The main advantage of CSV is that it is human-readable. But if you have this much data, then why not use some binary format that you can just read in as a blob, and don't even need to parse?

Also, the approach mentioned here is only useful if reading+parsing the data is the bottleneck (or close to it). For example, if reading+parsing takes only 10% of the total processing time, then optimizing this stage will only give at most a 11% increase in performance.