Modern scientific research and big data questions are urgent and require high-performance computing (HPC) system horsepower for answers. While recent advances in microprocessor, GPU, and ultra-low latency networking technologies continue to increase HPC performance, it's largely monolithic and designed for specific workloads, making it difficult to repurpose as new projects materialize.
The increased ubiquity of AI compounds this problem. Unpredictable Nonlinear AI and machine learning workloads differ vastly in their resource requirements from one point in the process to the next. Data ingest, for example, is storage-intensive, whereas inference relies heavily on GPU to make calculations. As the amount of data increases exponentially daily, so does this problem. Further, as workload requirements can change quickly, it’s difficult to predict hardware needs. Accelerators such as GPU and FPGA are difficult to add to systems as required, and traditional architectures do not lend themselves well to sharing these valuable resources, often leaving them overtaxed or underutilized.
Software-defined platforms are increasingly utilized to work around the inherent limits of physical hardware, and composable disaggregated infrastructure solutions are emerging as the platform of choice for HPC operations that rely on AI to solve big problems.
Composable infrastructure from Liqid enable IT users to compose precise HPC configurations from pools of compute, GPU, FPGA, NVMe and storage-class memory over multiple fabrics in seconds, and granularly scale as needed. Upon project completion, Liqid allows you to move resources back into pools, ready to be redeployed for future workloads.
To learn more, view this webinar, “Composable Supercomputing for HPC Workloads,” and find out how some of the most prestigious research laboratories in the world deliver smarter HPC infrastructure with composable infrastructure from Liqid. These critical functionalities help organizations accelerate time-to-value, improve IT agility and increase resource utilization of their HPC environments.