Qualcomm Adreno 32.0.181.0 Driver Causes Snapdragon X2 Elite Local AI and Premiere...
A Reddit user's testing found that after the Snapdragon X2 Elite was upgraded to Adreno driver 32.0.181.0, local LLM inference and Premiere Pro video rendering performance dropped sharply, with declines of up to 87%. When running Qwen3 series models with llama.cpp using the OpenCL backend, Qwen3-0.6B's generation speed fell from 98.1 tokens per second to 12.8 tokens per second, Qwen3-4B from 29.5 to 9.0, and Qwen3-Coder-30B-A3B-Instruct from 26.8 to 5.9. Another user reported that the export time for a 7-minute 1080p video increased from 1 minute 40 seconds to about 10 minutes. After rolling back to driver 32.0.172.2, performance returned to normal, with the CPU and NPU not noticeably affected. Qualcomm has not yet issued a statement on the matter.
Why read this
AI score 58
User testing shows that the new Adreno driver significantly slows down GPU-accelerated tasks on the Snapdragon X2 Elite, offering insight into the driver compatibility risks of Windows on Arm.
Topics
Body · AI translation
According to test results posted by users on the Reddit forum, the problems are mainly concentrated in two major scenarios: Adobe Premiere Pro video rendering and local large language model (LLM) inference. Among them, a user using an ASUS Zenbook A16 equipped with a Snapdragon X2 Elite processor stated that after upgrading to driver 32.0.181.0, the export time for a video about 7 minutes long at 1080p resolution increased significantly from the original 1 minute 40 seconds to about 10 minutes, with rendering speed dropping by nearly 6 times. The user believes that after the new driver was installed, the operating efficiency of content creation software was clearly affected, and suggested temporarily avoiding upgrading to this version.
Another user conducted a more systematic test of the Snapdragon X2 Elite platform on a Microsoft Surface Laptop 8th generation device. The test device was equipped with an Adreno X2-85 GPU and used llama.cpp combined with the OpenCL backend to run multiple Qwen3 series models. The test results showed that, compared with the previous 32.0.172.2 driver, after upgrading to 32.0.181.0, the text generation speed of multiple models declined significantly.
Specifically, the generation speed of the Qwen3-0.6B model dropped from 98.1 tokens per second to 12.8 tokens; the Qwen3-4B model dropped from 29.5 tokens per second to 9.0 tokens; and the Qwen3-Coder-30B-A3B-Instruct model dropped from 26.8 tokens per second to 5.9 tokens. Depending on differences in models and workloads, the overall performance decline reached as high as 87%.
To confirm the source of the problem, the user carried out repeated verification. After uninstalling driver 32.0.181.0 and restoring the older version, performance immediately returned to normal; after reinstalling the new driver, the performance decline appeared again. During testing, the hardware configuration, operating system version, software environment, and model files all remained consistent, and the device was always running connected to power, thereby ruling out the influence of other variables to the greatest extent possible.
It is worth noting that the problem appears to be mainly concentrated in GPU-related computing tasks. User feedback shows that the CPU and NPU performance of the Snapdragon X2 Elite were not significantly affected; the performance decline mainly appeared in AI inference and video rendering workloads that rely on Adreno GPU acceleration. This means the problem is more likely related to OpenCL or GPU computing-related optimizations in the new Adreno driver, rather than a performance defect in the processor itself.
As of now, Qualcomm has not issued an official statement regarding the relevant user feedback, nor has it confirmed whether a driver defect exists. For users who rely on running local AI models, video editing, or other GPU computing tasks, continuing to use earlier driver versions such as 32.0.172.2 may be a more prudent choice before more investigation results are published.
This incident also once again highlights the software compatibility challenges faced by the Windows on Arm ecosystem during its rapid development. As the Snapdragon X2 Elite continues to penetrate the creator and AI developer markets, driver-level stability and continuous optimization capabilities will become important factors affecting the overall platform experience.