Introduction

There has been a lot of buzz recently in the industry regarding SmartNICS. But before we go any further, we need to understand the genesis of smartnics. With Moor’s law slowing down and Computing moving to Data Centers. There is need for more packet processing closer to the Packet IO and relive the CPU for more arithmetic workloads.

SmartNICs and NPU’s of the old era are remarkably similar in Outcome they want to achieve…  You might have heard of Network Processors (NPU). The earlier Designed Cisco and Juniper Routers have plenty of them. Then what changed now…?

• Cloud computing happened
• iOS & android based Smartphones have come.
• 4 billion users have “Always ON” Internet Connectivity on their Smartphones.
• Data Consumption have exploded 100times since the last decade.
• Low Latency Applications with massive scale have come.
• Autonomous Car are fiddling around the horizon.
• 5G & Edge computing is at horizon of 5 yrs.

How SmartNIC is different from Normal NICs

Traditional NICs cards rely on Interrupt based mechanism for pkt IO. The approach has not changed since last 40yrs into industry. This approach takes a heavy toll on the CPU performance especially when there is high packet throughput. We are talking about 20Gbps and higher (up to 200gbps).

SmartNICS do 2 things differently

1. They significantly reduce the CPU-interrupt.
2. A large amount of pkt processing is Done in the NIC card itself, Be it Vlan/VXLAN, tcp connection maintenance, packet crypto processing to name a few.

SmartNIC market size

Smart NIC market is projected to become a $600 M market by 2024, or 23% of the total Ethernet adapter market. Vendors have developed or are developing innovative solutions to gain entry in the expanding Cloud data center market, and the emerging telco edge market.(src)

 SmartNIC Hardware Categories

SmartNICs hardware have broadly 4 categories

ASIC

These smartnics have excellent price performance, Though development cost is high but cheap to manufacture. Programmability is limited. All those Nics below 40G typically fall in this category. An example would be i40e Intel Nic card.

There are high ends NICs too… Consider Netronome Agilio CX 2x25GbE

ASICs are the staple of Nics. 90% of all NIC cards fall in this category.

Power Efficiency and low heat dissipation are the great advantage for ASIC designed NICs.

With advent of WiFi connectivity for workstation and personal use, The wire based connectivity is not limited to Server rooms, Private/Public clouds and Data Centers.

FPGA

FPGAs are unique in their ability due to programmability.

A FPGA based NIC once shipped out can stilled be programmed, consider this like circuity update every year or so, This is a heck of a capability as knocks off the ASIC limitation.

Increases the shelf life of the NIC devices… But FPGA is pricier as compared to ASIC. The cost absorption has to be done at both the ends… (NIC Vendor & the Data Center/Telecom Provider)

Since packet Processing adheres to relatively fixed standards & specifications. New networking features can be added, and firmware can be upgraded to support new features. This enhances the Shelf life of NIC-cards.

• IP cores developed by 3rdparty vendors can be Programmed to FPGA-SmartNIC card and you have relative less vendor lock-in. You can buy various fpga-ip blocks few listed here…
• Higher-self life
• Vendor supported firmware upgrades.

give an example Xilinx Alveo U25, Alpha Data 9v3

FPGA + Processor

With growing variety of use cases and increasing feature support in Openstack and SDN world.

1. There is requirement for programming the ovs rules directly into the NIC cards via SDN Controllers. This requires a support for bare backbone Linux into the system, hence, NIC cards support processors.

• Incase of complex VNF Orchestration, a SmartNICs can be instructed to provide enhanced

IO virtualization and Network slicing. Typically, cases for 5G deployments.

• Bare Metal Provisioning via Openstack ironic would require SmartNICs to be configure directly by Neutron components via OpenFlow messages as well as to support pixie boot mechanism, As the Server rack is completely without OS, all onus likes on SmartNICs, with Processors like arm/Zynq along with Linux.

• Some SmartNICs vendor like to provide telnet/ssh interface within SmartNICs itself to execute commands, retrieve pkt stats, and other ipmi related capabilities. Complex capabilities like bonding, doing BAR memory realignment, requires running some OS interface.

Processor + ASIC

This band of NIC cards have All the Ethernet based capabilities built in Asic and the Acceleration capability is powered by Processor. They give you the best of both the worlds… And they kind of form the sweet spot.

1. Allow you to run this Hardware as traditional & Capable ASIC based high performance ethernet card, with kernel and user-space mode (read DPDK) with advanced features like SRIOV.
2. Allows you to accelerate packet processing using pkt accelerator algorithms programmed in processors. Example ConnectX®-5 InfiniBand and Ethernet network controller ASIC with a state-of-the-art FPGA.
3. Bare Metal Provisioning via Openstack ironic would require SmartNICs to be configure directly by Neutron components via OpenFlow messages as well as to support pixie boot mechanism, As the Server rack is completely without OS, all onus likes on SmartNICs, with Processors like arm/Zynq along with Linux.
4. OpenFlow Offloading into the HW itself.
5. Complete IPSEC offloading including IPsec SA negotiation, In general only ipsec Phase SA policies and, key are offloaded. But All IPSEC phase-1 and Phase-2 neg

SmartNICs Evolution

• Asic based Std NIC cards with kernel drivers
• Upgraded NICs with DPDK Support.
  • Added support for SRIOV,
• Crypto Offload & Inline Processing
• SDN Compatible, Service chaining
• OVS flow offload & Rte Flow offload

 SmartNIC features

SmartNIC Software architecture