Lenovo Yoga RX 7900 XT AI rig hits a system RAM wall

A home-built local AI project ended the way many ambitious hardware experiments do: with a practical compromise. After trying and failing to use a Lenovo M910Q desktop, the builder shifted to a Lenovo Yoga laptop, wired an AMD Radeon RX 7900 XT to the machine through the laptop’s M.2 slot using an external PCIe cable, and powered the graphics card with a separate desktop PSU. The goal was to run local chatbots on the GPU rather than depend on cloud services.

The setup was never a normal desktop build. The laptop’s M.2 slot was no longer available for its original storage duties, so a USB SSD was used to boot the system. For a short while, the arrangement behaved like a science project that had escaped the lab: the external GPU was recognised, a display output was possible, and the rig was visually striking enough to look like a laptop that had been turned into a PCIe development board.

Several elements of the project are clearly established. The graphics card involved was an AMD Radeon RX 7900 XT with 20GB of VRAM, the power supply used for the setup was a DeepCool PL750D, and the connector hardware was an ADT-Link PCIe external cable assembly. The builder also removed the bottom panel of the laptop to make the connection work, which is the kind of step that makes the entire project feel equal parts clever and reckless.

Once everything was connected, the machine was able to output a picture through the external GPU. One reported test run with Furmark was described as reaching 500 FPS at 1080p, which shows that the link-up itself was functioning well enough for graphics output. However, that success did not translate into a smooth local-AI machine, because the display framebuffer also consumed some of the very VRAM that was supposed to be reserved for model inference.

The bottleneck was not only the graphics card. The builder tried running fairly large local language models, including Qwen 35B A3B and GLM-4.7 Flash, alongside a 128K-token context window. That combination demands more than GPU VRAM alone, because a long context window also needs a large amount of supporting data to live in system memory. In this case, the laptop’s available DRAM was the limiting factor.

As the work continued, the machine exhausted its RAM and began leaning on swap space. Once that happened, performance fell sharply and the GPU stopped staying busy. The reported behaviour was that the graphics card became idle except for brief bursts, which is exactly what happens when a system is waiting on slow storage-backed swap instead of keeping the inference pipeline fed with fast memory.

For Indian technology buyers, the story is less about spectacle and more about system balance. A strong GPU is only one part of a local AI setup. If the plan is to run chatbots on-device, the whole platform needs enough RAM, sensible storage, stable power delivery and a motherboard or enclosure that can support the workload without resorting to awkward workarounds. The builder’s experience is a reminder that a laptop can be a useful test bench, but it is rarely the best long-term host for a desktop-class GPU.

This also matters because local AI has become a serious hobbyist and productivity trend. People want privacy, lower latency and the ability to work offline. But the hardware stack has to match the ambition. In practical terms, that usually means enough system memory and a proper desktop chassis will matter more than a dramatic cable trick. For buyers in India comparing upgrade paths, a conventional desktop build may deliver a far cleaner result than trying to force a thin-and-light machine into acting like an AI tower.

Some details remain unspecified in the available evidence. The exact Lenovo Yoga model is not identified, and the laptop’s DRAM capacity is not given. It is also not clear which operating system or local AI software stack was used to manage the models, beyond the model names themselves. That matters because different runtimes and memory-management approaches can change how well a build survives under load.

The reported $550 purchase price for the Radeon RX 7900 XT is also a single data point rather than a verified market-wide benchmark. It should be treated as a reported purchase price, not as a general India market reference. Likewise, the Furmark result is a reported test outcome from the build, not a standardised benchmark that can be used to compare this setup with a proper desktop workstation.

The biggest lesson is simple: local AI depends on more than the graphics card. Even a 20GB GPU can be held back if the machine does not have enough system RAM or if the storage path becomes too slow once swap is involved. If you are planning to run local models, the safer approach is to think in terms of a complete platform: GPU memory, DRAM, CPU support, cooling, storage speed and power delivery all need to scale together.

For Indian buyers, that usually means spending more time on the full build sheet and less time chasing a single headline component. A used tower, a properly sized motherboard and enough memory will often make more sense than adapting a laptop that was never meant to hold a desktop graphics card. The result may be less dramatic on a desk, but it is far more likely to deliver the day-to-day stability that local AI actually needs.

This Lenovo Yoga build was never meant to be sensible. It was an experiment in making a laptop behave like a desktop GPU host, and for a while it succeeded in the most entertaining way possible. The Radeon RX 7900 XT could be wired in, the screen could light up, and local AI chatbots could at least be brought into the conversation. But the system’s limited DRAM eventually turned the whole project into a slow-motion lesson in memory hierarchy.

In the end, the builder did what many enthusiasts eventually do when a proof-of-concept outgrows the host machine: they moved the expensive parts into a more appropriate desktop platform. For readers in India, the story is a useful reminder that local AI enthusiasm is best matched with balanced hardware, not just an impressive GPU and a lot of optimism.