Gemma-4-31B-it on Jetson AGX Thor — BF16, vLLM vision-language inference benchmark

Open this configuration in the VLM Inference Benchmark Explorer

Model	Gemma-4-31B-it
Parameters	31B
Device	Jetson AGX Thor
Quantization	BF16
Image size	720p
Images per camera	1
Max Cameras (response ≤ 3 s)	8
Engine	vLLM

Cameras	Response (s)	TPS (tok/s)	Target
1	0.98	3.9	≤3 s
2	1.31	3.8	≤3 s
4	1.52	3.7	≤3 s
8	2.11	3.5	≤3 s
16	4.38	3.2	>3 s
32	6.53	2.7	>3 s
48	8.94	2.3	>3 s
56	9.62	2.2	>3 s
60	10.3	2.0	>3 s
64	—	—	Failed

Images per request	Response (s)	TPS (tok/s)	Target
1	0.98	3.9	≤3 s
2	1.44	3.8	≤3 s
4	2.74	3.8	≤3 s
8	5.51	3.8	>3 s
16	11.8	3.8	>3 s
32	33.0	3.8	>3 s
64	124.1	3.7	>3 s
128	—	—	Failed

Gemma-4-31B-it (31B parameters), served in BF16 format with vLLM on Jetson AGX Thor, starts answering a single camera that sends one 720p image after 0.98 s, then writes 3.9 tok/s.

With 64 images per camera at 720p, the answer starts after 124.15 s. With one 2K image, after 1 s.

At the default target of an answer starting within 3 s (720p, one image per camera), it keeps up with 8 cameras at once.