Read an article in Äripäev yesterday about new computer-cluster of Tartu University. It’s old-fashioned classical 42-node 1U 2 x 4core setup. each node has 2 x Intel Quad Xeon 5400 = 8 cores per node, 32 GB ram and 500GB HDD, interconnected via InfiniBand. TU reports total performance 0,84 TFLOPS.
Some math (and physics):
- CPU power consumption for CPU system
42 x 2 x 80 Watts = 6 720 Watts - GPU power consumption for GPU system
4 x 285 Watts = 1 140 Watts - Performance per Watt for CPU system
840 GFLOPS / 6 720 W= 0,125 GFLOPS / Watt - Performance per Watt for GPU system
7 200 GFLOPS / 6 720 W= 6,315 GFLOPS / Watt
Conclusion: GPU system uses 50 times less energy per GFLOPS than CPU system used by Tartu University.
So, its just 334 cores against my current configuration of 1440 cores. I really believe, that their choice was driven by status quo, sad, too mutch money has been spent to achieve only mediocre performance.
Read Tartu University news article
Advertisement
2 Comments
Comments RSS TrackBack Identifier URI
One size fits few, but not all. Most of the scientists making large computations are not computer scientists. They are not expert programmers nor are they dreaming to become one. Most of them are using PCs with Linux, doing science. Transition to some other architecture (like GPUs or Cell processors) would require rewriting a lot of code, and that means devoting some resources. Most of currently used proprietary code does not make use of GPUs.
That was the understanding of users needs when Tartu University started the HPC Center. It was decided, that the first machine for HPC Center would be general-purpose PC-based cluster with large memory, fast interconnect and with plenty of storage space. Its success would be measured by how many scientist start using it and how useful the machine can be for them. The actual CPU performance is quite secondary in this case.
Anyway, I hope that some computations will be too big for our new machine. If you don’t object then those jobs I would like to hand over to your GPU-based cluster.
P.S. In HPC Center’s new cluster, E5420 CPUs are used, therefore the power consumption per CPU is max 50W.
So, what would be the fortran compiler for GPU system? I would not mind testing my program somewhere on GPU cluster if possible.