I Doubled My GPU’s Memory…

Summary

Luke travels to Michigan to have GPU Lab upgrade two RTX 4090 graphics cards from 24GB to 48GB of VRAM each. Together, the cards provide 96GB—enough memory to run larger local AI models that might otherwise require much more expensive professional GPUs. The video frames the project as part of growing interest in local AI, self-hosting, and keeping computing and data under users’ control.

GPU Lab’s process is a specialized BGA rework, not a simple memory upgrade:

• The original GPU cores and memory are removed from the donor cards.
• The cores are transferred to custom PCBs designed for 48GB configurations. The memory layout is described as similar to the RTX 3090’s: chips on both sides of the board share data and address pins, with the GPU selecting between them.
• Memory chips are prepared, inspected, reballed, and soldered onto the new boards using flux, stencils, infrared reflow, and a BGA rework station. Greg notes that sourcing memory has become difficult, so the team uses tested chips salvaged from other boards.
• After the GPU core is optically aligned and soldered back on, the cards are cleaned, dried, assembled with compact, high-density coolers, and tested.

The work requires precision equipment and considerable hands-on expertise. The replacement coolers are smaller than standard 4090 coolers but use fast, loud fans; water blocks are another option. The cards still draw about 450W.

Testing and services

The upgraded cards boot successfully and pass a short Vulkan memory test. GPU Lab reports roughly 850–900GB/s under stress on Windows, compared with a theoretical maximum of around 1,050GB/s for the memory. Greg says this is within expectations and comparable to a standard 4090; Linux may report higher figures.

The video is a behind-the-scenes demonstration rather than a DIY repair guide. Greg says customers would send their cards to GPU Lab rather than perform the work themselves. He also mentions 4080 Super upgrades from 16GB to 32GB, A100 conversions, and future 170HX modifications.

Tool mentioned: Vulkan Memory Test, available free on GitHub.

Sponsor segment

The video also includes a RingConn promotion. It describes the Gen 2 ring as tracking sleep and breathing insights, and the Gen 3 as adding vascular trends and notifications for prolonged inactivity or other concerns. The ad mentions battery life of up to two weeks and no subscription requirement.

Main speakers and sources

• Luke — host and participant in the GPU upgrade.
• Greg (GPU Lab) — explains and performs the PCB, memory, and GPU-core rework.
• RingConn — sponsor segment about its Gen 2 and Gen 3 health-tracking rings.