MoE prefill on AMD Radeon 780M (gfx1103) via Vulkan — any real numbers?

[realaaronmorris]

Hardware: Ryzen 7 7840HS / Radeon 780M (gfx1103), 32GB DDR5-4800, Linux, Mesa RADV.

I'm trying to decide whether a 780M mini-PC is viable as a small always-on inference node, and the deciding factor is MoE prefill, which I can't find a single measurement of on this hardware.

What I've found for dense models on 780M (Gemma 4 12B QAT Q4_0, 7.6GB):

Backend	Prefill	Decode
ROCm	54 tok/s	4.5
Vulkan RADV	39 tok/s	4.7
CPU 8-thread	14 tok/s	5.3

So the iGPU gives ~3x prefill and essentially nothing on decode, which makes sense — decode is bandwidth-bound and the iGPU shares the same DDR5.

My question is about sparse MoE models (Qwen3-30B-A3B, Qwen3.6-35B-A3B, gpt-oss-20b — ~3B active params):

1. Does anyone have prefill numbers for an A3B-class MoE on a 780M (or any RDNA3 iGPU) via Vulkan?
2. Is GPT-OSS 20B and Qwen3 30B A3B prefill so slowly #15163 ("GPT-OSS 20B and Qwen3 30B A3B prefill so slowly") an Intel-specific issue, or does the same prefill collapse happen on AMD iGPUs?

Why prefill specifically: my workload sends ~20k-token prompts. At 150 tok/s prefill that's ~2 minutes; at the ~20 tok/s reported in #15163 it's ~17 minutes, which changes the buying decision entirely.

I've separately measured Qwen3.6-35B-A3B at 45.8 tok/s decode on an M1 Max (Metal), and on CPU-only I measured MoE running slower than a comparable dense model — presumably expert-gather destroying cache locality. I don't know whether that penalty carries over to an iGPU via Vulkan.

Any real numbers, even rough ones, would be appreciated. Happy to run and post benchmarks if I end up buying the hardware.

[Rickenbacker]

Here are some Vulkan and ROCm numbers from a MINISFORUM UM790 Pro. Numbers are for 0 and 32768 depth so your 20K prompt should land somewhere in between.

ROCm - build: c841aee (10687)

llama-bench -fa 1 -m gpt-oss-20b-mxfp4.gguf -d 0,32768

model	size	params	backend	ngl	fa	test	t/s
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	ROCm	-1	1	pp512	532.73 ± 4.92
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	ROCm	-1	1	tg128	26.03 ± 0.00
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	ROCm	-1	1	pp512 @ d32768	224.59 ± 3.49
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	ROCm	-1	1	tg128 @ d32768	15.28 ± 0.00

llama-bench -fa 1 -m ggml-org-Qwen3.6-35B-A3B-Q8_0.gguf -d 0,32768

model	size	params	backend	ngl	fa	test	t/s
qwen35moe 35B.A3B Q8_0	34.36 GiB	34.66 B	ROCm	-1	1	pp512	338.44 ± 2.31
qwen35moe 35B.A3B Q8_0	34.36 GiB	34.66 B	ROCm	-1	1	tg128	15.81 ± 0.02
qwen35moe 35B.A3B Q8_0	34.36 GiB	34.66 B	ROCm	-1	1	pp512 @ d32768	176.09 ± 2.29
qwen35moe 35B.A3B Q8_0	34.36 GiB	34.66 B	ROCm	-1	1	tg128 @ d32768	13.19 ± 0.01

Vulkan- build: c841aee (10687)

llama-bench -fa 1 -m gpt-oss-20b-mxfp4.gguf -d 0,32768

model	size	params	backend	ngl	fa	test	t/s
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	-1	1	pp512	403.22 ± 2.46
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	-1	1	tg128	29.39 ± 0.05
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	-1	1	pp512 @ d32768	62.31 ± 0.84
gpt-oss 20B MXFP4 MoE	11.27 GiB	20.91 B	Vulkan	-1	1	tg128 @ d32768	20.84 ± 0.02

llama-bench -fa 1 -m ggml-org-Qwen3.6-35B-A3B-Q8_0.gguf -d 0,32768

model	size	params	backend	ngl	fa	test	t/s
qwen35moe 35B.A3B Q8_0	34.36 GiB	34.66 B	Vulkan	-1	1	pp512	325.12 ± 2.65
qwen35moe 35B.A3B Q8_0	34.36 GiB	34.66 B	Vulkan	-1	1	tg128	20.48 ± 0.02
qwen35moe 35B.A3B Q8_0	34.36 GiB	34.66 B	Vulkan	-1	1	pp512 @ d32768	131.61 ± 5.37
qwen35moe 35B.A3B Q8_0	34.36 GiB	34.66 B	Vulkan	-1	1	tg128 @ d32768	16.68 ± 0.01