Offloading TCP into hardware
Many applications in industry can benefit from Ethernet and TCP/IP as it is a well-known and supported networking standard. More and more, these industrial applications require higher bandwidth and lower latency. This means that it is becoming a challenge not to overload the CPU with a TCP/IP stack running at maximum bandwidth. These increasing requirements make the processor spend more time handling data rather than running your application.
Easics' TCP Offload Engine (TOE) can be used to offload the TCP/IP stack from the CPU and handle it in FPGA or ASIC hardware. This core is an all-hardware configurable IP block. It acts as a TCP server for sending and receiving of TCP/IP data. Because everything is handled in hardware very high throughput and low latency are possible. The IP block is completely self-sufficient and can be used as a black box module which takes care of all networking tasks. This means that the rest of the system is free to use its processing power purely for application logic. In some use cases, integrating our full-hardware TCP/IP stack eliminates the need for any built-in embedded processor at all.
The easics TCP Offload Engine is available as a 1 Gbit/s or 10 Gbit/s version. Both versions support Ethernet packets, IP packets, ICMP packets for ping, TCP packets and ARP packets. The 10 Gbit/s version additionally supports pause frames.
IP Core Architecture
The figure below shows the core’s building blocks and its four most important interfaces. The first of these is an industry-standard (X)GMII interface which communicates with a 1(0) Gbit PHY. The second is situated on the application side: two FIFOs with a simple push/pop interface, one for RX and one for TX. These FIFO interfaces, as well as an internal TCP block, communicate with a memory system which is to be provided outside of the core (the third interface). The size and type of memory can be selected by the user. ARM’s AMBA AXI4 2.0E is the protocol used for this communication. Various FPGA vendors, such as e.g. Xilinx, Intel, provide building blocks to interface internal block RAM, SRAM, or DRAM with an AXI bus. The fourth and final interface is used to configure various networking parameters and to read status info.
Performance
Following data throughput numbers have been measured on Xilinx ZC706:
| MTU | TX | CPU (%) | RX | CPU (%) |
|---|---|---|---|---|
| 1500 | 905 | 0 | 949 | 0 |
| MTU | TX | CPU (%) | RX | CPU (%) |
|---|---|---|---|---|
| 1500 | 9.18 | 0 | 9.35 | 0 |
| 9000 | 9.69 | 0 | 9.73 | 0 |
The data throughput is thus higher for MTU=9000 (jumbo frames). CPU load is 0% since the full TCP/IP connectivity is in the FPGA (full hardware acceleration).
Following latency numbers have been measured for the 10G TOE in simulation, making use of the Xilinx transceiver models (responsible for 160 ns latency):
- TX latency = 656 ns
- RX latency = 640 ns
- Round Trip Time (RTT) = 1.3 µs
Resource Usage & Scalability
- 1G (1 TCP socket) ~ 12K LUTs
- 10G (1 TCP socket) ~ 34K Zynq LUTs / 24K Arria 10 ALMs
- 10G (4 TCP sockets) ~ 39K Zynq LUTs
Our design is such that for each additional TCP socket the number of additional LUTs / ALMs is negligible, only extra memory must be foreseen.
Application Areas
Some of the main application areas are listed below. Don't hesitate to contact us for feasibility advice of your use case.
- networking in an industrial environment, for real-time automation and process control ('Industrial Ethernet' e.g. EtherCAT, PROFINET, POWERLINK, Modbus TCP, ...)
- machine vision / multi-camera monitoring (e.g. GigE Vision)
- test & measurement connectivity
- networked storage, such as iSCSI
- electronic trading, using Financial Information eXchange (FIX) protocol
- ...
Documentation
Refer to the easics' TCP Offload Engine product brief for an extensive list of all features.
Xilinx Zynq-7045 Evaluation System
Targeted to Xilinx Zynq-7045 on ZC706
Dual ARM Cortex-A9 core processors are not used for handling TCP/IP. They are free to be used for your application.
SFP+ for 10GigE
Supports Vivado design flow
10 GigE Resource count (64-bit @ 156.25 MHz)
- Vivado 2017.1 targeting XC7Z045
- Full Stack (1 TCP socket) : 34K LUTs
- Multisocket Full Stack (4 TCP sockets) : 39K LUTs
Runs on Xilinx ZC706 Development Kit
Intel Arria 10 Evaluation System
Targeted to Intel Arria 10 SoC SoM by ReflexCES plugged in on ReflexCES' PCIe Carrier Board Arria 10 SoC SoM
SFP+ for 10GigE
Supports Quartus design flow
10 GigE resource count (64-bit @ 156.25 MHz)
- Quartus Prime Pro 17.0 targeting 10AS027H3F34E2SG
- Full Stack (1 TCP socket) : 24K ALM
- Multisocket Full Stack (4 TCP sockets) : 27K ALM
Runs on ReflexCES PCIe Carrier Board Arria 10 SoC SoM Development Kit
Xilinx Virtex-6 FPGA ML605 Evaluation Kit
Targeted to Xilinx Virtex-6 on ML605
RJ-45 for 1GigE
Supports ISE design flow
1GigE resource count (32-bit @ 31.25 MHz)
- Synplify H-2013.03 targeting XC6VLX240T
- Full Stack (1 TCP engine) : 12K LUT
Runs on Xilinx Virtex-6 FPGA ML605 Evaluation Kit
Easics' Artix-7 FPGA Demo Board
Targeted to Xilinx Artix-7 on the easics demo board
RJ-45 for 1GigE
Supports Vivado design flow
1GigE resource count (32-bit @ 31.25 MHz)
- Vivado 2015.1 targeting XC7A100TFTG256-2
- Full Stack (1 TCP socket) : 5K LUT
Runs on the easics Artix-7 FPGA demo board