University of MaineElectrical and Computer Engineering

DINGO one memory shared across a fleet of edge devices, that learns where data is needed

ACM Great Lakes Symposium on VLSIGLSVLSI 2025
Same request, repeated
The QA station keeps asking for the thermal frames of layer 212. They were captured by the thermal camera.
Standard database
Data stays where it was sensed.
–device hops
TIME
hops per request
DINGO
Moves data toward the device that keeps asking for it.
–device hops
TIME
hops per request
Every device hop is a network round trip. The paper charges 100 ms per hop on top of the local search time.
Simulated fleets of 2 to 32 devices, 3,200 searches each, versus Milvus and Chroma
Device hops per search, 32 devices
STANDARD
31
DINGO
22
Time per search, 32 devices
STANDARD
3.1 s
DINGO
2.2 s
about 30% faster
One graph that spans every device, searched as a whole
No central server and no global sync. Each device keeps working on its own slice when the link drops
Reinforcement learning keeps moving data toward where it is asked for, as the workload shifts
Numbers are from the GLSVLSI 2025 paper: virtual devices with a simulated network and a 100 ms cost per device hop, retrieval accuracy on par with Milvus and Chroma. The print cell shown here is an illustration of the intended use, not an experiment from the paper.
Funded by the National Science Foundation
For more details, contact
Sarah Glatter, lead student researcher sarah.glatter@maine.edu·Prabuddha Chakraborty, SIEGE Lab PI prabuddha@maine.edu