Closed mrgloom closed 6 years ago
This is relative CPU speedup for p3.8xlarge over Baseline, so it is for (p3.8xlarge's CPU) vs (Baseline's CPU) - i.e. (32 vCPUs) vs (i7 5960X). Measured workload wasn't using GPU at all.
So it's literally just x2.1 speedup of 32 CPU server over top i7 CPU? looks weird. Is it because of virtualization or is it something special about Photoscan?
I don't know for sure, but it seems that aws vCPUs are hyperthreaded, so 32 vCPUs ~ 16 cores/32 threads, while 5960X is 8 cores/16 threads, so x2 speedup looks quite expected.
But there are also possible memory bandwidth differences (number of channels), possible overhead of multiple sockets and many other possible reasons.
How CPU speed was measured?
For example for p3.8xlarge x2.1 speedup is for (4xV100 + 32 vCPUs) vs (GTX1080 + i7 5960X) ? or (32 vCPUs) vs ( i7 5960X)? or (1 vCPUs) vs ( i7 5960X)?