A Surface Laptop 8th Edition owner has reported a sharp slowdown in local language models after updating the Adreno graphics driver on a Snapdragon X2 Elite system. Moving from version 32.0.172.2 to 32.0.181.0 reduced GPU text-generation speed through llama.cpp’s OpenCL backend by 69–87% across three Qwen3 models in the user’s tests.
Qwen3-0.6B fell from 98.1 to 12.8 tokens per second, Qwen3-4B from 29.5 to 9, and Qwen3-Coder-30B from 26.8 to 5.9. The laptop had an Adreno X2-85 GPU and 64 GB of RAM and was connected to AC power. Rolling back restored performance, while installing the new driver again reproduced the slowdown.
CPU and NPU results remained broadly unchanged in the same checks. Qwen3-0.6B running through the alternative WebGPU graphics interface also held steady at around 90 tokens per second on both drivers.
