随着人工智能计算需求持续增长，中国出现了将英伟达GeForce RTX 5090游戏显卡大规模改造为AI加速卡的生产线。根据Wccftech及其他硬件媒体援引的最新视频资料，相关工厂正在拆解零售版RTX 5090显卡，将GPU核心和显存芯片转移到专门设计的服务器电路板上，再配备适合数据中心使用的散热系统，使原本面向游戏玩家的高端显卡能够用于AI计算。
最新流出的视频由香港硬件媒体HKEPC分享。画面显示，一处规模较大的改装车间内堆放着大量零售版RTX 5090显卡，工人按照流水线流程拆卸显卡、移除原有散热器，并将GPU和显存芯片转移至定制电路板上。视频中可以看到来自PNY、技嘉、华硕等厂商的显卡，甚至包括英伟达自家的Founders Edition公版产品。
从现场画面来看，这并非普通电脑爱好者在工作台上进行的个别改装，而是具有一定规模的生产作业。大量显卡被集中处理，改装后的产品还会经过测试，以确认其能够正常运行。原始视频据称曾在社交媒体上发布，随后被删除，但HKEPC保留了相关影像。
改装的核心原因在于，普通游戏显卡的设计与数据中心的需求存在明显差异。RTX 5090采用体积庞大的多风扇散热器，通常占据三个甚至更多扩展槽位。在普通台式机中，这种设计有利于散热，但在需要密集安装多张GPU的服务器机箱中，却会占用过多空间，并使机箱内部的气流组织更加困难。
因此，改装厂会将原有GPU核心和显存芯片从零售版电路板上拆下，再焊接到重新设计的定制PCB上。新电路板通常采用更紧凑的双槽设计，并配备涡轮式散热器，将热空气集中排向机箱外部。这样一来，原本为游戏电脑设计的显卡就能更适合安装在多GPU服务器或AI计算系统中。
这类改造并不是简单地更换散热器。GPU核心和显存芯片采用复杂的封装与焊接工艺，需要使用专业设备进行拆卸、清理、重新植球和焊接。定制电路板还必须重新设计供电线路、元件布局、散热结构和接口位置，以满足高功耗GPU的运行要求。改装完成后，产品还需要经过功能测试和稳定性验证，才能用于实际计算任务。
RTX 5090采用英伟达Blackwell架构，搭载GB202 GPU核心，并配备32GB GDDR7显存。其显存容量和计算能力，使其能够承担部分大型模型推理、模型微调以及其他AI计算任务。对于无法通过正规渠道获得某些数据中心GPU的企业或研究团队而言，将消费级GPU改造成服务器加速卡，可能成为获取计算资源的一种替代途径。
不过，RTX 5090改装卡并不能完全等同于英伟达专门设计的数据中心GPU。两者在显存容量、散热冗余、可靠性设计、服务器管理功能、软件支持和长期运行保障方面可能存在差异。改装后的RTX 5090仍然主要保留原有的32GB GDDR7显存，并不会因为换上定制电路板，就自动获得更大容量或数据中心产品的全部功能。
近期流传的部分说法称，中国厂商能够将RTX 5090的显存扩展至64GB甚至96GB，但目前尚无可靠公开证据证明这些高容量改装版本已经实现稳定运行。最新视频展示的产品，按照现有资料判断，仍采用原本的32GB显存配置。
这一现象也与美国对先进计算芯片的出口限制有关。美国对华出口管制限制了部分高性能计算芯片进入中国市场，而消费级RTX 5090的计算性能又具有较高吸引力。因此，相关产品通过拆解和重新设计电路板，转变为适合AI服务器使用的加速卡，引发了外界对出口管制实际效果的讨论。
但仅凭视频中的产品外观和数量，还无法确定这些显卡的具体采购渠道、交易是否合法，或是否违反了相关出口管制规定。改装本身与芯片的来源、跨境运输方式及最终用途属于不同问题，需要结合具体事实判断。
与此同时，RTX 5090在全球市场的供应情况也受到关注。近期有报道称，这款旗舰游戏显卡在部分市场变得极为难买，零售价格明显高于官方建议售价。英伟达高管杰夫·费舍尔在10月中旬曾表示，RTX 5090并未停产，公司仍将继续供货，但供应链产能确实受到限制。因此，当前显卡供应紧张不能简单归结为所有产品都已停止生产。
如果大量RTX 5090被用于AI计算，而非进入普通游戏电脑市场，可能进一步加剧玩家购买旗舰显卡的困难。不过，视频并未提供这家工厂的实际产量、采购规模和市场占有率，因此尚无法量化其对全球零售供应的影响。
对于英伟达而言，这种改装也体现了消费级GPU与AI加速器之间日益模糊的界限。RTX 5090虽然面向游戏市场，但其强大的并行计算能力同样适用于部分AI工作负载。只要GPU核心和显存能够稳定工作，重新设计电路板和散热系统，就有机会让它进入原本并非为其设计的服务器环境。
总体来看，最新视频展示了中国厂商正在以较为工业化的方式，将零售版RTX 5090改造为更适合AI服务器使用的加速卡。这不仅反映出市场对高性能GPU计算能力的需求，也说明消费级硬件能够通过重新设计进入专业计算领域。不过，改装产品的实际规模、芯片来源、长期可靠性及其对游戏显卡供应的影响，目前仍缺乏足够公开数据。
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。
As the demand for artificial intelligence computing continues to grow, China has seen a large-scale transformation of the GeForce RTX 5090 game GPU into an AI Accelerator Card production line. According to the latest video material cited by Wccftech and other hardware media, the plant is dismantling the retail version of RTX 5090GPU, transferring the GPU core and VRAM chips to specially designed server circuit boards equipped with heat-dispersion systems suitable for use in data centres, so that the high end of GPU, which was intended for game players, can be used in AI calculations.
The latest video is shared by HK hardware media HKEPC. The images show that a large retrofit workshop contains a large number of retail editions of RTX 5090GPU, where workers dismantle GPU, remove the original radiator and transfer GPU and VRAM chips to custom circuit boards in accordance with the flow line process. The video shows GPUs from PNY, Tweedy, Huashu, etc., and even includes the home-based FOODERS EDITION products.
From the scene, this is not an individual modification made by ordinary computer fans on the workstation, but a production operation of a certain scale. A large number of GPU have been centrally processed and the modified products are tested to confirm that they are functional. The original video, allegedly on social media publish, was subsequently deleted, but HKEPC retained the relevant images.
The central reason for the modification is that the design of the ordinary game GPU differs markedly from the needs of the data centre. RTX 5090 uses a large multi-zeal radiator, which usually occupies three or more extended slots. In common desktops, this design facilitates heat dispersion, but in server boxes requiring intensive installation of multiple GPUs, it takes too much space and makes it more difficult to organize the air flow inside the boxes.
The retrofitting plant will therefore remove the original GPU core and VRAM chips from the retail circuit board and weld them to the redesigned customized PCB. New circuit boards are usually designed with more compact double-train designs and are equipped with turbines to concentrate thermal air outside the aircraft. As a result, GPU, originally designed for video games computers, would be more suitable for installation on multi-gPU servers or AI computing systems.
Such modifications do not simply replace radiators. The GPU core and VRAM chips use complex sealing and welding processes and require specialized equipment for dismantling, cleaning, replanting and welding. Custom circuit boards must also re-engineer power lines, component layouts, heat-dispersion structures and interface locations to meet the operational requirements of high-capacity GPS. Once the modifications have been completed, the product will also need functional testing and stability certification to be used for the actual calculation tasks.
RTX 5090 uses the Blackwell structure of Weidar, which carries the GB202 GPU core with 32GB GDDR7Z82Z. Its VRAM capacity and computing capabilities enable it to undertake some large model reasoning, model fine-tuning and other AI computing tasks. For enterprises or research teams that do not have formal access to GPUs for some data centres, the conversion of consumer-grade GPUs into server accelerators could be an alternative to computing resources.
However, the RTX 5090 retrofitting card is not exactly equivalent to GPU, a specially designed data centre in England. There may be differences between the two in VRAM capacity, heat redundancy, reliability design, server management functionality, software support and long-term operating security. The modified RTX 5090 retains the original 32GB GDDR7Z82Z and does not automatically acquire the full functionality of a larger capacity or data centre product by changing custom circuit boards.
Some of the recent rumours that Chinese manufacturers were able to extend RTX 5090 VRAM to 64 or even 96GB, there is no reliable public evidence that these high-capacity variants have stabilized. The products of the latest video display, judged by the information available, still use the original 32GBVRAM configuration.
This phenomenon is also related to United States export restrictions on advanced computing chips. U.S. export controls on China limit the entry of some high-performance computing chips into the Chinese market, while consumer-grade RTX 5090 is more attractive. Thus, the transformation of the products into a speed card suitable for the use of the AI server by dismantling and redesigning circuit boards has triggered external discussions on the practical effects of export controls.
However, on the basis of the appearance and quantity of the products in the video alone, it is not possible to determine whether the specific procurement channels, transactions or violations of the relevant export control provisions of GPU are legal. The modifications themselves are different from the source, cross-border transport and end-use of the chips and need to be judged in relation to specific facts.
At the same time, the supply of RTX 5090 in the global market was also of concern. There have been recent reports that the flagship game GPU has become extremely difficult to purchase in some markets, with retail prices significantly higher than those officially recommended. In mid-October, Executive Director Jeff Fisher stated that RTX 5090 had not ceased production and that the company would continue to supply but that supply chain capacity was indeed limited. As a result, the current supply strain of GPU cannot simply be attributed to the fact that all products have ceased to be produced.
If a large number of RTX 5090 are used in AI calculations instead of entering the ordinary game computer market, it may further aggravate the difficulty for players to purchase the flagship GPU. However, the video does not provide the actual production, procurement scale and market occupancy of the plant, and therefore the impact of quantization on global retail supply is not yet available.
In the case of Weeda, this modification also reflects an increasingly blurred line between consumer-grade GPU and AI accelerators. RTX 5090, while oriented towards the game market, its strong parallel computing capability applies equally to part of the AI workload. As long as GPU core and VRAM are able to stabilize their work and redesign circuit boards and heat-dispersion systems, there is an opportunity for them to enter a server environment that was not originally designed for them.
Overall, the latest video shows Chinese manufacturers being converted from retail version RTX 5090 to an accelerator more suitable for the use of the AI server in a more industrial way. This reflects not only the market demand for high performance GPU computing capacity, but also the ability of consumer-grade hardware to enter the field of professional computing through redesign. However, the actual size of the retrofit product, the source of the chip, long-term reliability and its impact on the supply of the game GPU still lack sufficient open data.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.