### 🔀 [#145](https://github.com/ikawrakow/ik_llama.cpp/pull/145) - IQ3_K_R4

| **Author** | `ikawrakow` |
| :--- | :--- |
| **State** | ❌ **Closed** |
| **Created** | 2024-12-17 |
| **Updated** | 2024-12-17 |

---

#### Description

Adding `IQ3_K` with 4 interleaved rows.

We get very signifiant performance gains on `ARM_NEON` and more modest gains on `AVX2/Zen4`. Overall slower than other `_R4` quants, which is expected as 3-bit quantization is always kind of slow. 

Here is `PP-512` for LLaMA-3.1-8B on `Zen4` (Ryzen-7950X), `ARM_NEON` (M2-Max) and `AVX2` (Ryzen-5975WX)

| Platform |  Threads | IQ3_K | IQ3_K_R4 | Speedup |
| ---: | ---: | ---: | ---: | ---: |
| ARM_NEON |  8 |  54.94 ± 0.79  | 93.83 ± 0.09 | 1.708 |
| Zen4            | 16 | 180.13 ± 0.48 | 230.33 ± 0.13  | 1.279 |
| AVX2           | 32 | 197.59 ± 0.43 |  253.36 ± 0.50  | 1.282 |

We get decent performance gains for TG as well.
Here results for TG-128 on LLaMA-3.1-8B with different numbers of threads:

| Platform |  Threads | IQ3_K | IQ3_K_R4 | Speedup |
| ---: | ---: | ---: | ---: | ---: |
| ARM_NEON | 2 |  5.84 ± 0.00 | 6.71 ± 0.05 | 1.149 |
|                      | 4 | 11.14 ± 0.00 | 12.83 ± 0.01 | 1.152 |
|                      | 8 | 20.59 ± 0.17 | 23.07 ± 0.16  | 1.120 |
| Zen4            | 1 |  5.06 ± 0.00  | 5.64 ± 0.00  |  1.115 |
|                      | 2 |  9.58 ± 0.01 | 10.50 ± 0.01  |  1.096 |
|                      | 4 |  16.56 ± 0.05  | 16.77 ± 0.32  |  1.013 |
| AVX2           | 2 | 4.45 ± 0.00   | 6.83 ± 0.00 | 1.535 |
|                     | 4 | 8.24 ± 0.00   |  12.51 ± 0.00 | 1.518 |
|                     | 8 |  14.59 ± 0.04  | 16.23 ± 0.00  | 1.112 |