Using voltage glitching to unleash the A100 in my mining card
335 points
• 1 day ago
• Article
Link
一名研究人员发现了 NVIDIA GA100 die 上的一个电压故障注入漏洞，该漏洞可实现对 eFuses 的未授权修改。利用这一缺陷，能够把受限硬件（例如 CMP 170HX 挖矿卡）变成完全可用的 A100-80G 数据中心级 GPU 。这些解锁后的显卡甚至可能超过零售版 A100，因为该漏洞可让此前因良率管理而被禁用的所有 streaming multiprocessors 重新启用，使性能大约达到标准量产机型的 1.2 倍。
该漏洞的关键在于 NVIDIA 如何进行产品分层。 GA100 die 主要使用两类保险丝：标准物理 eFuses 以及所谓的 fuseless fuses 。物理保险丝是永久性的，而 fuseless fuses 更像一种脚本式机制，用于告知硬件哪些功能需要启用或关闭。研究人员通过搭建自定义的电压故障注入方案，绕过安全校验，并把新的配置写入 fuseless fuse 区域。这样就相当于推翻了把高性能 die 限制成挖矿卡的约束，从而使 PCIe Gen4 、 NVLink 以及完整 framebuffer 容量等能力得以恢复。
要实现这种绕过，必须对 GPU 的电压调节器进行精确控制。研究人员设计了一种故障注入机制，用于接入 PWMVID 信号，从而在 GPU 的启动过程中操纵供电电压。通过找到"Goldilock's glitch"这一状态，安全敏感的 option words 就能在不触发防御性锁定的情况下被清除。障碍解除后，用户即可对硬件进行编程，使其忽略人为设定的限制。整个过程高度专业，通常需要进行硬件改装，包括拆除特定电阻并进行人工焊接连接，才能成功与显卡的供电系统建立接口。
这项研究带来的结果是显著的硬件能力提升。除了承载众多此类挖矿卡上本就具备的完整 80GB 内存外，用户还能够启用 Multi-Instance GPUs 与 NVDEC 等更高级功能，但前提是他们完成所需的硬件与固件调整。该研究在发表前已向 NVIDIA 的安全团队披露，然而相关技术已公开发布，目的是加深对 GA100 架构的理解，并展示那些因市场分层而本应长期处于沉睡状态的强大功能，仍可能被解锁。
A researcher has identified a voltage fault injection vulnerability on the NVIDIA GA100 die that allows for the unauthorized modification of eFuses. By exploiting this flaw, it is possible to convert restricted hardware, such as the CMP 170HX mining card, into a fully functional A100-80G datacenter-grade GPU. These unlocked cards can even outperform retail A100 units because the exploit enables all available streaming multiprocessors that were previously disabled for yield management, resulting in roughly 1.2 times the performance of standard production models.
The core of the vulnerability lies in how NVIDIA manages product segmentation. The GA100 die uses two main types of fuses: standard physical eFuses and so-called fuseless fuses. While physical fuses are permanent, fuseless fuses act as a script-like system that tells the hardware which features to enable or disable. By using a custom voltage glitching setup, the researcher was able to bypass security checks and write new configurations to the fuseless fuse region. This effectively overrides the restrictions that turn a high-performance die into a limited mining card, allowing for the restoration of features like PCIe Gen4, NVLink, and full framebuffer capacity.
Implementing this bypass requires precise control over the GPU's voltage regulator. The researcher developed a glitching mechanism that taps into the PWMVID signal to manipulate the supply voltage during the GPU's boot sequence. By finding the "Goldilock's glitch" state, the controller's security-sensitive option words can be cleared without triggering defensive locks. Once these barriers are removed, the user can program the hardware to ignore its artificial limitations. This process is highly technical and requires hardware modifications, including the removal of specific resistors and manual soldering of connections, to successfully interface with the card's power delivery system.
The end result of this research is a significant transformation in hardware capability. Beyond simply accessing the full 80GB of memory found on many of these mining cards, users can now enable advanced functionalities such as Multi-Instance GPUs and NVDEC, provided they perform the necessary hardware and firmware adjustments. While the research was disclosed to NVIDIA's security team prior to publication, the techniques have been released publicly to foster deeper understanding of the GA100 architecture and to demonstrate the potential for unlocking powerful features that would otherwise remain dormant due to market segmentation.
130 comments • Comments Link
• 向销售端禁用特定功能硬件的做法，是一种成熟的经济策略。通过制造单一的高规格设计并加以细分，公司最大限度地降低了与多条生产线、掩模以及库存周转相关的物流复杂度与成本。
• 硬件分档（binning）有双重目的：一方面让制造商能够回收存在部分缺陷的硅片与内存；另一方面实现更精确的市场细分。这样，本来可能会被丢弃、但仍带有部分可用功能的组件仍能产生收益，同时让高端 SKU 保持更高的利润率。
• 认为这些显卡只是为了市场细分而被"削弱"的普遍说法，忽视了良率（制造良好率）的现实。此类产品往往需要针对真实的内存或逻辑缺陷做出相应限制。因此，"解锁"过程天然存在不稳定风险，因为用户可能绕过了旨在隔离故障硬件区域的安全机制。
• "熔断"保险丝（fuse）的决策体现了一种更复杂、也更鲁莽的硬件破解方式。与可撤销或可修补的基于软件的解锁不同，这种手段会通过永久性的物理状态改变，以设计上不可逆转的方式改造设备。
• 对这类系统进行逆向工程极具吸引力。当小众硬件（例如带有过剩 HBM 的矿卡）被认为可能存在"受限"情况时，巨大的利润空间会促使社区投入长期、集中的攻关，以绕开制造商设置的限制。
• 关于产品细分的道德问题，存在明显的意识形态分歧。有人认为这本质上是剥削，用户应该拥有所购买硬件的全部功能；也有人认为这是降低门槛的必要机制，能压低预算有限的消费者的入门价格，否则他们将被排除在市场之外。
• 像 CMP 170HX 这类企业级硬件对散热要求极高。许多用户反馈，这些原本为高风量服务器机箱设计的显卡在家庭环境里难以妥善管理：一旦超出其额定限制运行，就需要定制散热方案或高噪音配置来避免过热。
• "解锁"社区的经验表明，对于某些工作负载（如 AI 推理或大语言模型运行），一定程度的不稳定性或少量位错误是可以容忍的。通过在软件层面重新初始化或管理内存故障，发烧友能够从本应被视为过时、等待被丢弃的硬件中挖掘出可用价值。
• 硅片生产的经济学决定了：一旦芯片制造完成，完全利用内存分配的边际成本通常可以忽略不计。相较之下，创建并维护多个性能较低的产品变体所带来的开销更高。因此，公司更倾向于先生产一个"最大化"版本，再按需对其进行限制。
• 这场讨论凸显了硬件发烧友为最大化利用二手市场组件所采取的极端手段。矿机硬件向 AI 爱好者领域的转移，反映出更普遍的趋势：人们正在努力规避上一代硬件的高昂成本，从而重新利用大量过剩的企业级库存。
本次讨论聚焦于企业产品细分策略与社区主导的回收被抑制硬件能力之间的紧张关系。尽管不少参与者认为硬件的"限制"是不道德的做法，但从技术与经济角度来看，主流共识是：这种模式是对良率与规模经济挑战的一种务实、节约成本的回应。在矿卡上解锁高带宽内存的现象，既体现了发烧友社区的创造力，也揭示了将回收或分级后的组件用于高强度工作负载所固有的风险。归根结底，这段对话强调了企业硬件的二级市场如何为高性能计算提供可及的切入口——即便这个过程需要显著的技术专长，并愿意接受潜在的不稳定风险。 • The practice of selling hardware with artificially disabled capabilities is a well-established economic strategy. By manufacturing a single, high-spec design and then segmenting it, companies minimize the logistical complexity and costs associated with managing multiple production lines, masks, and inventory streams.
• Hardware binning serves a dual purpose: it allows manufacturers to salvage silicon and memory with partial defects while simultaneously enabling precise market segmentation. This ensures that otherwise discarded, partially functional components remain profitable while premium SKUs maintain higher margins.
• The widespread assumption that these cards are purely "nerfed" for market segmentation overlooks the reality of manufacturing yields. Many of these units likely contain genuine memory or logic defects that necessitated the restriction; consequently, the "unlocking" process carries inherent risks of instability, as users may be bypassing safety gates intended to isolate faulty hardware regions.
• The decision to "blow fuses" represents a sophisticated, albeit reckless, form of hardware hacking. Unlike software-based unlocking, which can be undone or patched, this approach involves permanent physical state changes, transforming the device in ways that are intended to be irreversible by design.
• Reverse-engineering these systems is a highly incentivized endeavor. When a niche piece of hardware—such as mining cards with surplus HBM—is identified as potentially "hamstrung," the potential for significant profit drives a concentrated, long-term community effort to bypass manufacturer restrictions.
• There is a vocal ideological divide regarding the ethics of product segmentation. Some view it as inherently exploitative, arguing that users should have access to the full functionality of the hardware they purchase, while others contend it is an essential mechanism that lowers the entry price for budget-conscious consumers who would otherwise be priced out of the market entirely.
• The cooling requirements for enterprise-grade hardware like the CMP 170HX are substantial. Many users report that these cards, designed for high-airflow server chassis, are difficult to manage in home environments, necessitating custom cooling solutions or high-noise setups to prevent overheating when pushed beyond their rated, restricted specifications.
• The "unlocking" community demonstrates that for certain workloads, such as AI inference or large language model execution, a minor degree of hardware instability or a few bit errors may be tolerable. The ability to reinitialize or manage memory faults in software allows enthusiasts to extract utility from discarded hardware that would otherwise be considered obsolete.
• The economics of silicon production dictate that once a chip is fabricated, the marginal cost of the full memory allotment is often negligible compared to the overhead of creating and maintaining multiple, less-capable product variants. As a result, companies prefer to produce one "maximum" version and throttle it as needed.
• This discussion highlights the extreme lengths to which hardware enthusiasts will go to maximize value from secondary market components. The transition of mining hardware into the AI hobbyist space reflects a broader trend where massive, surplus enterprise inventory is repurposed by individuals seeking to circumvent the high costs of current-generation hardware.
The discussion centers on the tension between corporate product segmentation strategies and the community-led efforts to reclaim suppressed hardware capabilities. While many participants view the "hamstringing" of hardware as an unethical practice, the consensus from a technical and economic perspective is that this model is a pragmatic, cost-saving response to the challenges of manufacturing yields and economies of scale. The phenomenon of unlocking high-bandwidth memory on mining cards illustrates both the ingenuity of the enthusiast community and the risks inherent in using salvaged or binned components for high-intensity workloads. Ultimately, the conversation underscores how secondary markets for enterprise hardware create accessible entry points for advanced computing, even when that process requires significant technical expertise and the acceptance of potential hardware instability.