mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-02-20 13:14:09 +00:00
21 lines
693 B
Markdown
21 lines
693 B
Markdown
### 🔀 [#120](https://github.com/ikawrakow/ik_llama.cpp/pull/120) - Q8_0_R4
|
|
|
|
| **Author** | `ikawrakow` |
|
|
| :--- | :--- |
|
|
| **State** | ❌ **Closed** |
|
|
| **Created** | 2024-12-03 |
|
|
| **Updated** | 2024-12-03 |
|
|
|
|
---
|
|
|
|
#### Description
|
|
|
|
Following PR #118, #119: `Q8_0` repacked with 4 interleaved rows.
|
|
|
|
PP-512 for LLaMA-3.1-8B for `ARM_NEON` (M2-Max), `Zen4` (Ryzen-7950X) and `AVX2` (Risen-5975WX):
|
|
|
|
| Platform | Threads | Q8_0 | Q8_0_R4 | Speedup |
|
|
| ---: | ---: | ---: | ---: | ---: |
|
|
| ARM_NEON | 8 | 83.69 ± 1.53 | 112.95 ± 0.17 | 1.350 |
|
|
| Zen4 | 16 | 175.61 ± 0.71 | 268.98 ± 0.31 | 1.532 |
|
|
| AVX2 | 32 | 213.95 ± 0.44 | 234.40 ± 0.60 | 1.096 | |