Intel : Smart, Programmable Interconnects Unleash Compute Performance

[Link]

Hong Hou Corporate Vice President General Manager, Connectivity Group

By Hong Hou

Today, the amount of data moving through the data center and networks is growing at about 25% per year¹, and it shows no sign of slowing. The ability to move, store and process this data efficiently requires highly optimized, high-bandwidth, low-latency connectivity solutions.

I frequently talk with customers about the growing demands on their networks, and our discussions inevitably focus on three areas:

• Maintaining network performance as data movement increases.
• 'Delivering the agility to adjust to their customers' dynamic network requirements.
• Increasing visibility and control to manage the growing complexity of the network.

Addressing these questions often comes down to interconnect and the ability to efficiently move data through the network. Interconnect is one of Intel's Six Pillars of Technology Innovation and is a critical part of our strategy to deliver leadership products to customers across our portfolio.

More:Interconnect Technologies (Press Kit)

At Intel, we realize the critical role that interconnects play in delivering unbridled compute and storage performance at ever-increasing scale, and it is our strategy to deliver these innovations to our customers.

Over the years, Intel has talked in detail about interconnect technologies and their ability to address our customers' most critical network and data center infrastructure requirements.

Our interconnect vision focuses on four key areas:

• Increasing Network Programmability
• Optimizing Networks End-to-End
• Proliferating Silicon Photonics
• Deploying Intel® SmartNICs as Network Accelerators

Focusing on these areas allows us to help our customers accelerate network performance, improve network agility and increase network visibility, while achieving end-to-end compute performance at scale.

Baidu is one customer that recently deployed our interconnect technologies. The company leveraged Intel SmartNICs innovations based on Intel SoC Field Programmable Gate Arrays (FPGAs) and Intel Ethernet 800 Series adapter with Application Device Queues technology to improve its virtualization and workload performance, while accelerating data processing speeds.

Deploying Network Programmability

We embrace the vision that in order to increase network agility, we need to make networks more programmable. The technology we acquired from Barefoot Networks in 2019 and our portfolio of Intel Xeon Scalable processors, FPGAs, Foundational network interface cards (NICs) and SmartNICs play a critical role in our ability to enable network programmability. For example, Intel Tofino series of high-performance Ethernet switch ASICs and Intel FPGAs deliver full data-plane programmability with the open and community-owned P4 programming language.

Advantages of a programmable data-plane include greater diagnostic visibility, reduced feature complexity, and more portability and modularity. Intel Ethernet 800 Series provides a fully programmable pipeline through dynamics device personalization to support the growing number of network protocols for both network function virtualization and network edge use cases.

In the future, deploying greater levels of programmability within the network will enable customers to optimize the movement of data based on their requirements.

Optimizing Networks End-to-End

Unleashing computing performance requires end-to-end network optimizations Intel takes a holistic view of network infrastructure from the NIC to the switch and across the data center. Our vision centers around how we deliver additional value by leveraging and combining the capabilities in our portfolio. An example includes P4 programmability and In-band Network Telemetry on Ethernet switch silicon and bringing these technologies to the Ethernet controller. This, in combination with our Intel® Deep Insight Network Analytics Software, enables real-time visibility into the network's behavior.

Proliferating Silicon Photonics

High-speed optical interconnects provide the bandwidth needed by cloud service providers and large-scale data center operators. Intel pioneered silicon photonics technology, from the architecture to design to manufacturing, as a reliable way to move high-speed signaling from the die to the network. Use of silicon photonics engines has proven to keep board design complexity and costs down, while offering an interconnect that can be used for future growth in speeds and needs of programmable networking.

As bandwidth demands grow, silicon photonics, and optical interconnects will provide operators of hyperscale data centers with the best service level performance for cloud service uptime and latency. In 2020, Intel demonstrated the industry's first co-packaged optics Ethernet switch that integrated a 1.6 Tbps silicon photonics engine with Intel Tofino 2, a 12.8 Tbps P4-programmable Ethernet switch. In future networks, co-packaged optics like this will offer power and density advantages for switches at 25 Tbps and higher, and ultimately will be necessary for bandwidth scalability.

Deploying Intel SmartNICs as Network Accelerators

Ethernet-based NICs are broadly used across the network today to provide endpoint connectivity to servers, CPUs and other processing elements in the network. Intel Ethernet-based NICs provide a comprehensive level of interoperability across a range of media types and Ethernet speeds. They are highly optimized to unlock network performance while providing flexible and scalable I/O to meet network infrastructure needs.

As the network has evolved over time, the infrastructure needs of customers have also evolved as they seek the flexibility through technologies like SmartNICs to accelerate their networking functions for existing and emerging use cases.

Today, Foundational NICs from Intel, such as the Intel Ethernet 700 and 800 Series, are delivering proven, reliable high-performance connectivity for the majority of server and storage

deployments. SmartNICs are becoming critical in managing and avoiding congestion in the network.

Intel has significant investments in SmartNICs and a long-term roadmap of new products to continue its leadership. We collaborate on focused solutions with key cloud and communications customers to address their unique needs. In addition to our many custom engagements, we are delivering the Intel FPGA PAC N3000 SmartNICs for communications service providers with several systems partners, including Nokia, Altiostar, HCL, F5 and Affirmed.

We see SmartNICs playing an even greater role in data centers as a network accelerator and we will collaborate with key cloud service providers and communications service providers to address their unique needs.

Building the Networks of the Future

Delivering true end-to-end network performance requires a programmable interconnect platform built with higher levels of intelligence and open software frameworks. Continuous network innovation is required to unbridle compute and storage at ever-increasing scale.

In the future we see a network that is a programmable platform, where SmartNICs and switches become more critical, silicon photonics technology is integrated and ubiquitous, there is end-to-end optimizations across Ethernet-based controllers, and adapters and switches will shape the traffic and deliver performance at scale.

Hong Hou is corporate vice president and general manager of the Connectivity Group at Intel Corporation.

¹ Source: Cisco Global Cloud Index, 2016-2021