I am looking forward to consumer / producer grade Tensor Processing Units and holding off on a desktop / server purchase until we maybe see something like this.

I am looking forward to consumer / producer grade Tensor Processing Units and holding off on a desktop / server purchase until we maybe see something like this. With inference tasks it'd be nice to have something not as performance heavy and instead like 24gb vram or more. I know there are tricks in CUDA to allow access to system ram as a proxy but limited success in duplicating this across various setups.