Hello all, I have a few questions about FP64 compute power.
Looking at the Xeon Phi processors, it looks like they offer high double precision TFLOPS. However, many times I've seen it suggested that GPUs are better at very large matrix multiplication in double precision, such as HPC/scientific compute. One of my problems with that statement at the moment is that it's been difficult for me to find any specific information on CPUs VS GPUs in terms of compute per watt per dollar. I can get a lot of metrics for GPU TFLOPS per watt per dollar, but they're all manufacturer reported. I can also find that many modern CPUs have AVX 2.0, allowing large integer calculations, but I'm not sure how capable they are of FP64 vs FP32 (i.e. in a i7-4770k or i7-6700k).
Some of the answer's I've found only caused more questions. For instance:
In FEMAP, FRRD is good to run on a GPU(specifically a W8100), but other modules are not.
the W8100 is fully FP64 functional core, but there are no specifics on CPUs. My assumption here is that the whole CPU is FP64 oriented, but the CPU isn't as powerful per watt per dollar as a GPU due to general compute capabilities GPUs do not have.
If you know of a table of single and double precision compute performance for modern CPUs, that would really helpful!
Looking at the Xeon Phi processors, it looks like they offer high double precision TFLOPS. However, many times I've seen it suggested that GPUs are better at very large matrix multiplication in double precision, such as HPC/scientific compute. One of my problems with that statement at the moment is that it's been difficult for me to find any specific information on CPUs VS GPUs in terms of compute per watt per dollar. I can get a lot of metrics for GPU TFLOPS per watt per dollar, but they're all manufacturer reported. I can also find that many modern CPUs have AVX 2.0, allowing large integer calculations, but I'm not sure how capable they are of FP64 vs FP32 (i.e. in a i7-4770k or i7-6700k).
Some of the answer's I've found only caused more questions. For instance:
In FEMAP, FRRD is good to run on a GPU(specifically a W8100), but other modules are not.
the W8100 is fully FP64 functional core, but there are no specifics on CPUs. My assumption here is that the whole CPU is FP64 oriented, but the CPU isn't as powerful per watt per dollar as a GPU due to general compute capabilities GPUs do not have.
If you know of a table of single and double precision compute performance for modern CPUs, that would really helpful!