2 min read
Add as a preferred source on Google

Google’s AI-Rewritten AV1 Decoder Ran Up to 30% Faster in Tests

The Rust version matched output across a broad conformance suite. The first Rust port ran about four times slower than the original C++ code.

Two test boards display matching paused landscape frames / TokenPost.ai
Two test boards display matching paused landscape frames / TokenPost.ai

Google’s Antigravity /teamwork system rewrote libgav1, the company’s AV1 decoder, in Rust, and its optimized version ran 20% to 30% faster than the original C++ and hand-optimized assembly on sample clips.

The rewrite contains about 170,000 lines of Rust. Its first version ran about four times slower than the C++ original. Google then profiled and optimized the code with a focus on speed across sample clips, testing the result on x86_64 and ARM systems.

The optimized decoder produced bit-for-bit matching output across a broad conformance-test suite. That result covers the reported suite; it does not establish identical output for every possible AV1 input. The benchmark details do not name processor models, compiler settings, the clips, sample counts, thread counts or timing distributions. The test date is also unspecified, and the available details are not enough to reproduce the performance result.

This result differs from Google’s Sept. 30 Argon effort. That work replaced 32,000 lines of SIMD code in an existing Rust version of libgav1, making it 2.7 times faster than an earlier Rust port while preserving video output. The September comparison used an earlier Rust implementation as its baseline; the November result compares the rewrite with the original C++ and hand-optimized assembly. The two speed figures describe separate comparisons.

The public libgav1 build requires a C++17 compiler. Google has not announced a production deployment or identified a Rust release.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…