一项由社区开发者推动的新研究表明，NVIDIA GeForce RTX 30系列和RTX 40系列显卡隐藏着远超官方规格的计算潜力。通过对底层固件与硬件配置的解锁，开发者成功释放了原本被限制的部分计算能力，使部分型号在人工智能和矩阵运算场景下获得最高接近两倍的性能提升。
报道称，RTX 30系列“Ampere”架构显卡所使用的核心与NVIDIA专业级GPU本质上采用相同硅片，但消费级产品的部分计算能力被软件和固件策略限制。一位研究人员开发的解锁方案能够提升FP/BF16 FMA到FP32运算的指令发射速率，适用于GA102、GA103、GA104和GA106核心。
在实际测试中，RTX 3090 Ti的BF16矩阵乘法性能最高提升约80%。由于矩阵乘法是运行绝大多数神经网络模型的重要基础运算，这一提升对于本地AI推理和机器学习任务具有明显意义。
与此同时，另一项针对RTX 40系列“Ada Lovelace”架构的解锁项目也取得突破。该方案实现了FP16和BF16 Tensor计算吞吐量提升约两倍，同时保留FP32累加精度。
以RTX 4090为例，在FP16 Tensor吞吐量测试中，显卡原本可达到165.6 TFLOPS，而在启用解锁后，性能提升至331.6 TFLOPS，几乎翻倍。这一水平已经接近甚至达到部分专业级产品，例如RTX 6000 Ada所具备的计算能力。
研究人员认为，这表明AD102核心在消费级产品中存在人为设置的计算限制，尤其是在矩阵乘法和累加等关键AI运算领域。这些限制会直接影响大型语言模型等AI应用的运行效率。目前，该解锁项目已确认支持AD102、AD103及AD106核心型号。
值得关注的是，针对RTX 30系列和RTX 40系列的两项解锁方案几乎同时出现，分别覆盖Ampere和Ada Lovelace两代GPU架构。开发者表示，这些修改能够通过NVIDIA驱动程序的完整验证机制，同时成功绕过PCIe功能级重置（Function Level Reset，FLR）限制，因此仍可使用官方原版驱动程序，并正常运行CUDA相关测试软件。
测试数据显示，无论是GA102还是AD102显卡，解锁前后均保持相同的功耗限制。例如在450W总板卡功耗设定下，性能提升完全来自被释放的计算能力，而非额外提高功耗或频率。
对于使用本地大语言模型、生成式AI以及其他高强度计算任务的用户而言，额外释放的Tensor计算能力能够提升提示词预处理和推理阶段的效率，从而获得更高的Token吞吐量和更快的响应速度。
不过开发者也提醒，这类修改属于非官方行为，可能导致显卡保修失效。此外，相关模块需要在系统启动阶段率先加载，以便在NVIDIA驱动接管GPU之前完成解锁操作。因此，这类方案更适合具备一定技术能力并愿意承担风险的高级用户。
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。
A new study driven by community developers shows that the NVIDIA GeForce RTX 30 series and RTX 40 series GPU hide the computing potential of far beyond official specifications. By unlocking the bottom solids and hardware configuration, the developer successfully released the originally restricted partial computing capability, resulting in up to twice the performance of some models under artificial intelligence and matrix computing scenarios.
It was reported that the core used in the RTX 30 series of “Ampere” architecture GPU was essentially the same as the NVIDIA professional GPU, but that part of the computing capacity of consumer products was limited by software and solidware strategies. A locking programme developed by a researcher could enhance the command launch rate of the FP/BF16 FMA to FP32, applicable to GA102, GA103, GA104 and GA106 cores.
In the actual test, RTX 3090 Ti ' s BF16 matrix multiplication performance increased by a maximum of about 80%. As matrix multiplication is an important base calculation for operating most of the neuronetwork models, this upgrade is clearly relevant for local AI reasoning and machine learning missions.
At the same time, there was a breakthrough in another lock-out project targeting the RTX 40 series “Ada Lovelace”. The programme achieved an increase of about twice in the volume calculated by FP16 and BF16 Tensor, while retaining FP32 cumulative accuracy.
In the RTX 4090 case, GPU could have reached 165.6 TFLOPS in the FP16 Tensor Bumping Test, while the performance increased to 331.6 TFLOPS, almost double, after the unlocking was enabled. This level is close to, and even reaches, some professional-level products such as RTX 60000 Ada ' s computing capability.
According to researchers, this suggests that there are artificial computational limits for AD102 cores in consumer-grade products, especially in key AI fields such as matrix multiplication and accumulation. These limitations have a direct impact on the operational efficiency of AI applications such as large language models. Currently, the lock-out project has identified support for core models AD102, AD103 and AD106.
Of concern are the almost simultaneous emergence of two lock-outs for the RTX 30 series and the RTX 40 series, covering the two-generation GPU structures of Ampere and Ada Lovelace, respectively. The developers indicated that these changes could be verified through the complete verification mechanism of the NVIDIA driver, while successfully bypassing the PCIE functional level replacement (Function Level Reset, FLR) restrictions, thus still using the official original driver and running the CUDA-related testing software.
Test data show that both GA102 and AD102GPU maintain the same working limits before and after the unlocking. For example, under the 450 W master card power set, performance enhancement is entirely derived from the ability to calculate release, rather than from additional power consumption or frequency.
For users who use local large-language models, generating AI and other high-intensity computing tasks, additional released Tensor computing capabilities can increase the efficiency of the prompt pre-processing and reasoning phase, thereby obtaining higher Token throughput and faster response speed.
However, developers also cautioned that such modifications were unofficial and could result in the lapse of GPU warranties. In addition, the modules need to be first loaded during the system start-up phase in order to complete the unlocking operation before the NVIDIA driver takes over GPU. As a result, such programmes are better suited to senior users with a certain technical capacity and a willingness to take risks.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.