Intel's Open-Source VP9 Video Encoder Just Scored A Massive ~3x Performance Boost

Intel's open-source team continues showing the power of optimizations... Or rather in this case, a three fold performance improvement due to previously limiting an AVX-512 routine that also works on AVX-2 CPUs. SVT-VP9 is now a lot faster on AVX2 CPUs from both Intel and AMD.

We were alerted today to this change to Intel's SVT-VP9 video encoder. Oddly enough the title is "Fix the perf gap for Epyc CPU" And, yes, the fix was contributed by an Intel developer.

This fix for the performance gap to AMD EPYC CPUs is basically opening up some motion estimation optimizations that were previously just limited to AVX-512 to now also work on AVX2 CPUs. So it's just not about AMD CPUs in reality, where there isn't yet any AVX-512 support, but also older Intel CPUs that are limited to AVX2.

That code went in nonchalantly but it turns out that it's to massive benefit for AVX2 CPUs including from Intel. As was pointed out our daily benchmarking of various open-source video encoders already flagged a massive shift in performance prior to myself manually looking through the data.

The Xeon E5-2687W v3 (Haswell), Xeon E5-1680 v3 (Skylake), and Core i7 7740X (Skylake-X) systems I have dedicated to this daily benchmarking of different video encoders all saw around a 3x speed-up thanks to this single code change. The i7-7740X went from 30 FPS to 120 FPS, the E5-1680 v3 from 38 to 113 FPS, and the E5-2687W v3 from 46 to 150 FPS. Damn!

This is easily the largest speed-up we've seen to SVT-VP9 in its history, at least as far as AVX2 CPUs are concerned. Unfortunately I don't have any AMD CPUs doing daily benchmarking of the SVT-VP9 code due to power/resource constraints, but given the interesting changes we've seen out of these Intel video encoders this year, I guess I'll add it to my TODO list / budget for adding some additional systems for this daily benchmarking. (Those wishing to support this independent effort can do so by joining Phoronix Premium or making a PayPal tip and making it known your interest in our video encoder benchmarking.)

SVT-VP9 was already performing well against other VP9 encoders so now is likely to win by a landslide victory. I'll be working on some tests on that front shortly. This is another example on Intel's way for aiming at a 1000x performance advancement from hardware to software work.