A faulty Snapdragon X2 Elite driver update has triggered severe performance drops of up to 87 percent in local artificial intelligence workloads, according to tests on Qualcomm laptops this week. The Adreno GPU software update slows OpenCL tasks and significantly increases video export times in Adobe Premiere Pro.
Key points
- Qualcomm’s Adreno GPU driver version 32.0.181.0 reduces local large language model generation speeds by up to 87 percent on Snapdragon X2 Elite hardware.
- Testing across open-source models shows the issue is isolated to OpenCL pipelines, leaving CPU, NPU, and WebGPU performance largely untouched.
- Adobe Premiere Pro export times have increased up to sixfold on affected machines, prompting users to roll back to driver version 32.0.172.2.
The benchmark findings
The extent of the drop was documented by a user testing a Microsoft Surface Laptop 8th Edition, which pairs the Snapdragon X2 Elite system-on-chip with Qualcomm’s Adreno X2-85 integrated graphics processor. According to detailed performance testing posted on Reddit, moving from driver version 32.0.172.2 to version 32.0.181.0 caused text generation speeds in local large language models to plunge by 69 to 87 percent.
The benchmarks evaluated text generation using the popular llama.cpp framework running over an OpenCL backend under Windows. Across multiple model sizes, the newer driver introduced immediate degradation. On the compact Qwen3-0.6B model, output fell from 98.1 tokens per second on the older software to just 12.8 tokens per second on version 32.0.181.0, representing an 87 percent loss in throughput.
Intermediate and larger models suffered comparable declines. Generation on Qwen3-4B dropped from 29.5 tokens per second to 9.0 tokens per second. Meanwhile, the coding model Qwen3-Coder-30B-A3B-Instruct collapsed from 26.8 tokens per second down to 5.9 tokens per second. The tester verified that the operating system, host machine, software versions, and model weights remained identical throughout testing, with the laptop connected continuously to mains power.
To confirm that the driver was the sole variable, the user rolled the software back to version 32.0.172.2, which immediately restored full inference speeds. Upgrading a second time to version 32.0.181.0 reproduced the performance collapse, establishing a direct link between the software package and the drop in computing speed.
Video editing and render delays
The issue is not confined to machine learning inference. A separate report highlighted severe bottlenecks in creative software running on Snapdragon hardware. According to an initial user report on Reddit, an Asus Zenbook A16 running the same 32.0.181.0 driver suffered extreme slowdowns during video export in Adobe Premiere Pro.
The user reported that a standard seven-minute video file at 1080p resolution previously completed export in one minute and 40 seconds. Following the installation of driver 32.0.181.0, the identical export job required roughly ten minutes to finish, representing a sixfold increase in render duration. The poster cautioned other video editors against applying the update, pointing out that editing suites on the platform frequently encounter instabilities following graphics driver updates.
These experiences illustrate the continued fragility of the graphics software stack on Windows on Arm devices. While hardware specifications on paper match demanding creative tasks, driver changes can introduce regressions that disrupt time-sensitive production workflows.
Why the OpenCL pipeline broke
Detailed telemetry collected during the language model tests indicates that the bug is confined to specific execution pathways rather than the entire silicon package. The Snapdragon X2 Elite incorporates custom Oryon CPU cores, an upgraded Hexagon neural processing unit, and the Adreno X2-85 GPU. The benchmark data shows that the CPU and NPU escaped the issue entirely.
When running inference on the Oryon CPU cores, token generation rates shifted by only 1 to 4 percent between the two driver packages, well within normal run-to-run variance. Similarly, workloads directed to the onboard NPU remained stable, moving by only a fraction of a percent either way. This isolation confirms that memory controllers, power management profiles, and general operating system scheduling were operating normally.
Even within the Adreno GPU itself, the breakdown appears selective. When the tester ran the Qwen3-0.6B model through a WebGPU runtime rather than OpenCL, performance held steady at approximately 90 tokens per second on both driver revisions. GPU prompt processing times also stayed unchanged on the smaller models. However, prompt ingestion on Qwen3-Coder-30B-A3B-Instruct deteriorated by 73 percent.
Because llama.cpp relies on OpenCL to interface with Adreno silicon under the Windows environment, the tester concluded that driver version 32.0.181.0 introduced an instruction handling or memory dispatch bug specifically inside Qualcomm’s OpenCL runtime. The failure of prompt processing on larger models suggests that memory management under heavier memory pressure may be failing to allocate or transfer weights efficiently across the shared memory bus.
What users should do next
The performance defects have been escalated directly to Qualcomm engineers through the company’s official developer channels. A report detailing the reproducible drops in OpenCL execution and Premiere Pro rendering was submitted to the Qualcomm Developers Discord server, where technical staff monitor community bugs.
Until Qualcomm provides a hotfix or an updated driver release that repairs the OpenCL dispatch logic, users relying on Snapdragon X2 Elite laptops for local AI experimentation or video production should avoid version 32.0.181.0. Those who have already installed the update can recover their previous rendering speeds by manually rolling back to driver version 32.0.172.2 through the Windows Device Manager.
The regression arrives at a sensitive moment for Qualcomm as it attempts to solidify its presence in the PC workstation and AI laptop markets. While Oryon processor cores have proven competitive against established x86 platforms from Intel and AMD, sustained adoption among developers and power users depends on consistent, trouble-free graphics and compute drivers.
