How far can CUDA go, and can CUDA replace CPUs

CUDA will go on forever, as long as NVIDIA doesn't go out of business. CUDA will get better and better, and it will be more and more convenient to use. The latest CUDA SDK supports Unified Memory Management, which makes CUDA's programming closer to C++, and will definitely be better in the future.

I'd love to know about that, but there's no good alternative to him at the moment.

OpenCL performance is poor.

So for a while, CUDA still has a place.

Hello, cuda is used to do parallel computing, for example, you have a for loop, and the operations between each loop are independent, then this loop can be parallelized with cuda. I don't know much about the three names you mentioned, but as long as they all have parts that can be executed in parallel, then CUDA can definitely be used.

Note: CUDA is mainly based on C C++, and there is support for other languages, but it is not fully functional.

CUDA can be supported.

There is no need to set it separately in the graphics card driver, and the graphics card acceleration needs to be enabled in the software.

In GTX1080 discrete graphics mode, CUDA can be run.

OK! Most platforms don't check if you're on display.

As long as you have cuda, it's fine.

It's just that some have requirements for architecture, such as kepler later, or pascal .........

CUDA is co-processing, the GPU cooperates with the CPU to perform tasks together, the CPU will give Kernal** to the GPU to execute, and the control will be returned to the CPU after the end, so how can there be no CPU? (You try to use CUDA optimization to know what's going on) In the current computing mechanism is a must have a CPU, if the GPU as the core, you may be able to consider developing a new architecture (not feasible), after all, the CPU and the GPU instruction set are different, using the GPU as the core, whether from the hardware or software, is a challenge for developers.

At least not in the near future. GPUs are mainly strong in parallel numerical processing, and their ability to perform ordinary serial programs is far worse than that of general-purpose CPUs. Not to mention the operating system and software support.

1 can't run all, CUDA is nvidia's technology, AMD's graphics card will not support competitors' technology, otherwise it is strengthening the other party and weakening itself (even if NVIDIA authorizes AMD, AMD will not want it, just like NVIDIA will not want AMD's firefight technology).