Re: Is one wu gpu better then one wu nacI/smp for research?
Posted: Fri Apr 22, 2016 7:07 pm
Though you didn't mention the Android client, I'll include it in my answer.
The value to science depends on the hardware capabilities. Slower {CPUs/GPUs} with fewer threads produce less than faster [CPUs/GPUs] with more threads. As a general statement, GPUs are more powerful than CPUs which are more powerful than the processor in a typiclal phone.
The nacl client is a SMP client of sorts. I'm not sure how it establishes how many CPUs to use. The SMP client can be managed explicitly so you may be able to get more work out of the same hardware. Both of them suffer if you allocate N CPUs if, because of some other long-running process(es) only a smaller number of CPUs are available on a regular basis. GPUs generally don't have the same limitation.
At one time, the GPU client was limited to certain types of analysis whereas the CPU client was more versatile. The vsrsatility of the FAHCores for GPUs has been improved very significantly and I believe they're now equivalent.
When developing a new project, some consideration is also given to maintain a reasonable balance between projects available for client X and the resources being donated by donors running that client. Then, too, the project's due-date may be considered based on the ability of that client to complete the project on schedule.
The value to science depends on the hardware capabilities. Slower {CPUs/GPUs} with fewer threads produce less than faster [CPUs/GPUs] with more threads. As a general statement, GPUs are more powerful than CPUs which are more powerful than the processor in a typiclal phone.
The nacl client is a SMP client of sorts. I'm not sure how it establishes how many CPUs to use. The SMP client can be managed explicitly so you may be able to get more work out of the same hardware. Both of them suffer if you allocate N CPUs if, because of some other long-running process(es) only a smaller number of CPUs are available on a regular basis. GPUs generally don't have the same limitation.
At one time, the GPU client was limited to certain types of analysis whereas the CPU client was more versatile. The vsrsatility of the FAHCores for GPUs has been improved very significantly and I believe they're now equivalent.
When developing a new project, some consideration is also given to maintain a reasonable balance between projects available for client X and the resources being donated by donors running that client. Then, too, the project's due-date may be considered based on the ability of that client to complete the project on schedule.
