Hi folks, I currently crunch for MilkyWay project but I like Einstein too. Strong FP64 is pretty much a requirement for MilkyWay, is this also true for Einstein tasks?
If you are talking about the FGRPB1G search using GPUs, only the final followup stage (<10% of the calculation) uses double precision. If the GPU doesn't have DP capability, the followup stage would use a CPU core - and take a lot longer. So DP capability is quite desirable. My impression is it doesn't need to be particularly 'strong'.
At the moment on an AMD HD7850 GPU running 2 concurrent tasks, the followup stage takes around 20-30 secs out of an overall crunch time of perhaps 28 mins (two tasks completed). This is about a fifth of what it used to be on previous data files so the nature of the data has changed somewhat recently. The overall crunch time has also reduced, but not so dramatically. The 28 mins used to be around 36 mins or so before the change.
If you are talking about the
)
If you are talking about the FGRPB1G search using GPUs, only the final followup stage (<10% of the calculation) uses double precision. If the GPU doesn't have DP capability, the followup stage would use a CPU core - and take a lot longer. So DP capability is quite desirable. My impression is it doesn't need to be particularly 'strong'.
At the moment on an AMD HD7850 GPU running 2 concurrent tasks, the followup stage takes around 20-30 secs out of an overall crunch time of perhaps 28 mins (two tasks completed). This is about a fifth of what it used to be on previous data files so the nature of the data has changed somewhat recently. The overall crunch time has also reduced, but not so dramatically. The 28 mins used to be around 36 mins or so before the change.
Cheers,
Gary.