Someone's turning dead $500 GPUs into 22GB VRAM monsters — here's the catch nobody's mentioning

A Hong Kong-based eBay seller is taking seven-year-old RTX 2080 Ti cards, physically doubling their memory, and selling them for about $500 with 22GB of VRAM on board — nearly double what Nvidia ever shipped on this GPU, and more than a stock RTX 4090 owner gets stock. Tom's Hardware confirmed the listings, and the seller's feedback shows at least 38 buyers have already received one. VRAM-hungry local AI users are the obvious target market. Whether that's smart depends entirely on what you're planning to run.

How you double VRAM on a GPU Nvidia never built for it

The RTX 2080 Ti shipped with 11 GDDR6 chips, each holding 1GB, wired across a 352-bit bus. The trick — and it long predates this particular seller — is that the TU102 GPU die underneath was never hard-limited to that capacity. Swap each 1GB chip for a 2GB module, reflow it back onto the board, and the silicon can physically address the extra memory. The problem is that the GPU doesn't know the chips changed unless something tells it.

That something is a set of strap resistors on the PCB — tiny components that tell the BIOS how much memory to expect at boot. Modders desolder and re-route them from a low/low/low configuration to high/high/low, and the board reports 22GB instead of 11GB. It's the same class of trick hardware modders have used for years on mining-era cards, just applied here specifically for the local-AI resale market. Nvidia never validated any of this, and there's no official firmware path for it.

The catch: Turing is seven years old

My honest take: the memory swap is clever and the bandwidth genuinely doesn't change, so once it boots, it boots at the same speed the card always ran at — just with more room. That's the whole value proposition. The risk sits entirely in three places: whether the specific board you receive was reflowed cleanly, whether Nvidia's driver stack keeps supporting Turing at all as newer architectures dominate the release notes, and whether llama.cpp and friends keep optimizing for a GPU generation that's rapidly aging out of active development. None of that is disqualifying. It's just the honest price of $500 for 22GB.

Run that table straight and the RTX 3090 wins almost every category that matters for local AI — it's faster, newer, and comes from an actual factory line. The only place the modded 2080 Ti wins is upfront cost, and even that gap shrinks fast once you factor in the real chance of a dead card with nobody to call. If you can stretch to a 3090, that's still the better $/performance buy on this site's own numbers; we walked through that exact trade in our used RTX 3090 vs. RTX 5060 Ti 16GB comparison.

Where the modded card actually makes sense: you already own a 2080 Ti, or you're specifically trying to fit a 20B–34B model at Q4 quantization that won't squeeze into 12GB, and $500 genuinely is the ceiling of your budget. In that narrow lane, 22GB opens doors 11GB or 12GB simply can't — check what fits at each quant level in our Q4 vs Q5 vs Q8 breakdown before you buy. Outside that lane, it's a novelty, not a plan.

The bigger story here is that demand for cheap VRAM has gotten strong enough that a cottage industry of GPU modders now exists specifically to serve it. That's not going away — if anything, expect more of these listings as GDDR7 prices keep climbing on new cards. Just go in knowing exactly what you're buying: a used, hand-modified, seven-year-old GPU with no factory backing. For most people reading this, our datacenter-GPU install guide or a straight used-3090 purchase is still the less stressful $500–800 local-AI move.