1 min read
White Paper: GPU-Initiated, Liquid-Cooled, Ultra-High-Density Storage for Next-Gen AI
This paper introduces a paradigm shift in storage architecture designed to overcome the CPU-centric data path bottlenecks in modern AI workloads. By...
Press
Updated on May 28, 2024
Air cooling is the traditional way to remove heat generated from processing units by directing cool air through hot surfaces to dissipate heat.
In recent years, the rapid expansion of AI and HPC is in demand of higher performance in both the processing units and in heat dissipation. Aimed to address the heat dissipation limit of air cooling, liquid cooling is introduced as an advanced cooling method to have coldplate array dissipate heat directly from the processing units.
With significant research into AI training system and other data center applications, we came up with multiple swappable cooling solutions for cooling hot processing units (OCP accelerator module) within the same AI training system. This swappable cooling solution design can provide quick and easy upgrade of cooling in most scenarios.
Register to Download the whitepaper!
1 min read
This paper introduces a paradigm shift in storage architecture designed to overcome the CPU-centric data path bottlenecks in modern AI workloads. By...
1 min read
This paper explores a disaggregated key-value (KV) storage architecture designed to efficiently offload KV cache tensors for generative AI workloads.
1 min read
This paper explores an advanced framework designed to automate the extraction of important attributes from unstructured part datasheets. By...