A proposed Linux 7.5 change would add a ChaCha path that processes up to 16 blocks at a time with 512-bit AVX-512 registers. Phoronix reported gains of up to 99% on a Ryzen AI Max Strix Halo with Zen 5, but that figure belongs to a specific ChaCha benchmark, not general CPU or application performance. On November 9, 2026, Phoronix reported that the patch was queued in Eric Biggers’s libcrypto-next branch and was expected to be submitted for the Linux 7.5 merge window. (Phoronix)

How the proposed 16-block path works

ChaCha is a stream cipher, and the proposed code processes more of its data blocks in parallel. The existing path handles up to eight blocks at a time in 256-bit ymm registers; the new path would handle up to 16 using 512-bit zmm registers. The patch description says the wider routine would be dispatched for inputs larger than eight blocks, with X86_FEATURE_PREFER_YMM included in its feature-selection logic. (Phoronix)

The narrower path had a specific historical rationale: Martin Willi’s patch description says 512-bit zmm instructions caused heavy core down-clocking on Skylake-X. The proposed wider option is intended for processors where that penalty is gone, while the existing path remains in place.

What the reported performance figures measure

Phoronix reported up to 99% improvement on a Ryzen AI Max Strix Halo with Zen 5 using the full 512-bit path. It also reported up to 39% on Zen 4 using a 256-bit path. (Phoronix)

A separate measurement came from Antonio Quartulli’s October 25, 2026, Linux kernel mailing-list message. For chacha_crypt() inputs of 1 KiB or more on a Ryzen 9 9950X, he reported throughput rising from about 5.2 GB/s to about 10.2 GB/s. Those rounded figures amount to roughly a 96% increase. (Linux-Kernel Archive)

For WireGuard, end-to-end tunnel throughput is the relevant measure; the figures above describe ChaCha benchmarks. They do not tell you how much faster a complete tunnel workload would run.

Treat the 99% figure as a result for one ChaCha benchmark on one named Zen 5 system, not as a forecast for your whole workload.

Separate VM results and a proposed test suite

Quartulli also reported a virtual-machine benchmark associated with work on OpenVPN data-channel offload. Its ChaCha20-Poly1305 throughput was about 10% higher at a 1500-byte MTU and 35% higher at a 9000-byte MTU. He said he withdrew his duplicate implementation patch in favor of Willi’s approach and asked that his separate KUnit test-and-benchmark patch be considered on its own. (Linux-Kernel Archive)

The proposed KUnit tests compared chacha_crypt() with chacha_block_generic() for lengths up to 4,100 bytes and random lengths up to 12 KiB. The test plan also included guard pages, in-place and split calls, and the RFC 7539 test vector.