2 ms·

Are TPUs particularly useful for this kind of workload, compared to specialized encoders/decoders available on GPUs?