IQ quant performance

[ThatGuyWhoAsked]

Prior sentiment was that IQ quants were slower on apple silicon, however is that still true? My benchmarks show it being FASTER then the similar quality quantisation:

ggml_metal_library_init: using embedded metal library
ggml_metal_library_init: loaded in 0.007 sec
ggml_metal_device_init: GPU name:   Apple M3
ggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)
ggml_metal_device_init: simdgroup reduction   = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory    = true
ggml_metal_device_init: has bfloat            = true
ggml_metal_device_init: has tensor            = false
ggml_metal_device_init: use residency sets    = true
ggml_metal_device_init: use shared buffers    = true
ggml_metal_device_init: recommendedMaxWorkingSetSize  = 12713.12 MB
| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen3vl 8B IQ4_XS - 4.25 bpw   |   4.24 GiB |     8.19 B | Metal,BLAS |       4 |           pp512 |        190.79 ± 2.77 |
| qwen3vl 8B IQ4_XS - 4.25 bpw   |   4.24 GiB |     8.19 B | Metal,BLAS |       4 |           tg128 |         19.25 ± 0.13 |

and

ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices
ggml_metal_library_init: using embedded metal library
ggml_metal_library_init: loaded in 0.006 sec
ggml_metal_device_init: GPU name:   Apple M3
ggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)
ggml_metal_device_init: simdgroup reduction   = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory    = true
ggml_metal_device_init: has bfloat            = true
ggml_metal_device_init: has tensor            = false
ggml_metal_device_init: use residency sets    = true
ggml_metal_device_init: use shared buffers    = true
ggml_metal_device_init: recommendedMaxWorkingSetSize  = 12713.12 MB
| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen3vl 8B Q4_K - Small        |   4.47 GiB |     8.19 B | Metal,BLAS |       4 |           pp512 |        187.90 ± 0.98 |
| qwen3vl 8B Q4_K - Small        |   4.47 GiB |     8.19 B | Metal,BLAS |       4 |           tg128 |         18.68 ± 0.04 |

build: ecf74a841 (7220)

I also read that apply family 9 (m3 and m4) are faster at this ( I have m3) does anyone have an idea as to why it is faster?
Prior Discussion: #5617