Composable infrastructure is the industry term for virtualizing pools of physical resources into logical servers that are optimized for the applications they run. The goal is higher performance at lower cost, by tuning server configurations to application requirements and reducing the idle time of costly physical resources.
For example, if you are training a machine learning model, you may need multiple neural processors or GPUs, plus a large store for the training corpus. If you're running a lot of containerized apps, you may need a few hundred gigabytes of memory and plenty of CPU cycles.
With traditional servers, what's in the box is what you get. With composable servers, you choose the physical resources your application requires, knit them together with software, run the jobs, and when done, return the resources to the pool.
Think of it as cloud computing for the rest of us. Or Intel's rack scale integration (RSI) concept made real.
Liqid offers a low-latency, PCIe fabric to connect the pools of disaggregated GPUs, FPGAs, CPUs, NVMe storage and Intel Optane non-volatile memory. IT users can orchestrate balanced systems for each AI phase of data ingest, cleaning/tagging, training, and inference, while minimizing the data center footprint.
"Many organizations are looking for ways to integrate AI and machine learning into their IT infrastructure while avoiding the hardware sprawl and inefficiencies that often come with it," said Ron Pugh, Vice President, Dell Technologies OEM & IoT Solutions. "Liqid now can provide its composable solutions for AI, based on trusted Dell EMC PowerEdge infrastructure and support, for IT users seeking to improve utilization and efficiency for data-intensive applications."
With the advent of warehouse-scale computers, it soon became obvious that traditional 2 to 6u building blocks needed to scale up as well. Thus Intel's RSI initiative.
But, as always, the interconnects were both costly and slow - until now. With PCIe v4 and, soon, v5, and NVMe, we finally have the high-bandwidth and low-latency links needed to realize Intel's vision. But once you are installing racks of CPUs, etc., how do you turn them into usable systems?
HPE's Synergy product line is a single-vendor answer to that question, while Liqid is a vendor-independent option. Dell's blessing will give more CIOs the confidence to explore the benefits of composable infrastructure, and bring important applications back in-house from the cloud.