DISCLAIMER: Expressed views on this blog are my own.
I wonder if it is worth the effort to use an Nvidia Tesla for automatic anomaly detection in time series data. Obviously, you do not want detection for all time series data because it has to be meaningful, but if you had a way to combine multiple anomalies with a rule based system then it should be effective for large amounts of data.
My idea is to throw this beast into 1-10 machines and have data fed to them from a TSDB in real time (as they are committed to the TSDB), so rules can be evaluated ASAP. TSDBs nowadays focus on you querying them for data instead of pub/sub. It is not to say querying isn't useful as it gives you history, but threshold based checks can be super quick. Pub/sub is for streaming and I think that is the way to go for building real time services on top of a TSDB.