SAM 3 throughput on a laptop GPU: about 7.8 s per 360 scan point (4 faces x 4 prompts, fp16)
Measured with transformers Sam3Model/Sam3Processor in fp16 on a laptop NVIDIA GPU: a 261-point Matterport model (1,044 skybox faces resized to 1024 px, 4 text prompts each) took 2,031 s, about 7.8 s per scan point or about 0.5 s per face-prompt including download. VRAM held at about 2.4 GB. Each prompt is a separate forward pass, so time grows linearly with prompts; drop prompts that don't work (see the reliability note) before adding hardware. For a building with 7-9 scanned suites this is 1-2 hours, so run it ahead of time and save per-model JSON so a restart skips finished models.
Update: 2.2x faster with identical boxes. Most of the 7.8 s was re-running the image encoder once per text prompt. Encode each face once, then run each prompt on the cached features: vis = model.get_vision_features(pixel_values=x['pixel_values'].half()); then per label: o = model(vision_embeds=vis, input_ids=t['input_ids'], attention_mask=t['attention_mask']) with t = proc(text=label, return_tensors='pt') tokenized once and cached. post_process_instance_segmentation works unchanged on o. On 12 faces (4 prompts each, fp16, RTX 5060 Laptop, transformers 5.2.0) the time went from 42.2 s to 19.1 s, and 12/12 faces gave the same boxes. So a 360 scan point now takes about 3.5 s instead of 7.8 s. Pass .half() yourself: get_vision_features does not cast pixel_values to the model dtype.
Measured end to end after the encode-once change, over a real batch: two worker processes on the same RTX 5060 Laptop GPU (fp16, transformers 5.2.0, 4 skybox faces x 4 text prompts per scan point), each looping over its own list of Matterport scans. Over 1119 scan points in 30 buildings, each worker averaged 5.85 s per point, including downloading the faces, the color check and saving annotated JPEGs. Combined that is about 2.9 s per point, roughly 1200 points an hour. A second process helps because one process leaves the GPU idle while it fetches images over the network. Correction to my previous comment: the 3.5 s per point I gave there was the old 7.8 s divided by the 2.2x speedup, not a measurement; the relative speedup (same boxes, 2.2x) was measured. The 90 buildings (about 3000 points) finished overnight with no failures.