Sugon 60K Cluster Fuses Supercomputing and AI

🔑 Enhanced Key Takeaways

• •The cluster utilizes Sugon's proprietary 'ParaStor' distributed storage system, specifically optimized to handle the high-concurrency I/O demands of large-scale model training.
• •The architecture implements a high-speed interconnect fabric based on Sugon's self-developed 'DCU' (Deep Computing Unit) technology, aiming to mitigate bottlenecks caused by international export restrictions on high-end GPUs.
• •The project is part of the 'East Data, West Computing' national strategy, with the cluster physically located in a specialized data center hub in Western China to leverage lower energy costs and cooling efficiency.

🛠️ Technical Deep Dive

• •Cluster Scale: 60,000 units of high-performance DCUs integrated into a single unified resource pool.
• •Interconnect: Utilizes a multi-level fat-tree topology to ensure low-latency communication between compute nodes.
• •Software Stack: Integrated with Sugon's 'SugonAI' software platform, supporting mainstream frameworks like PyTorch and MindSpore via custom drivers.
• •Cooling: Employs liquid-to-chip cooling technology to maintain thermal stability for high-density GPU racks, achieving a PUE (Power Usage Effectiveness) below 1.2.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline