Hacker News Reader

Yes, and it requires quite a low-level understanding of the architecture to "think GPU". SIMD, warps, blocks, threads, different memory types, no branching/identical branching per core, ... Some of this could probably be abstracted away but you definitely need to be aware and adjust algorithms appropriately. You can't just convert code and hope for the best, unless you just want a slow co-processor.

That's my experience in a nutshell. The cost of doing a cudamemcpy() far outweighs the advantages for computationally small tasks. The surprising bit for me was what's classed as a small task.

Decompressing a 5MP jpg then applying various filters is too lightweight a task to benefit. I thought that would be more or less a perfect GPU task, not so.

Running on OpenCL, the CPU with vector instructions horses a small GPU performance wise for this problem.

Probably not, typical java code is dominated by branching indirect code. Such code typically operates on mutable data which is not what GPUs are designed to do efficiently.

This should be a performance nightmare.