YOLOv4
- Latency
- 2.39ms6.6× vs 15.8
- mAP@50
- 0.639+12 vs 0.519
- Energy
- 0.21J2.2× less
Chip
Tenstorrent Wormhole n150
tenstorrent/wormhole-n150- Compute cores
- 64 of 72 Tensix
- On-chip SRAM
- 96 MB
- Memory
- 12 GB GDDR6
- Card
- PCIe, 160 W
Our run used 3.6% of matrix peak; 66% of conv time waits on DRAM weights.
Coming soon: Tenstorrent Blackhole p150, Axelera AI Metis
Live gateway at zerolat-api.fly.dev. The Tenstorrent backend runs during demo windows; the key below is a placeholder, ask us for a real one.
Model ID
zerolat/yolov4-wormholeBase URL
https://zerolat-api.fly.dev/v1API key demo
zl_demo_••••••••••••B3fAcurl https://zerolat-api.fly.dev/v1/detect \ -H "Authorization: Bearer $ZEROLAT_API_KEY" \ -F model=zerolat/yolov4-wormhole \ -F image=@street.jpg \ -F conf=0.25POST /detect parameters
imagefile- requiredJPEG or PNG, sent as multipart form data
modelstring- requiredModel ID, e.g. zerolat/yolov4-wormhole
confnumber- 0.25Drop boxes scoring below this
iounumber- 0.45IoU threshold for non-maximum suppression
max_detinteger- 100Most boxes returned per image
{ "id": "det_8f3k2m9q", "model": "zerolat/yolov4-wormhole", "chip": "tenstorrent/wormhole-n150", "latency_ms": 2.39, "boxes": [ {"class": "person", "score": 0.91, "box": [412, 188, 538, 471]}, {"class": "bicycle", "score": 0.84, "box": [398, 302, 590, 489]} ], "proof_id": "prf_yolov4_wh_coco1k"}