Chutes AI trains 8B model on 240 RTX 5090 GPUs for $6,500 via Bittensor network
- —240 RTX 5090 GPUs across 30 hosts in 13 countries trained an 8B-parameter model
- —Training cost about $6,500, or roughly $11 per 1 billion tokens
- —The 8B model generates 59.6 tokens/s on a smartphone CPU; a 40.75B model hits 26.9 tokens/s
- —Training relied on Parallax, a distributed system built on a libp2p P2P architecture
Why it matters: The demo suggests consumer GPUs in distributed networks can cut AI training costs, reinforcing the decentralized-compute narrative around Bittensor.
Source: TokenPost