Dell "Unconditional high performance? 'TPO' is the answer for AI storage as well"

TPO (Time, Place, Occasion), which means "appropriate to the time, place, and situation," is becoming important in selecting artificial intelligence (AI) storage. As the selection of storage infrastructure that can efficiently process the rapidly increasing amount of unstructured data, such as images and videos required for AI learning, is emerging as a key element of AI projects, Dell's customized storage strategy for each AI workload is attracting attention.



Jang Se-jae, managing director of Dell Technologies UDS business unit. [Photo = Reporter Yoon So-jin]



"AI storage, cost reduction and management efficiency are key"



Jang Se-jae, managing director of Dell Technologies UDS (Unstructured Data Solutions) business unit, said in a recent interview with iNews24, "The most efficient storage method varies depending on the AI ​​workload," and "There is no need to insist on the highest performance unconditionally." He explained that, unlike the development of large-scale AI models, storage that is easy to manage and has excellent scalability is more efficient for the AI ​​workloads performed by most companies.



"Only 5% of AI workloads involve the development of large-scale training models," said Jang. "Most companies use already trained models to fine-tune or perform inference tasks, which do not necessarily require high-performance storage such as parallel file systems (PFS)." This means that



just as you don't need an F1 racing car to go to the convenience store in front of your house, purpose-built storage is efficient for AI development. "Rather than PFS, which is complex to manage and expensive, scale-out NAS can be a more realistic option," he explained. He



then suggested Dell's scale-out NAS (Network Attached Storage) solution, PowerScale, as the answer. PowerScale is a storage platform that can expand large-scale data into a single volume, and is the first in the industry to receive Ethernet-based NVIDIA DGX SuperPOD certification. It features low-cost deployment by using general-purpose Ethernet instead of expensive InfiniBand network equipment, while providing sufficient performance.



Director Jang said, "PowerScale supports GPUDirect Storage technology that directly communicates with NVIDIA's graphics processing units (GPUs) to increase data processing speed," and "It showed performance equivalent to PFS in NVIDIA's AI performance test, MLPerf." He added, "It shows strengths in data feeding (supplying training data to GPU) and checkpoint (mid-training backup) performance, which are important in AI workloads."



"Maximizing AI utilization by integrating structured and unstructured data"



Last year, Dell also launched 'Dell Data Lakehouse', which integrates and manages structured and unstructured data. This platform combines the advantages of existing data warehouses and data lakes, and shows the characteristic of enabling analysis without data movement.



Director Jang explained, "In the AI ​​development process, both structured and unstructured data must be utilized, but in the past, they had to be managed with different systems. Data Lakehouse integrates and manages both of these, while allowing immediate analysis and utilization in the necessary form."



It saves time and cost by being able to store and utilize data coming in through various protocols such as S3 (cloud-based object storage), NFS (network file system), and SMB (server message block) without conversion. Dell developed this platform in collaboration with data analysis specialist Starburst, and Starburst's data acceleration technology, 'Warp Speed', is an exclusive function provided only to Dell Data Lakehouse.



"It is now important to go beyond simply introducing GPUs and build an infrastructure that can efficiently operate the entire AI workload," said Managing Director Jang. "Dell's strength is providing a portfolio for the entire AI workload process from data management to AI model development and inference."





https://www.inews24.com/view/blogger/1811952