I would like to set up a new Linux test system/server to evaluate HSA and Boltzmann (hipify) software with regards to multi-GPU computing. Ideally, I would like it to be an APU + multi-dGPU system (as many GPU resources as possible!).
- Does HSA software support APU + multiple dGPU?
- What limitations are there? Can you operate on data between two or multiple dGPUs? How about between an APU and a dGPU?
- Are there requirements in HSA software that would limit the choices in my host configuration?
- I want as high performance as possible (e.g., low latency, direct memory accesses, unified memory addressing).
- What type of system should I build? The plan right now is Godaveri APU with many Radeon R9 Nano GPUs, however I am unsure how much support is available for Kaveri/Godaveri and whether I should wait for the desktop Carrizo APU/Bristol Ridge. Limitations on Linux support?
- Which AMD HPC GPUs include HSA support? Commercial APU support (are there any)?
- Are there any pitfalls or issues that I should be aware about?
- Do any of the questions or answers change when we talk about evaluating OpenCL instead of HSA?
- What about OpenCL and HSA applications running concurrently on the same system?