I think one of the promising technology in the next couple of years is the use of GPU for accelerating any kinds of job.
![](https://dl.dropbox.com/s/ujj6bhujt73uk57/mapd.jpg)
One of the company follows the direction is OmniSci (formerly MapD). They have a live demo showing how fast GPU processes almost 400 million tweets and visualizes them geographically in less than 1 second (Check this out: http://scl2-04-gpu03.mapd.com:8003/ ). @szilard covers the benchmark that indicates significant speedup.
![](https://i2.wp.com/github.com/szilard/benchm-databases/raw/master/plot.png?w=640&ssl=1)
I was wondering whether this blitz performance of GPU databases was true. For that reason, I need to benchmark one of them. OmniSci was chosen because it currently has a cloud option (so I don’t have to buy GPU to try it out). I was so lucky getting an OmniSci cloud account (2 GB GPU RAM, 8 GB system RAM) due to education affiliation. I used 7 million rows 2008 flight data (source: https://lnkd.in/eATNq-M). The query test I did is to aggregate those data with sufficient cardinality.
Here is the result:
1) harddisk-based database (mysql): 13 sec (min)
2) gpu-based database (omnisci):
– fully execute on cloud: 40 msec (max)
– execute on cloud, retrieve the result on local machine (+transport time from US to NL): 8.5 sec (max)
– execute on local machine’s container (docker omnisci community ed.) using NVIDIA GTX 650Ti: 400 msec (max)