ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 09:09:50 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
main
ik_llama.cpp/github-data/pull_requests/143 - Slightly faster IQ4_XS_R4 on AVX2.md
Thomas eaa2510a28 Add GitHub data: filename sanitization (#640)
2025-07-23 13:31:53 +02:00

661 B
Raw Permalink Blame History

🔀 #143 - Slightly faster IQ4_XS_R4 on AVX2

Author ikawrakow
State ❌ Closed
Created 2024-12-16
Updated 2024-12-16

Description

PPL-512(LLaMA-3.1-8B) on Ryzen-5975WX goes to 262.2 t/s up from 248.2 t/s.

On AVX2/Zen4 it is much better to interleave 8 rows - see this branch. We get 284 t/s on Zen4 and 275 t/s on AVX2. But the ARM_NEON implementation becomes extremely messy, and we get ~1-2% lower performance. Hence sticking with 4 interleaved rows for now.