Cost pressure is rising for wearables that use large amounts of DRAM as memory supply shortages persist. AI device makers are reducing DRAM per device through code optimisation. Qualcomm is also lowering the burden by supporting memory-sparing 1-bit models on the Snapdragon AR1 family.

Ziad Asghar (지아드 아스가르), executive vice president at Qualcomm and head of the XR, wearable and personal AI division, said at a Snapdragon Summit personal AI and sound media Q&A on Sept. 23 local time that the industry broadly has memory problems but small devices use small memory components. He said earbuds were barely affected because memory accounts for a small share of electronic component costs, while devices that use more DRAM were hit more. AI glasses fall in between.

AI glasses makers are responding by refining software to reduce memory requirements. Asghar said many glasses used 4 GB of DRAM in the past, but new models increasingly use 2 GB, and efforts are under way to cut it further. He said all companies are currently focusing on code optimisation.

Qualcomm's 1-bit model is applied after code optimisation. The approach reduces model size by lowering numerical precision used in AI model computation to 1 bit. Asghar said the 1-bit model is especially important for small devices with tight memory limits, adding that sharply lower memory requirements can cut component costs for such devices.

But lowering precision comes with the issue of reduced accuracy. Asghar said Qualcomm partners' technology keeps accuracy loss very limited even when converting higher-precision models to 1 bit.

The first device to apply the 1-bit model is AI glasses. Qualcomm demonstrated a 1-bit multimodal model in the summit demo room that recognises objects in front of the glasses and answers questions. Asghar said distinguishing the colour of an apple may be easy for some people but can be a major challenge for the visually impaired, adding that there is meaning in a small device handling such processing on its own. Translation models are also getting smaller. He added that some partners are developing ultra-small models that translate only a single language, and that these models along with speech recognition and speech synthesis models can run on a micro NPU.

Camera-equipped earbuds were designed from the start to use less memory. Cleer camera earbuds have no space to store photos. Asghar said the camera provides an agent with only visual information such as whether there are many people in a room, whether it is an office, or whether it is a clear day, adding that higher resolution or additional information is not needed.

The earbud industry is also seeing moves toward platforms that use less memory. Andrew Leister (앤드루 레이스터), senior product manager at Qualcomm, said memory shortages are instead becoming a positive catalyst for low-power audio devices. Some companies have used DRAM-based Linux and Android platforms because they are easy to develop on, but Qualcomm now provides a similar development environment on lightweight platforms, creating room to switch.

Leister said moving to a lightweight platform cuts memory costs and overall product costs and reduces power consumption. He added that on-chip AI processing can improve voice quality, eliminating the need to increase the number of microphones, and that microphone costs will not rise either.

Qualcomm is supporting the shift with Snapdragon Sound Apps and Agent. Snapdragon Sound Elite second generation adds external memory to shared memory within the platform, allowing large models that do not fit entirely on the chip to be loaded and run from external memory. Leister said some companies are considering whether now is the time to move to a lighter platform, adding that Qualcomm makes the transition easier.