DanteGPU

Demo

About

The problem it solves ***DanteGPU*** democratizes access to high-performance computing by enabling individuals and organizations to monetize their unused GPU resources. Through a blockchain-powered marketplace, AI developers and researchers can directly access these distributed resources, ensuring efficient and equitable computing power utilization.

How it's made

Challenges we ran into **1- Large model weight downloads on limited bandwidth:** Downloading multi-gigabyte AI model weights over a 5–10 Mbps link repeatedly stalled my setup. How I overcame it: spun up a local HTTP mirror and used aria2c for segmented, parallel downloads—cutting total download time by ~60%. **2- Telemetry floods saturating the network:** Streaming per-second GPU metrics as JSON was clogging the client’s upstream bandwidth and causing UI lag. How I overcame it: switched to Protocol Buffers with LZ4 compression and batched updates every 5 s, which slashed bandwidth usage by ~75% while keeping the dashboard snappy. **3- Distributed dataset sync choking the link:** Transferring large training data shards between nodes would max out our 10 Mbps uplink. How I overcame it: adopted rsync over SSH with Zstd compression and enabled delta transfers—reduced sync payloads by ~80%. **4- Bulky Docker image pulls:** Pulling 4 GB container images on new GPU nodes repeatedly strained the band. How I overcame it: switched to a slim base image, leveraged multi-stage builds, and hosted images on a LAN registry—cut image size in half and pull times by ~65%. **5- Service discovery chatter:** Frequent UDP heartbeats for node discovery caused packet storms on constrained networks. How I overcame it: moved to a gossip protocol with adjustable TTL and backoff intervals, trimming discovery traffic by ~70%.