Measured Database — The Numbers, and What Did Not Work

2026年9月4日

This page contains advertising (affiliate links). See our Privacy Policy for details.

This page collects the numbers we actually measured, together with the setups that did not work. It is tables only, so you can compare the figures without reading the articles. Each row links to the article the number came from.

Measured numbers

Numbers measured on this blog, pulled out of the articles and put into one table. Each row links back to the article that produced it.

The same GPU is treated as a different entry when the link type differs (direct slot / OCuLink / Thunderbolt). Rows with an empty GPU field were measured per machine, not per GPU. Empty cells mean not measured – they are never filled with estimates.

Row	Measured	Machine	GPU	Link	Backend	Model	Context	Prefill [tok/s]	Decode [tok/s]	TTFT [s]	Load [s]	VRAM [GB]	Power [W]	n	Stat	Article	Raw data	Note
0001	2026-07-20	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	phi4-mini	1024	858.9	80.4		10.1		73.7	3	median	Article (JA)	`b580_full_suite.jsonl`
0002	2026-07-20	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	nemotron-3-nano:4b	1038	1424.4	93.7		5.6		88.6	3	median	Article (JA)	`b580_full_suite.jsonl`
0003	2026-07-20	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:8b	1031	1090.2	60		14		149.6	3	median	Article (JA)	`b580_full_suite.jsonl`
0004	2026-07-20	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	hf.co/deepreinforce-ai/Ornith-1.0-9B-GGUF:Q4_K_M	1021	617.6	47.5		91.7		135.1	3	median	Article (JA)	`b580_full_suite.jsonl`
0005	2026-07-20	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	gemma4-12b-fixed	1036	159	29.8		27		133	3	median	Article (JA)	`b580_full_suite.jsonl`
0006	2026-07-20	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	1031	385.7	9.1		29.2		61.9	3	median	Article (JA)	`b580_full_suite.jsonl`
0007	2026-07-20	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	nemotron-3-nano:30b	1038	42	13.7		43.7		44.5	3	median	Article (JA)	`b580_full_suite.jsonl`
0008	2026-07-23	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	phi4-mini	1024	2380.4	79.8		5.3		81.7	3	median	Article (JA)	`b580_full_suite.jsonl`
0009	2026-07-23	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	nemotron-3-nano:4b	1038	1650.8	95.5		5.6		85.2	3	median	Article (JA)	`b580_full_suite.jsonl`
0010	2026-07-23	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:8b	1031	1192.9	60.1		11.5		148.1	3	median	Article (JA)	`b580_full_suite.jsonl`
0011	2026-07-23	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	hf.co/deepreinforce-ai/Ornith-1.0-9B-GGUF:Q4_K_M	1021	1163.4	48		108.6		146.8	3	median	Article (JA)	`b580_full_suite.jsonl`
0012	2026-07-23	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	gemma4-12b-fixed	1036	396.9	29.9		17.2		132.9	3	median	Article (JA)	`b580_full_suite.jsonl`
0013	2026-07-23	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	1031	388.7	9.1		17.2		61.9	3	median	Article (JA)	`b580_full_suite.jsonl`
0014	2026-07-23	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	nemotron-3-nano:30b	1038	44.1	14.1		41.1		45.4	3	median	Article (JA)	`b580_full_suite.jsonl`
0015	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	2048		36					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=f16 / flash_attention=off
0016	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	8192		28.2					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=f16 / flash_attention=off
0017	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	16384		12.3					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=f16 / flash_attention=off
0018	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	32768		9.1					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=f16 / flash_attention=off
0019	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	65536		7.9					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=f16 / flash_attention=off
0020	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	131072		8					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=f16 / flash_attention=off
0021	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	2048		36.1					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q8_0 / flash_attention=on
0022	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	8192		28.3					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q8_0 / flash_attention=on
0023	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	16384		12.4					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q8_0 / flash_attention=on
0024	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	32768		9.2					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q8_0 / flash_attention=on
0025	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	65536		8.1					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q8_0 / flash_attention=on
0026	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	131072		7.8					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q8_0 / flash_attention=on
0027	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	2048		36.2					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q4_0 / flash_attention=on
0028	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	8192		28					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q4_0 / flash_attention=on
0029	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	16384		12.4					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q4_0 / flash_attention=on
0030	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	32768		9.1					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q4_0 / flash_attention=on
0031	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	65536		8.2					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q4_0 / flash_attention=on
0032	2026-07-29	Desktop PC	Intel Arc B580 12GB	直挿し	Vulkan	qwen3:14b	131072		8.2					3	median	Article (JA)	`b580_kvcache_quant.jsonl`	KV cache=q4_0 / flash_attention=on
0033	2026-07-07	GMKtec EVO-X2				phi4-mini:latest		1610.8	59.8		1.8	20.4			median	no article (raw data only)	`speed_3machines.csv`
0034	2026-07-07	GMKtec EVO-X2				qwen3:8b		1237.4	38		3.7	11.5			median	no article (raw data only)	`speed_3machines.csv`
0035	2026-07-07	GMKtec EVO-X2				qwen3:14b		746.8	22.7		4.2	16.3			median	no article (raw data only)	`speed_3machines.csv`
0036	2026-07-07	GMKtec EVO-X2				qwen3.6:27b		269.9	12.1		7.7	34.7			median	no article (raw data only)	`speed_3machines.csv`
0037	2026-07-07	GMKtec EVO-X2				gemma4:31b		162.5	10.1		9.1	23			median	no article (raw data only)	`speed_3machines.csv`
0038	2026-07-07	GMKtec EVO-X2				qwen3.6:35b-a3b		984.5	55.3		9.5	29.5			median	no article (raw data only)	`speed_3machines.csv`
0039	2026-07-07	GMKtec EVO-X2				llama3.3:70b		163.6	4.9		19.2	85.6			median	no article (raw data only)	`speed_3machines.csv`
0040	2026-07-07	GMKtec EVO-X2				gpt-oss:120b		1218.7	33.2		304.3	65.2			median	no article (raw data only)	`speed_3machines.csv`
0041	2026-07-07	GMKtec EVO-X2				MichelRosselli/GLM-4.5-Air:Q4_K_M									median	no article (raw data only)	`speed_3machines.csv`
0042	2026-07-07	GMKtec EVO-X2				qwen3-235b-q2:latest									median	no article (raw data only)	`speed_3machines.csv`
0043	2026-07-07	GMKtec EVO-X2				minimax-m2.7:latest									median	no article (raw data only)	`speed_3machines.csv`
0044	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gemma4:latest		1806.9	112.4		12.6	3.3			median	no article (raw data only)	`speed_3machines.csv`
0045	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gemma4-12b-fixed:latest		1067.4	66.2		9.7	8			median	no article (raw data only)	`speed_3machines.csv`
0046	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gemma4:26b		1341.5	80.3		23	17.4			median	no article (raw data only)	`speed_3machines.csv`
0047	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gemma4:26b-a4b-it-qat		1764.7	110.5		30.5	15			median	no article (raw data only)	`speed_3machines.csv`
0048	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gemma4:26b-a4b-it-q8_0		1247.6	70.2		29.6	23			median	no article (raw data only)	`speed_3machines.csv`
0049	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gemma4:31b		426.8	24.8		30.5	20.2			median	no article (raw data only)	`speed_3machines.csv`
0050	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gemma4:31b-it-qat		620.8	37		26.3	19.2			median	no article (raw data only)	`speed_3machines.csv`
0051	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gemma4:31b-it-q8_0		140	8		41.4	22.8			median	no article (raw data only)	`speed_3machines.csv`
0052	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		qwen3.6:27b		590.3	38.6		18.6	16.6			median	no article (raw data only)	`speed_3machines.csv`
0053	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		glm-4.7-flash:Q4_K_M		2265.6	110.1		28.4	19.1			median	no article (raw data only)	`speed_3machines.csv`
0054	2026-07-07	Desktop PC	RTX 3090 24GB	直挿し		gpt-oss:120b		556.2	17.3		89.2	23			median	no article (raw data only)	`speed_3machines.csv`
0055	2026-07-07	Apple Mac mini				qwen3:8b		623.1	19.4		5.1	5.7			median	no article (raw data only)	`speed_3machines.csv`
0056	2026-07-07	Apple Mac mini				qwen3:14b		359.7	10.9		7.1	9.3			median	no article (raw data only)	`speed_3machines.csv`
0057	2026-07-07	Apple Mac mini				gemma4:26b-a4b-it-qat		394.9	32.5		19.4	14.9			median	no article (raw data only)	`speed_3machines.csv`
0058	2026-07-07	Apple Mac mini				qwen3.6:27b		59.8	5.4		19.3	16.5			median	no article (raw data only)	`speed_3machines.csv`
0059	2026-07-10	Desktop PC				qwen3:8b		1061	127.5	3.39	3.2	5.2			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0060	2026-07-10	Desktop PC				qwen3:8b		21089.9	132.7	0.25	0.2	5.2			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0061	2026-07-10	Desktop PC				qwen3:14b		904.4	76.6	3.77	3.6	9.3			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0062	2026-07-10	Desktop PC				qwen3:14b		12850.2	78.6	0.26	0.2	9.3			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0063	2026-07-10	Desktop PC				qwen3.6:27b		560	39.1	5.86	5.6	17.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0064	2026-07-10	Desktop PC				qwen3.6:27b		2471.6	39.1	0.53	0.5	17.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0065	2026-07-10	Desktop PC				gemma4:31b-it-q8_0		269.5	8.8	32.59	32.1	33.8			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0066	2026-07-10	Desktop PC				gemma4:31b-it-q8_0		590.2	8.8	0.94	0.7	33.8			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0067	2026-07-10	Desktop PC				llama3.3:70b		127.4	3.5	13.36	12.1	42.5			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0068	2026-07-10	Desktop PC				llama3.3:70b		544.6	3.5	0.59	0.3	42.5			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0069	2026-07-10	Desktop PC				gpt-oss:120b		30.6	12.5	95.27	88.2	65.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0070	2026-07-10	Desktop PC				gpt-oss:120b		3764.2	17.5	0.61	0.6	65.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0071	2026-07-10	Desktop PC				qwen3-235b-q2:latest		6.3	3.5	131.51	107.4	85.7			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0072	2026-07-10	Desktop PC				qwen3-235b-q2:latest		908.8	6	0.41	0.2	85.7			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0073	2026-07-10	Apple Mac mini				qwen3:8b		3094.3	19.7	0.15	0.1	5.2			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0074	2026-07-10	Apple Mac mini				qwen3:8b		186.7	19.7	1.82	1	5.2			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0075	2026-07-10	Apple Mac mini				qwen3:14b		1782.2	11.3	0.19	0.1	9.3			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0076	2026-07-10	Apple Mac mini				qwen3:14b		100.1	11.4	2.82	1.2	9.3			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0077	2026-07-10	Apple Mac mini				gemma4:26b-a4b-it-qat		1580.8	33.5	0.35	0.3	15.6			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0078	2026-07-10	Apple Mac mini				gemma4:26b-a4b-it-qat		234.2	33.8	3.31	2.7	15.6			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0079	2026-07-10	Apple Mac mini				qwen3.6:27b		250.8	5.8	0.78	0.2	17.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0080	2026-07-10	Apple Mac mini				qwen3.6:27b		43	5.8	14.4	11.2	17.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0081	2026-07-10	GMKtec EVO-X2				qwen3:8b		847.8	39.7	1.6	1.4	5.2			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0082	2026-07-10	GMKtec EVO-X2				qwen3:8b		5699.3	39.2	0.15	0.1	5.2			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0083	2026-07-10	GMKtec EVO-X2				qwen3:14b		516.1	23.4	2.24	1.9	9.3			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0084	2026-07-10	GMKtec EVO-X2				qwen3:14b		3534.7	23.2	0.21	0.2	9.3			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0085	2026-07-10	GMKtec EVO-X2				qwen3.6:27b		214.2	12.2	4.29	3.6	17.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0086	2026-07-10	GMKtec EVO-X2				qwen3.6:27b		1060.4	12	0.38	0.2	17.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0087	2026-07-10	GMKtec EVO-X2				llama3.3:70b		105.4	5.2	18.09	16.6	42.5			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0088	2026-07-10	GMKtec EVO-X2				llama3.3:70b		774.8	5.2	0.37	0.2	42.5			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0089	2026-07-10	GMKtec EVO-X2				gpt-oss:120b		246.4	36.7	22.54	21.7	65.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0090	2026-07-10	GMKtec EVO-X2				gpt-oss:120b		6855.4	34.3	0.29	0.3	65.4			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0091	2026-07-10	GMKtec EVO-X2				qwen3-235b-q2:latest		66.4	19.6	29.56	27.3	85.7			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=cold
0092	2026-07-10	GMKtec EVO-X2				qwen3-235b-q2:latest		2762.9	19.2	0.19	0.1	85.7			median	Article (JA) / Article (EN)	`loadtime_3machines.csv`	start=warm
0093	2026-07-11	GMKtec EVO-X2	RTX 5060 Ti 16GB	OCuLink		phi4-mini			133			2.5			median	Article (JA)	`gpu_decode_compare.csv`
0094	2026-07-11	Desktop PC	RTX 3090 24GB	直挿し		phi4-mini			173			2.5			median	Article (JA)	`gpu_decode_compare.csv`
0095	2026-07-11	Desktop PC	RTX 3060 12GB	直挿し		phi4-mini			97			2.5			median	Article (JA)	`gpu_decode_compare.csv`
0096	2026-07-11	GMKtec EVO-X2	RTX 5060 Ti 16GB	OCuLink		qwen3:8b			78			5.2			median	Article (JA)	`gpu_decode_compare.csv`
0097	2026-07-11	Desktop PC	RTX 3090 24GB	直挿し		qwen3:8b			128			5.2			median	Article (JA)	`gpu_decode_compare.csv`
0098	2026-07-11	Desktop PC	RTX 3060 12GB	直挿し		qwen3:8b			61			5.2			median	Article (JA)	`gpu_decode_compare.csv`
0099	2026-07-11	GMKtec EVO-X2	RTX 5060 Ti 16GB	OCuLink		Ornith-1.0-9B			69			5.6			median	Article (JA)	`gpu_decode_compare.csv`
0100	2026-07-11	Desktop PC	RTX 3090 24GB	直挿し		Ornith-1.0-9B			97			5.6			median	Article (JA)	`gpu_decode_compare.csv`
0101	2026-07-11	Desktop PC	RTX 3060 12GB	直挿し		Ornith-1.0-9B			49			5.6			median	Article (JA)	`gpu_decode_compare.csv`
0102	2026-07-11	GMKtec EVO-X2	RTX 5060 Ti 16GB	OCuLink		qwen3:14b			45			9.3			median	Article (JA)	`gpu_decode_compare.csv`
0103	2026-07-11	Desktop PC	RTX 3090 24GB	直挿し		qwen3:14b			77			9.3			median	Article (JA)	`gpu_decode_compare.csv`
0104	2026-07-11	Desktop PC	RTX 3060 12GB	直挿し		qwen3:14b			35			9.3			median	Article (JA)	`gpu_decode_compare.csv`
0105	2026-07-11	GMKtec EVO-X2	RTX 5060 Ti 16GB	OCuLink		qwen3.6:27b			24			17.4			median	Article (JA)	`gpu_decode_compare.csv`
0106	2026-07-11	Desktop PC	RTX 3090 24GB	直挿し		qwen3.6:27b			39			17.4			median	Article (JA)	`gpu_decode_compare.csv`
0107	2026-06-26	GMKtec EVO-X2	Radeon 8060S (内蔵)	内蔵		gpt-oss:120b		1651.1	32.9		63.6			3	median	Article (JA)	`evox2-speed-20260626.csv`
0108	2026-06-26	GMKtec EVO-X2	Radeon 8060S (内蔵)	内蔵		qwen3.6:35b-a3b		1557.4	57.6		10.2			3	median	Article (JA)	`evox2-speed-20260626.csv`
0109	2026-06-26	GMKtec EVO-X2	Radeon 8060S (内蔵)	内蔵		gemma4:31b		250	9.9		19.4			3	median	Article (JA)	`evox2-speed-20260626.csv`
0110	2026-06-26	GMKtec EVO-X2	Radeon 8060S (内蔵)	内蔵		qwen3.6:27b		419.9	11.9		7.7			3	median	Article (JA)	`evox2-speed-20260626.csv`

The numeric data on this page is reusable under CC BY 4.0. Please credit "Kimono Oyaji's Lab" and the URL of this page.

Built: 2026-08-30

What did not work

Combinations that failed are kept as they are. Knowing that a given connection does not work is more useful before you buy than after.

This is not a record of things that worked. It lists only the combinations we bought, connected and failed to run.

The same GPU is treated as a different entry when the link type differs. The "Raw log" column points at the measurement log. Where that column says no dedicated log exists, the row was reconstructed from working notes.

Row	Measured	Machine	GPU	Link	What was tried	Where it failed	Error	Recovered?	Known cause	Article	Raw log
0001	2026-07-18	GMKtec EVO-X2	RTX 5060 Ti 16GB	Thunderbolt/USB4 (AG03)	Run CUDA compute (load a model with ollama)	Compute (detection and full 16311MiB VRAM were fine)	SSH応答消失・ping不達（ホストごとフリーズ）	No – required a physical reboot. But changing the OS worked: on Windows the same machine over the same link held 14,162MiB and generated (internal GPU plus two external cards)	The same card runs fine over OCuLink (78-133 tok/s, no hang). Not a bad card or driver – specific to this machine's USB4/TB path combined with NVIDIA compute DMA	Japanese article ／ English article	`aeb-results/evox2_tb_5060ti.md ／ aeb-results/bench_win_vs_linux_20260731.jsonl`
0002	2026-07-19	GMKtec EVO-X2	Intel Arc B580 12GB	Thunderbolt/USB4 (AG03)	Boot and bind the driver	Driver probe (PCI enumeration and the full 16G VRAM BAR succeeded)	Unable to change power state from D3cold to D0, device inaccessible / Failed to reset GuC state (-EPROTO) / probe failed with error -71	No – remove/rescan/manual bind did not recover it	PCIe config reads back as all-ones through the TB tunnel, so the device is unreachable; xe then mistakes it for SR-IOV VF mode and the GuC reset fails. Note OCuLink could not even enumerate it, so TB got further	Japanese article ／ English article	`aeb-results/evox2_tb_b580.md`
0003	2026-07-19	GMKtec EVO-X2	Intel Arc B580 12GB	OCuLink (AG01)	Connect and enumerate	Enumeration – it never appeared in lspci	（デバイスが列挙されない）	No	The link cannot provide the large BAR (ReBAR equivalent) that the B580 requires	Japanese article ／ English article	`aeb-results/evox2_tb_b580.md`
0004	2026-07-21	GMKtec EVO-X2	RTX 3060 12GB	Thunderbolt/USB4 (AG03)	Initialise the GPU	GPU init (enumeration, PCIe 4.0 x4 and driver bind all succeeded)	NVRM: osIovaMap: failed to map allocation (status = 0x59) → RmInitAdapter failed! (0x22:0x59:894) / nvidia-smi: No devices were found	No – but the host stayed up, unlike the 5060 Ti case. Changing the OS worked: the same card over the same link ran at 99.76 tok/s on Windows, 90.4% of OCuLink on Linux	Mapping into the I/O virtual address space fails. Dropping from Blackwell to Ampere did not help over the TB path	Japanese article ／ English article	`aeb-results/evox2_tb_3060_ampere.md ／ aeb-results/bench_win_vs_linux_20260731.jsonl`
0005	2026-07-18	Apple Mac mini M4	RTX 5060 Ti 16GB	Thunderbolt (TinyGPU / tinygrad 0.13.0)	Run a matmul through tinygrad	GPU hardware init (TB link, driver extension and NVDevice creation all succeeded)	AssertionError: Must be table pt=0x0 … pt.entry(pte_idx)=0xffffffffffffffff	No	An unresolved upstream tinygrad bug (issue #14338, closed but unfixed). GSP firmware 570.144 does not know the USB4 bus type, the VRAM training microcontroller fails, and the page table reads back as all-ones. Specific to GB206 – the RTX 5090 is reported working	Japanese article ／ English article	`aeb-results/mac_egpu_tinygpu.md`
0006	2026-07-20	Apple Mac mini M4	RTX 3060 12GB	Thunderbolt (TinyGPU / tinygrad 0.13.0)	Retry the same thing with an Ampere card	GPU hardware init (enumeration, PCIe4 x4 and device probe succeeded)	（5060 Ti と完全に同じシグネチャ）AssertionError: Must be table pt=0x0 … 0xffffffffffffffff	No	A record of our own prediction being wrong. On 07-18 we wrote that Ampere would likely work; it failed at exactly the same point in exactly the same way	Japanese article ／ English article	`aeb-results/mac_egpu_3060_ampere.md`
0007	2026-08-29	GMKtec EVO-X2	Radeon 8060S（内蔵・gfx1151）	内蔵	Run it on ROCm 10 (torch 2.10.0+rocm7.0)	At run time	Segmentation fault ／ /opt/amdgpu/share/libdrm/amdgpu.ids が存在しない	No – two workarounds were tried and neither helped. Our recommendation is to drop it	Not determined. On ROCm 6.4 the same machine returned is_available()==True but failed at run time with HIP error: invalid device function	no article (raw data only)	`専用の一次記録なし（snapshot.md 2026-08-30 に記載）`
0008	2026-08-24	GMKtec EVO-X2	Radeon RX 9060 XT	OCuLink	Attach it as an external GPU	Detection	（認識されない・OCuLink が Gen1 に落ちる）	Yes – removing the pci=nocrs kernel parameter brought the 9060 XT up and moved OCuLink from Gen1 to Gen4	The real cause was the pci=nocrs kernel parameter left over from the B580 experiments. We first misdiagnosed it as amdgpu.gttsize, which turned out to be unrelated	Japanese article ／ English article	`専用の一次記録なし（memory / snapshot.md に記載）`
0009	2026-07-31	GMKtec EVO-X2	RTX 3060 12GB	Thunderbolt/USB4 (AG03)	Move the IOMMU group to identity (pass-through) at run time as a workaround	Applying the setting, with both the GPU and its audio function unbound from their drivers	EPERM (errno 1) Operation not permitted ／書き込みは通るがグループは DMA のまま	No – but the original state was restored and verified, drivers rebound	EPERM is neither EBUSY (in use) nor EINVAL (structural limit). The kernel deliberately forbids taking a device behind an external link out of the translation domain	no article (raw data only)	`aeb-results/egpu_tb_retry_20260731.jsonl`
0010	2026-07-31	GMKtec EVO-X2	RTX 3060 12GB	Thunderbolt/USB4 (AG03)	Turn the IOMMU off entirely (reboot with amd_iommu=off) and attach the card again	Authorising the device – before enumeration; the GPU never appeared	認可を書いた直後に ssh 切断・ping 不達（ホストごと応答停止）	No – required a physical power cycle. The setting was rolled back to return to the safe-failure state	A record of our reasoning being inverted. The IOMMU was not blocking the eGPU, it was containing it: with the IOMMU on, the card failed safely at 0x59; with it off, the host froze at the authorisation step. Removing the guard is not a path to making it work on this machine	no article (raw data only)	`aeb-results/egpu_tb_retry_20260731.jsonl`
0011	2026-08-20	GMKtec EVO-X2	Radeon 8060S（内蔵・統合メモリ）	内蔵	Load a 133GB model (Qwen3-235B-A22B Q4_K_M) into 119GB of free unified memory	Model load – it ended after 250 seconds with rc=137	radv/amdgpu: Failed to allocate a buffer: size 948211712 bytes ／ rc=137	No – the next step (internal GPU plus the 9060 XT) took the whole host down and a physical power cycle was needed	The obvious wall: a model larger than the free memory does not load. The cost is high, though – no remote recovery. Because each step was flushed to disk, the stopping point survived	no article (raw data only)	`aeb-results/t27-235b-q4-20260820-2031.csv ／ aeb-results/evox2_t27_20260820.jsonl`
0012	2026-07-31	GMKtec EVO-X2	Radeon 8060S（内蔵・統合メモリ）	内蔵	Run a 108GB model (minimax-m2.7) on the internal GPU alone	Model load	OOM（内蔵のみの上限は約90GB）	Yes, in another configuration: on Windows with the internal GPU plus two external cards it generated at 8.03 tok/s. But 78% was handled on the CPU side – it runs, it is not usable speed	Adding external GPUs is not only about speed – it also puts models beyond the internal limit within reach	Japanese article ／ English article	`aeb-results/bench_win_vs_linux_20260731.jsonl`
0013	2026-07-29	Desktop PC	Intel Arc B580 12GB	PCIe スロット直挿し	Measure with the context stretched to 32,768 tokens (f16 KV cache)	Context creation – the model itself had loaded	failed to create context with model	No – dropping the KV cache to q8_0 still failed at 65,536	A model fitting does not mean its context fits. There was no room left in 12GB for the KV cache	no article (raw data only)	`aeb-results/b580_kvcache_llamacpp.jsonl`
0014	2026-07-27	Desktop PC	Intel Arc B580 12GB	PCIe スロット直挿し	Load a 19.70GiB model (Ornith-1.0-35B Q4_K_M) on this card alone	Model load	failed to load（12GiB に収まらず）	Yes in another configuration: split across layers together with the 3090 it ran, and the answers matched the 3090 alone exactly (8/8 capability, 10/10 coding). Speed was 34.4% lower than the 3090 alone	A capacity wall. This is where we separated the two effects: pooling costs speed but not correctness	Japanese article ／ English article	`aeb-results/mixed_gpu_3090_b580.jsonl`
0015	2026-07-27	Desktop PC	RTX 3090 24GB	PCIe スロット直挿し	Benchmark a 39.59GiB model (llama3.3:70b Q4_K_M, 80 layers) with the default settings	Model load	OOM（24GiB に載る上限は 46層。48層で落ちる）	Yes, with a different invocation: name the number of layers explicitly and it runs. Pooled with the B580 (36GiB) the limit rose to 69 layers and generation went 1.91 to 3.33 tok/s (+74%)	The benchmark tool defaults to putting every layer on the GPU. This is a limit of how it was called, not of the machine	Japanese article ／ English article	`aeb-results/mixed_gpu_3090_b580.jsonl`
0016	2026-07-27	Desktop PC	（GPU以前の問題）	—	Read the gemma4 models already on disk as standalone model files	Reading the file	不完全なファイル（31b: expected 1189 got 833 ／ 26b-a4b-q8_0: expected 1014 got 658）	Worked around by switching models (llama3.3:70b). Our 24-36GiB candidates had been skewed towards gemma4	The distribution format cannot be used directly as a standalone model file. Nothing to do with the card or the machine	no article (raw data only)	`aeb-results/mixed_gpu_3090_b580.jsonl`
0017	2026-08-17	GMKtec EVO-X2	3種すべて（RX 9060 XT／RTX 5060 Ti／内蔵 Radeon 8060S）	Thunderbolt ／ OCuLink ／内蔵	Load qwen3.6:27b, gemma4:31b and qwen3.6:35b-a3b	Model load	load_failed_build_format_nonsupport_all_gpu（MoE 版は …(MoE)）	No – the same result on all three GPUs	Not the link and not the card: the build in use did not support that model format. When every GPU fails identically, suspect the software	no article (raw data only)	`aeb-results/9060_b580_suite_20260817.jsonl`
0018	2026-08-17	GMKtec EVO-X2	RX 9060 XT 16GB ／ RTX 5060 Ti 16GB	Thunderbolt ／ OCuLink	Load a 17.2GB model (qwen3-coder:30b) and a 70B model on an external GPU alone	Model load	capacity_wall_external_gpu_only（16GB に収まらない）	No, not on an external card alone – the same model fits in the machine's unified memory	17.2GB does not fit on a 16GB card. Adding an external GPU does not by itself let bigger models run – what matters is whether it fits on the card you added	Japanese article ／ English article	`aeb-results/9060_b580_suite_20260817.jsonl`
0019	2026-07-15	Desktop PC	RTX 3090 24GB	PCIe スロット直挿し	Use VRAM as a disk (vramfs, OpenCL plus FUSE)	Allocating the area and reading and writing to it	20GB 要求に対し実確保は約2.9GB（単一割当上限）／負荷をかけると std::bad_alloc で異常終了	No – a 16GB swap area could not be created, so the measurement itself never happened	There is a cap on the size of a single allocation. In this combination, a VRAM-backed disk layer is not usable	no article (raw data only)	`aeb-results/vram_as_ram_ubuntu.jsonl`
0020	2026-07-02	Desktop PC	RTX 3090 24GB（＋RAM 62GB＋SSD）	PCIe スロット直挿し	Force a 202GB model (GLM-5.2 745B at roughly 1.5-bit) to run	Start-up warm-up – five hours without a single token	（エラーは出ない。終わらないだけ）	Yes – with –no-warmup it loaded in 41.2 seconds and answered correctly at 0.21 tok/s	We first concluded this was the limit at which generation becomes unreachable. That was wrong: the default warm-up was touching all 202GB and never finished. The real limit was speed, not whether it starts	Japanese article ／ English article	`docs/kimono/muriyari-notes.md`
0021	2026-07	GMKtec EVO-X2	Radeon 8060S（内蔵・統合メモリ）	内蔵	Speed up the same 202GB model by keeping more layers resident on the GPU	At run time – it dies immediately once too much is pinned	OOM（-ncmoe を 42 以下に下げた条件）	Yes, within a safe range: -ncmoe 50 held at 0.62-0.69 tok/s, and 45 still worked	On a unified-memory machine, more layers on the GPU stops helping past a point. The exact date was not recorded in the primary note (it is after 2026-07-03 in the same file)	Japanese article ／ English article	`docs/kimono/muriyari-notes.md`

This table is reusable under CC BY 4.0. Please credit "Kimono Oyaji's Lab" and the URL of this page.