Building a deep learning workstation

Motivations

Daniel Whitenack embarked on building his own deep learning workstation, driven by both practical and personal motivations. He initially considered pre-built options from companies like Lambda Labs but found them costly, ranging from $10,000 to $12,000. Instead, he opted to build his own, inspired by blog posts suggesting significant savings. Daniel aimed to create a setup with two GPUs, expandable to four, to balance current needs and future scalability 1. He calculated that building his own workstation would be more cost-effective than relying on cloud services, estimating a break-even point over a summer of GPU usage 2. Additionally, the project rekindled his interest in computer building, a hobby he shared with his brother-in-law 3.

   

Challenges

Building the workstation came with its share of challenges and learning experiences for Daniel. One humorous issue was that his case couldn't close due to rigid power cables, which he rationalized as improving airflow 4. The scarcity of high-demand GPUs, like Nvidia's 30 series, added to the complexity, as Daniel and others struggled to secure these components 5. Despite these hurdles, he found confidence in using air cooling over liquid cooling, citing studies that suggested air coolers could outperform cheaper water coolers 6.

   

Hardware

Daniel's hardware choices were influenced by a mix of research and personal preferences. He selected a Ryzen Threadripper CPU, though he later considered switching to Intel based on his experiences 7. His build was guided by a blog post from Jeff Chen, which emphasized the importance of expandability and compatibility, particularly regarding PCIe lanes for GPUs 8. Daniel learned that each GPU requires a significant number of PCIe lanes, necessitating careful selection of CPU and motherboard combinations to ensure sufficient support for his setup 9.