GeForce RTX 2080 Ti gets 22GB upgrade to rival $1000 AI GPUs

A Japanese workshop offers to double the memory of an RTX 2080 Ti, from 11 to 22 GB for $282. A low-cost solution for AI, while recent dedicated GPUs cost three times as much.

22 GB of VRAM for $282, but not for gaming

In Saitama Prefecture, a workshop modifies GeForce RTX 2080 Ti cards by replacing their 11 1GB GDDR6 chips with Samsung 2GB modules. The process, priced at 45,000 yen (282 $ or 244 €), involves desoldering components, adjusting resistors on the PCB, and flashing a custom VBIOS. Not all PCB models are compatible.

The gain is zero for gaming: the memory bus remains at 352 bits, bandwidth at 616 GB/s, and the Turing GPU’s power (2018) stays the same. However, for AI, VRAM capacity is often the limiting factor. Upgrading to 22 GB allows loading quantized language models that exceeded the original capacity, avoiding slowdowns from spilling over to system RAM.

Unbeatable value for money… for AI

The workshop also sells pre-modified cards starting at $430. With a used RTX 2080 Ti at $250, the total comes to around $500, compared to over $1,000 for a recent GPU with 20 GB or more VRAM. On August 15, 2026, the workshop announced it had received 200 Samsung chips, enough to modify about 30 cards before restocking.

The cost per GB is $25, an attractive rate amid memory shortages. Yet, the aging Turing architecture lacks optimizations from newer generations (Ampere, Ada Lovelace). The modification remains a niche solution: useful for AI, but no miracle.

RecommendedGoogle unveils three new lightweight Gemini modelsNews · August 1, 2026

A second life for a legendary card

The RTX 2080 Ti, released in 2018 with an unusual 11 GB memory configuration, is particularly well-suited for this type of modification. Its PCB often supports doubling the VRAM, unlike other models. For developers or AI enthusiasts, it’s a cost-effective alternative to professional GPUs like the RTX A4000, which has less memory endurance but is more expensive.

This Japanese hack shows that innovation doesn’t always come from new products. By pushing the limits of an aging card, it provides a pragmatic solution to a current problem: the cost of local AI. It remains to be seen if other workshops will adopt the technique or if NVIDIA will respond.