.

Author: ahmadsharif1

README.md

@@ -120,6 +120,10 @@ The instructions below assume you're on Linux.
 
 ## Benchmark results
 
+The following results were obtained by running the [benchmark_decoders.ipynb](./benchmarks/decoders/benchmark_decoders.ipynb) on a lightly-loaded 22-core machine. We first get the operation latency for various seek and decode patterns in a loop from a single python thread and then compute the FPS (= 1 / latency). Error bars show the FPS of the p25 and p75 operation latency.
+
+If you are running multiple copies of the decoder (for example in a DataLoader) on multiple threads, it is best to use `num_threads=1` for the best performance. If you care about the operation latency (not throughput of many concurrent operations) you can set `num_threads=0` to utilize all cores of your computer.
+
 ![Benchmark Results](./benchmarks/decoders/benchmark_results.png)