高通近期面向搭载骁龙X与骁龙X2系列处理器的Windows on Arm设备发布了Adreno 32.0.181.0图形驱动，原本旨在提升游戏稳定性和应用兼容性。然而，多位用户在更新后发现，该驱动可能导致骁龙X2 Elite平台在部分专业应用和本地AI工作负载中出现显著性能下降，引发社区广泛关注。
根据用户在Reddit论坛发布的测试结果，问题主要集中在Adobe Premiere Pro视频渲染以及本地大语言模型（LLM）推理两大场景。其中，一位使用华硕Zenbook A16、搭载骁龙X2 Elite处理器的用户表示，在升级至32.0.181.0驱动后，一段时长约7分钟、分辨率为1080p的视频，导出时间从原先的1分40秒大幅增加至约10分钟，渲染速度下降接近6倍。该用户认为，新驱动安装后，内容创作软件的运行效率明显受到影响，并建议暂时避免升级至这一版本。
另一位用户则对微软Surface Laptop第八代设备上的骁龙X2 Elite平台进行了更为系统的测试。测试设备搭载Adreno X2-85 GPU，并使用llama.cpp结合OpenCL后端执行多个Qwen3系列模型。测试结果显示，与此前的32.0.172.2驱动相比，升级到32.0.181.0之后，多项模型的文本生成速度出现明显下滑。
具体来看，Qwen3-0.6B模型的生成速度从每秒98.1个Token下降至12.8个Token；Qwen3-4B模型从每秒29.5个Token下降至9.0个Token；Qwen3-Coder-30B-A3B-Instruct模型则从每秒26.8个Token下降至5.9个Token。根据不同模型和工作负载的差异，整体性能跌幅最高达到87%。
为了确认问题来源，该用户进行了反复验证。在卸载32.0.181.0驱动并恢复至旧版后，性能立即恢复正常；重新安装新驱动后，性能下降现象再次出现。测试过程中，硬件配置、操作系统版本、软件环境以及模型文件均保持一致，且设备始终连接电源运行，从而最大程度排除了其他变量的影响。
值得注意的是，问题似乎主要集中在GPU相关计算任务。用户反馈显示，骁龙X2 Elite的CPU与NPU性能并未受到明显影响，性能下滑主要出现在依赖Adreno GPU进行加速的AI推理和视频渲染工作负载中。这意味着问题更可能与新版Adreno驱动中的OpenCL或GPU计算相关优化有关，而非处理器本身的性能缺陷。
截至目前，高通尚未就相关用户反馈发布正式说明，也没有确认是否存在驱动缺陷。对于依赖本地AI模型运行、视频编辑或其他GPU计算任务的用户而言，在更多调查结果公布之前，继续使用32.0.172.2等较早版本驱动，或许是更为稳妥的选择。
此次事件也再次凸显了Windows on Arm生态在快速发展过程中面临的软件兼容性挑战。随着骁龙X2 Elite不断向创作者和AI开发者市场渗透，驱动层面的稳定性和持续优化能力，将成为平台整体体验的重要影响因素。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The recent publish-based Windows on Arm for the X and X2 series processors of the Quilloon X was designed to enhance game stability and compatibility with the Adreno 32.181.0 graphically driven. However, after updating, several users found that the driver could lead to thorium The DragonX2 Elite platform has experienced significant declines in performance in some professional applications and in local AI workloads, generating widespread community concern.
Based on the results of the user ' s tests at the Reddit Forum publish, the problems are mainly concentrated in Adobe Premiere Pro video rendering and in the local LLM model ' s reasoning. Among them, one user, using a Zenbook A16 with a Zenbook X2 Elite processor, indicated that, having been upgraded to 32.0.181.0, a video with a length of about 7 minutes and resolution with 1080p, the output time had increased significantly from 1 minute 40 seconds to about 10 minutes, and the rendering rate had dropped nearly six times. The user was of the view that the efficiency of the operation of the content creation software had been significantly affected by the installation of the new driver and suggested that it should be avoided for the time being.
Another user performed more systematic testing of the X2 Elite platform on Microsoft Surface Laptop's eighth generation equipment. The test equipment carries the Adreno X2-85 GPU and executes multiple Qwen3 series models using llama.cpp in conjunction with OpenCL backend. The test results showed a marked decline in the text generation rate of multiple models after upgrading to 32.0.181.0 compared with the previous 32.0.172.2 driver.
Specifically, the Qwen3-0.6B model decreased from 98.1 Tokens per second to 12.8 Tokens; the Qwen3-4B model from 29.5 Tokens per second to 9.0 Tokens; and the Qwen3-Coder-30B-A3B-Instract model from 26.8 Tokens per second to 5.9 Tokens. Based on differences between models and work loads, the overall performance dropped by up to 87 per cent.
In order to identify the source of the problem, the user was repeatedly validated. The performance returned to normal as soon as the 32.0.181.0 driver was unloaded and returned to the old version; the re-installation of the new driver led to a recurrence of the downswing. The hardware configuration, the operating system version, the software environment and the model file were consistent during the testing and the equipment was always connected to the power source, thus minimizing the impact of other variables.
It is noteworthy that the problem appears to be concentrated mainly on GPU-related computing tasks. User feedback indicates that the CPU and NPU performances of Tyrone X2 Elite have not been significantly affected, and that the downslide is mainly in the context of the accelerated AI reasoning and video rendering loads that rely on Adreno GPU. This means that the problem is more likely to be related to the optimization associated with the OpenCL or GPU calculations in the new version of the Adreno drive than to the performance defects of the processor itself.
To date, Chase has not provided feedback on publish for relevant users, nor has it confirmed the existence of driver deficiencies. For users who rely on local AI models, video editors or other GPU computing tasks, it may be a more prudent option to continue using earlier versions such as 32.0.172.2 until more results are available.
The event also highlighted the software compatibility challenge of the Windows on Arm ecology in a fast-growing process. As Tyrone X2 Elite continues to penetrate into the market for creators and AI developers, the stability and continuous optimization at the driving level will be an important factor in the overall experience of the platform.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.