The purpose of Intel’s Infrastructure Processing Unit (IPU) is to optimize data center performance by offloading infrastructure management tasks from server processors. The Intel IPU also provides enhanced security through isolation of tenant applications from provider services and enablement of virtual storage for increased datacenter flexibility.
The IPU itself connects to the server via a PCIe slot. This enables high-speed communication between the IPU and other components including the CPU, RAM and storage devices.
Once connected and configured, the IPU can offload packet routing and switching, load balancing, storage virtualization, and data encryption and compression. This saves resources that the CPU, GPU and other components would otherwise have to commit. The overall effect is increased efficiency and faster outcomes.
There are four key components of the Intel IPU:
1. The high-performance Network Interface supports advanced features like Remote Direct Memory Access (RDMA) and high-bandwidth Ethernet.
2. Programmable cores like Intel Xeon processors allow developers to tailor the IPU’s behavior for specific workloads.
3. Specialized hardware accelerators take care of various tasks, including encryption, compression, and packet processing.
4. Storage Integration helps manage NVMe storage, which enables disaggregation of compute and storage resources to increase efficiency and scalability.
Working as part of a comprehensive datacenter or edge solution, the Intel IPU can act as a hypervisor bypass, enabling IT managers to offload virtual machine (VM) networking and storage operations. And because the IPU employs a field programable gate array (FPGA), it allows cloud providers to create software-defined infrastructure specific to their unique requirements.
Intel’s datacenter-centric IPU is available as an add-on to various models in the Dell Technologies PowerEdge series of servers. The PowerEdge R760xa Rack Server, for example, makes use of the IPU to better manage AI-ML/DL training and inferencing, advanced analytics and VDI workloads.