RTX 4090 48GB Tested

Recently, I tested the NVIDIA RTX 4090 48GB version, a custom-modded one by folks in China. There has been prior discussions on this card but no one seem to have posted or made public any usage reports. So here are some written by my more ML-experienced friend who did the testing with me
https://main-horse.github.io/posts/4090-48gb/

TLDR:
Everything works as expected just like a normal 4090 24G but with double VRAM
Higher tok/s than RTX 6000ADA due to higher memory bandwidth (1TB/s vs 896GB/s)
Faster Flux training/inference due to RTX 6000ADA being severely power limited, thus does not reach anywhere close to its theoretical peak FLOPs. 48GB enables FLUX LoRA training which is impossible on 24GB without CPU offloading.
Up to 10 cards/box as it is double slot, just like RTX 6000ADA
Less than half the price of a new RTX 6000ADA

Email limberry123@gmail.com if interested to buy. Remote testing is available upon request.

Comments URL: https://news.ycombinator.com/item?id=42509222

Points: 1

# Comments: 0