The Intel SVT-VP9 Performance Boost Across 10 Intel/AMD Systems

As a follow-up from this weekend's article about Intel's SVT-VP9 video encoder running much faster on AVX2 CPUs from both Intel and AMD, here are the results now before/after for ten different systems with this open-source CPU-based VP9 video encoder.

The results shown this article as part of the daily performance tracker on LinuxBenchmarking.com while over the past two days did some additional SVT-VP9 benchmarks on ten different systems of both Intel and AMD processors of varying classes.

Besides the processors, there are hardware/software differences between the systems, so keep that in mind. This comparison isn't intended to be a scientific comparison between the CPUs but rather looking at the before/after performance on ten different systems for this quick, one-page article.

As you can see, the latest SVT-VP9 encoder is much faster on the AVX2 CPUs. With the Core i9 7980XE there is no change in performance since it is AVX-512 enabled and thus was already making use of the optimized SVT-VP9 code path that with this change just opened it up to running with AVX2. So with the latest Xeon Scalable changes, no performance changes are expected either - this optimization is about AVX2 CPUs like those from AMD and Intel processors without AVX-512. On the low-end the Core i5 8400 went from a 35 FPS encode speed to more than 120 FPS with the latest Git code while up the stack the Core i9 9900K went from 57 FPS to 176 FPS. AMD CPUs benefited greatly as well with the Threadripper 2950X going from 92 to 221 FPS or the EPYC 7502 2P from 189 to 340 FPS.

The SVT-VP9 performance for CPU-based VP9 encoding is looking incredibly competitive. A larger comparison will be coming in our next CPU review(s) on Phoronix.

For reference, here are the libvpx VP9 encode results on the same systems... Keep in mind this libvpx build is slightly older (current PTS test profile) and is of the first pass of the encode process and there may be default quality differences between the encode quality. (Also note the libvpx has an advantage in its default configuration of defaulting to -O3 over SVT-VP9 defaulting to -O2.) Again, today's testing is just a quick comparison. But taking that into account, SVT-VP9 is significantly faster than this reference VP9 encoder from Google on the larger systems. On the lower-end processors, there were some cases of the performance being similar but clearly SVT-VP9 is scaling better on the higher-end CPUs.

Overall, the direction of the SVT-VP9 performance is looking very positive. Fresh SVT-VP9 / SVT-HEVC / SVT-AV1 benchmarks, among other video encoder benchmarks, coming up soon in a proper multi-page featured article/review. The daily benchmarks continue to be done on LinuxBenchmarking.com via the Phoronix Test Suite.