The demo is nice. The other thing is that a mapd server with 8 nvidia k40s for 3000 dollars each costs around 30000 (with CPU and main memory). For 30000 you can also buy a 4 socket server with a terabyte of main memory. Even though the GPUs can access their 300gb of memory much faster and have more cores, that won't help you because database workloads (filtering, grouping, joining) is not compute-intensive enough to give the GPU an advantage over CPU cores.
My bet is that a main memory database optimized for modern CPUs will do better on even larger datasets at the same server costs.
OP here. The demo is running off of the equivalent of a single K40. We can use 8 K80s to scale to billions of records with the same response time in milliseconds.
And database workloads are often heavily memory bound so the terabytes per second of memory bandwidth you can get on a GPU server can give you a huge speed up, and of course you have the prodigious compute power of the GPUs on tap if needed.
Has anyone actually been able to contact MapD? I've emailed, called, and left voicemails without any response. I'm not sure what is going on, but it's strange for a potential customer to have such difficulty getting in touch with sales.
We tried and it seemed to be working fine. Please email us at info at mapd dot com and we'll make sure to send you one. Also would be good to know what steps you went through so we can fix if necessary.
My bet is that a main memory database optimized for modern CPUs will do better on even larger datasets at the same server costs.