You are correct in that 17.6 TFLOPS is an Rpeak figure. That is two FPUs, each returning 2 64-bit results per cycle, at 2 GHz time the 1100 G5s with two processors each.
What will be interesting to find is the actual Rmax result. That will, of course be dependent on the number of actual compute nodes that can be used (likely less than the full 1100), compiler and network efficiency. Cluster effieciency has improved with the improvements in software and networking, as well as CPU and memory performance. The better ones are getting over 50 percent efficiency (my term for Rmax/Rpeak). Just a couple of years ago that rate was under 30 percent. Really good clusters are approaching 60 percent. What will be exciting will be if the VT G5 cluster can gain an Rmax of over 10 TFLOPS. For the price/performance and short timeline of installation, that would be impressive.
Slashdot Mirror
You are correct in that 17.6 TFLOPS is an Rpeak figure. That is two FPUs, each returning 2 64-bit results per cycle, at 2 GHz time the 1100 G5s with two processors each. What will be interesting to find is the actual Rmax result. That will, of course be dependent on the number of actual compute nodes that can be used (likely less than the full 1100), compiler and network efficiency. Cluster effieciency has improved with the improvements in software and networking, as well as CPU and memory performance. The better ones are getting over 50 percent efficiency (my term for Rmax/Rpeak). Just a couple of years ago that rate was under 30 percent. Really good clusters are approaching 60 percent. What will be exciting will be if the VT G5 cluster can gain an Rmax of over 10 TFLOPS. For the price/performance and short timeline of installation, that would be impressive.