I’ve been working on whether advanced ML models used in computational biology can run outside centralized cloud infrastructure.
In a recent study, I evaluated running graph neural networks (GNNs) for protein–protein interaction analysis on GPU-enabled single-board computers (edge devices), instead of cloud GPUs. The goal was to understand feasibility, latency, and practical constraints rather than chasing benchmark scores.
What I observed:
Stable inference on edge hardware
Inference latency on the order of milliseconds
No dependency on cloud GPUs during execution
This raises some interesting questions:
Are edge devices underutilized for graph ML workloads?
Where does edge inference make sense vs. cloud execution for biological or scientific ML?
What trade-offs (graph size, memory, model depth) matter most in real deployments?
For context, here’s a longer write-up with code and system design notes:
https://dev.to/your-article-link
(replace with your Dev.to link)
In a recent study, I evaluated running graph neural networks (GNNs) for protein–protein interaction analysis on GPU-enabled single-board computers (edge devices), instead of cloud GPUs. The goal was to understand feasibility, latency, and practical constraints rather than chasing benchmark scores.
What I observed:
Stable inference on edge hardware
Inference latency on the order of milliseconds
No dependency on cloud GPUs during execution
This raises some interesting questions:
Are edge devices underutilized for graph ML workloads?
Where does edge inference make sense vs. cloud execution for biological or scientific ML?
What trade-offs (graph size, memory, model depth) matter most in real deployments?
For context, here’s a longer write-up with code and system design notes: https://dev.to/your-article-link (replace with your Dev.to link)
And the research paper (preprint): https://doi.org/10.21203/rs.3.rs-8645211/v1
Curious to hear thoughts from folks working on ML systems, edge computing, or scientific ML.