Linum reposted this
From the performing arts to frontier AI ➡️ Linum's co-founders built a 2B parameter text-to-video model from scratch. Two people. Petabyte-scale data. No egress fees eating into their budget. Here's how they did it. Linum streams training data live from Cloudflare R2 directly to NVMe drives on each node, skipping dataset migration entirely. That strategy only works if the cloud provider's public internet uplink can actually keep up. On Crusoe, it worked. Linum benchmarked multiple providers and found Crusoe's internet uplink the fastest they'd tested. The result: data-loading bottlenecks that would have made training 25% more expensive elsewhere; gone. 🤯 Crusoe's Slurm-managed clusters handled multi-node job orchestration. NVIDIA HGX H200 GPUs with 141GB HBM3e handled the massive context windows that come with video data. And when an outage hit during the holidays, the Crusoe team proactively communicated fault attribution so Linum could stand down instead of debugging their own code. Read the full case study: https://bit.ly/4aSdLLp