cuFFT INVALID_DEVICE on 470/570/670

My team wrote some algorithms utilizing cuFFT on some 980 GPUs. They seem to work well and the customers were wanting them on some “older” workstations with a 470, 570 or 670. When the code runs on the 470-670, the plan gives an INVALID_DEVICE error. I am digging into the result, but I am assuming it has something to do with the CUDA compute engine (470=v2.0, 570/670=v3.0 and 980=v5.2). I have updated CUDA to v7.5 on the other cards and recompiled (successfully), but the error still occurs.

Can someone verify/squash, my thinking while I research more?


Are you building the binary to contain all compute capabilities or just one for the 980?