vulkan: tune MMVQ for Intel Windows

[0cc4m]

Tune MMVQ use for Intel Windows according to #17628 (comment)

@savvadesogle Please try it and see if performance is good.

[savvadesogle]

@0cc4m omg, omg! Of course!! Thank you. I will try tomorrow ♥️

[savvadesogle]

@0cc4m Hello. Done!

And

qwen3_8b_benchmark_results.csv

[0cc4m]

Is that good performance? Please compare it to master.

[characharm]

master

model	size	params	backend	ngl	fa	test	t/s
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	99	1	pp512	963.70 ± 5.96
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	99	1	tg128	52.15 ± 0.08
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	99	0	pp512	926.94 ± 6.03
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	99	0	tg128	51.50 ± 0.02

pr

model	size	params	backend	ngl	fa	test	t/s
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	99	1	pp512	970.71 ± 10.03
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	99	1	tg128	60.53 ± 0.08
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	99	0	pp512	924.87 ± 3.88
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	99	0	tg128	59.50 ± 0.10

[0cc4m]

Great, thank you!

[Cath0deRay]

I am seeing some significant improvements in token generation speeds on Arrow Lake H, excellent work, many thanks for all your efforts !

Benchmark Model: .\GGUF\qwen\Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf

load_backend: loaded RPC backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Arc(TM) 130T GPU (32GB) (Intel Corporation) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-cpu-alderlake.dll

model	size	params	backend	ngl	mmap	test	t/s
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	1	pp512	222.90 ± 0.37
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	1	tg128	18.28 ± 0.10

build: 2afcdb9 (8185)

Benchmark Model: .\GGUF\qwen\Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf

load_backend: loaded RPC backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Arc(TM) 130T GPU (32GB) (Intel Corporation) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-cpu-alderlake.dll

model	size	params	backend	ngl	mmap	test	t/s
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	1	pp512	223.78 ± 0.21
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	1	tg128	22.37 ± 0.43

build: feefb92 (8187)

---

Benchmark Model: .\GGUF\Qwen3.5-35B-A3B-UD-Q5_K_XL.gguf

load_backend: loaded RPC backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Arc(TM) 130T GPU (32GB) (Intel Corporation) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-cpu-alderlake.dll

model	size	params	backend	ngl	mmap	test	t/s
qwen35moe ?B Q8_0	23.21 GiB	34.66 B	Vulkan	99	1	pp512	165.12 ± 0.88
qwen35moe ?B Q8_0	23.21 GiB	34.66 B	Vulkan	99	1	tg128	11.57 ± 0.06

build: 2afcdb9 (8185)

Benchmark Model: .\GGUF\Qwen3.5-35B-A3B-UD-Q5_K_XL.gguf

load_backend: loaded RPC backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Arc(TM) 130T GPU (32GB) (Intel Corporation) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-cpu-alderlake.dll

model	size	params	backend	ngl	mmap	test	t/s
qwen35moe ?B Q8_0	23.21 GiB	34.66 B	Vulkan	99	1	pp512	165.68 ± 0.37
qwen35moe ?B Q8_0	23.21 GiB	34.66 B	Vulkan	99	1	tg128	12.76 ± 0.00

build: feefb92 (8187)

---

Benchmark Model: .\GGUF\qwen\Qwen3.5-27B-Q4_K_M.gguf

load_backend: loaded RPC backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Arc(TM) 130T GPU (32GB) (Intel Corporation) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-cpu-alderlake.dll

model	size	params	backend	ngl	mmap	test	t/s
qwen35 ?B Q4_K - Medium	15.58 GiB	26.90 B	Vulkan	99	1	pp512	52.51 ± 0.61
qwen35 ?B Q4_K - Medium	15.58 GiB	26.90 B	Vulkan	99	1	tg128	2.38 ± 0.02

build: 2afcdb9 (8185)

Benchmark Model: .\GGUF\qwen\Qwen3.5-27B-Q4_K_M.gguf

load_backend: loaded RPC backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Arc(TM) 130T GPU (32GB) (Intel Corporation) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-cpu-alderlake.dll

model	size	params	backend	ngl	mmap	test	t/s
qwen35 ?B Q4_K - Medium	15.58 GiB	26.90 B	Vulkan	99	1	pp512	52.84 ± 0.18
qwen35 ?B Q4_K - Medium	15.58 GiB	26.90 B	Vulkan	99	1	tg128	3.07 ± 0.00

build: feefb92 (8187)

---

Benchmark Model: .\GGUF\GLM-4.7-Flash-Q4_K_M.gguf

load_backend: loaded RPC backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Arc(TM) 130T GPU (32GB) (Intel Corporation) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\AI\bin\llamacpp\Vulkan\8185\ggml-cpu-alderlake.dll

model	size	params	backend	ngl	mmap	test	t/s
deepseek2 30B.A3B Q4_K - Medium	17.05 GiB	29.94 B	Vulkan	99	1	pp512	161.99 ± 21.63
deepseek2 30B.A3B Q4_K - Medium	17.05 GiB	29.94 B	Vulkan	99	1	tg128	13.21 ± 1.55

build: 2afcdb9 (8185)

Benchmark Model: .\GGUF\GLM-4.7-Flash-Q4_K_M.gguf

load_backend: loaded RPC backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Arc(TM) 130T GPU (32GB) (Intel Corporation) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\AI\bin\llamacpp\Vulkan\8187\ggml-cpu-alderlake.dll

model	size	params	backend	ngl	mmap	test	t/s
deepseek2 30B.A3B Q4_K - Medium	17.05 GiB	29.94 B	Vulkan	99	1	pp512	175.88 ± 0.72
deepseek2 30B.A3B Q4_K - Medium	17.05 GiB	29.94 B	Vulkan	99	1	tg128	17.31 ± 0.02

build: feefb92 (8187)

---