National Tsing Hua University from Taiwan came into Dalian and grabbed the highest HPCG benchmark score, running away from the rest of the field like a scorched weasel.
The team scored an amazing 2,004.96 Gflop/s on a benchmark that is one of the most severe tests for any cluster. Team NTHU achieved this with their five node, 16 GPU system that was one of the burliest in the competition.
The high score from the previous cluster competition (SC18 in Dallas) was only 1,000 Gflop/s, posted by NTHU’s arch-rival Tsinghua. This is also an eyelash away from setting the all-time HPCG record of 2,080 Gflop/s set by Tsinghua at ASC18.
Beihang University and newcomer Southern University of Science & Technology (SUSTech) turned in the exact same score and thus tied for second place. It’s interesting that they both achieved the same number given that their clusters were completely different.
Team Beihang went with a three node system that was jammed full with 12 V100 GPUs while SUSTech built a seven node cluster only four V100s. Looks like SUSTech did a fantastic tuning job on HPCG, beating several teams who had much faster hardware. Congrats, SUSTech!
The mighty Tsinghua University grabbed fourth place with a HPCG score of 1,476.6, which is considerably lower than their worldwide record setting run of 2080.60. In fairness to them, they were running six nodes last year (vs. seven this year) and had 16 V100 GPUs in 2018 vs. 12 in 2019. So they were driving a less HPCG-centric box this year.
Fuzhou and Jinan also made the leader board with their scores of 1,373.90 and 1,345.01 respectively. Fuzhou brought the house when it comes to GPUs, running 16 on their six node cluster.
Fuzhou might have left some performance on the table. Given their configuration, you might expect them to get a bit bigger number and maybe even challenge for the top slot. But this is where experience comes in and Fuzhou is still a newish team when it comes to these competitions. Jinan’s configuration was a bit more modest – five nodes with 10 GPUs and it looks like they got everything they could from it. Good job, Jinan.
We have to tip our hats to home team Dalian for their HPCG performance. They finished in the top seven (out of 20), driving their four node, eight GPU cluster to a score 1,022. Great to see the home team show up on the leaderboard.
Next up we’re going to look at HPC applications they students ran and the results. Stay tuned for more exciting ASC19 coverage…..