The first task facing the university teams at the ASC18 Student Cluster Competition was running the venerable LINPACK benchmark.
It was a hard fought battle, but in the end, Taiwan’s National Tsing Hua University (NTHU) came out on top with a score of 42.99 teraflops.
Newcomer Qinghai University was a surprise second place with a 40.47 mark, with Sun Yat-Sen, a former Highest LINPACK winner taking home third place with a 37.42 score.
NTHU had an interesting configuration: four compute nodes with 12 total NVIDIA GPUs, evenly mixed between P100 and V100 models. In order to run both Tesla models at the same time, the students needed to slow down the V100s in order to get them in synch with the slower P100s.
Qinghai approached the competition with four nodes and ten NVIDIA V100s, but it wasn’t quite enough to overcome NTHU’s configuration and optimization. Sun Yat-Sen had a total of five nodes and 10 V100s, but the power required by the five nodes was just enough to keep the team from getting the full benefit of their accelerators.
Tsinghua University had a total of 16 NVIDIA V100s, which you’d think would be enough to lock down the number one slot for LINPACK, right? But the team was configured with 6 nodes, and since they had to power all the nodes whether they’re being used for LINPACK or not, this meant they couldn’t devote enough power to their GPUs to get the win.
From a historical perspective, the top score from ASC18 is a very good LINPACK. It’s a lot higher than the previous ASC LINPACK top score and also higher than the score set at ISC17.
However, it’s significantly below the 51.77 teraflops record set at SC17 by the mighty Nanyang Technological University.
Of course, Nanyang was packing a total of 16 P100s on only two nodes, so they had a configuration that was very well suited for the LINPACK benchmark.
In our next article, we’re going to take a look at how the ASC18 field dealt with HPCG – a benchmark that’s much tougher on the hardware and much more like real-life workloads. Stay tuned….