Memory and CPU usage by NuPic


I’m thinking about using NuPic in production for anomaly detection in various Graphite metrics and I would like to know how much memory and CPU should I expect the NuPic platform to use?


NuPIC works on sparse data structures. So the exact CPU and memory usage depends on your data and network parameters. You need to try to know how much resource it needs.

1 Like

Thanks a lot for your response!
This is totaly understandable. I was hoping to learn from other people experience. Can you please share some from your personal experience?


1 Like

I haven’t used NuPIC in a while really. (I’m an author of another HTM framework, so… yeah)

But in the absolute worst case of feeding totally random data. NuPIC’s Temporal Memory implementation can become quite slow, ~100ms with 1024 columns, 16 cells per column. But the situation is very rare. A average case (predicting characters in NLP) is about 30ms.

HTM is mostly bounded by the memory bandwidth available to the CPU. So your results may vary. If you are interested in HTM’s performance from a computer architecture’s view. I hope my project report of my HTM framework can help. Page 22 ~ 28 and contains a quick performance characteristic analysis and potential optimization proposals.