Ex-Twitter dev reminisces about finding 700 unused Nvidia GPUs after takeover

Tim Zaman, a former Twitter engineer now at Google DeepMind, recently shared his experience of discovering 700 unused Nvidia V100 GPUs in Twitter's data center shortly after Elon Musk's acquisition in 2022. The GPUs were found powered on but idle, described by Zaman as "the forgotten remains of an honest attempt to make a cluster within Twitter 1.0." This revelation highlights a significant waste of computing resources, as these powerful GPUs had been left without purpose for years. Zaman's post was inspired by news of xAI's Memphis Supercluster, which utilizes 100,000 Nvidia H100 accelerators for AI training. He noted the irony of Twitter having such a substantial amount of GPU power sitting idle while tech companies are racing to build larger AI training clusters. Zaman also pointed out that the GPUs were of the PCIe variety, which have lower bandwidth compared to the NVLink interfaced SXM2 form factor, raising questions about Twitter's decision-making in their procurement. He reflected on the challenges of managing large-scale GPU operations, emphasizing the importance of graceful failure management in such setups. The discovery serves as a reminder of the potential inefficiencies in tech companies' resource management, especially in the context of the growing demand for AI capabilities.