RTX 2080 Ti魔改22GB显存实测！不卡了、但帧率也没高多少

Kuai Technology reported on November 1 that PC Games Hardware conducted an actual test on a modified version of RTX 2080 Ti 22GB. After doubling the video memory from the original 11GB to 22GB, modern games and local AI have improved. The modification plan for this card is to replace the original 11 1GB Samsung GDDR6 video memory with 11 2GB Samsung GDDR6 modules. The video memory capacity is doubled to 22GB. The TU102 core, 352bit bit width and 616GB/s bandwidth remain unchanged. Just change a configuration resistor to allow the graphics card to recognize the new video memory. There is no need to modify the BIOS, and the standard GeForce driver can be used directly. The game test selected 9 games with high graphics memory usage. The overall difference under the 1440p DLSS quality level is not big, but the 22GB version takes more stable time to render each frame. Microsoft Flight Simulator 2024 and Marvel's Spider-Man 2 have greatly improved, and the frame rate of the original 11GB card is about 20% lower due to video memory limitations. The gap further widens with the 4K DLSS performance profile and FSR frame generation. Many games experience severe lag on the 11GB card, and some games directly crash. The Magic Card avoids these memory-related issues, but the GPU performance itself is still a bottleneck. Gears of War E-Day only has 17 FPS. In the actual measurement of video memory usage, Spider-Man 2 was allocated a maximum of 13.9GB, Pragmata 13.2GB, and Microsoft Flight Simulator 2024 12.9GB, indicating that the original 11GB is indeed not enough under these settings. However, the test deliberately focuses on high-graphics memory scenarios, which does not mean that all modern games require more than 11GB. The AI ​​test uses ComfyUI to run Krea-2 Turbo and Qwen 3-VL 4B to generate images. When the card occupies more than 11GB, the modified card has obvious advantages. Under the highest load, the 22GB version produces the first image in about 315 seconds, and the original version takes nearly 540 seconds. For comparison, the RTX 5070 Ti completed the same task in about 85 seconds. The gap is partly due to the architecture. Turing GPU does not have native FP8 support. The test work