After Virtual SC20: Liqid Talks Composable Infrastructure at HPC Scale and Recent DoD HPC Deployments

Print Friendly, PDF & Email

Letting the workload take the lead in drawing on the hardware, networking and software resources it needs, and doing this dynamically – that’s what composable infrastructure vendor Liqid is all about. In this virtual SC20 interview, we sat down with Liqid Vice President of Public Sector Sales Eric Oberhofer, and Vice President of Advanced Computing Matt Demas, to talk about the company’s HPC-class technical offering and recent installations for the Department of Defense’s High Performance Computing Modernization Program.

For the DoD, “those were very exciting wins and opportunities for Liqid as a company and show that composable infrastructure is not only ready for primetime, but we are ready for rack scale, multi-rack deployments…” Oberhofer said, “…being able to rapidly and dynamically reconfigure the hardware stacks based off of what the workload requirements are, fitting a hardware profile on the fly…”

He said the deployments are comprised of more than 200,000 Intel cores, 30 petabytes of usable all-NVMe composable storage in a parallel file system, more than 900 Nvidia A100 Tensor Core GPUs, all connected using Mellanox HDR 200 InfiniBand networking.

Compsability, he said, is a “shift from the kind of a static architectures of traditional HPC, where you’ve had different types of nodes that did not have the ability to change or accommodate different types of workloads.”

“There’s so many advantages to having this flexible hardware,” Demas said, “by allowing (customers) to let their software define what their hardware needs look like. It puts them in a scenario where they don’t have to go change their software, they’re not going to change their algorithms to meet what they bought three years ago. Instead, they say, ‘Hey, you know what, this works the best for me, this gives me the be the best result. So let me go ahead and re-outfit my hardware to go match exactly what I need.’ That’s from a from an accelerator perspective, that’s from a memory perspective, and from a networking and disc perspective, all of these things can be reprogrammed to give them exactly the resources they need, for the most effective result.”