I guess it is no secret, the reason why I’ve put so much effort into getting the Goldilocks 1284p board built. I was looking for a platform that would allow me to experiment with the uIP TCP/IP and UDP/IP stack with the most performance and flexibility possible while still being compatible with the huge range of sensors and actuator Shields that form the Arduino legacy. From the microprocessor view, the ATmega1284p used in the Goldilocks certainly achieves that goal.
I’ve written in a previous post about the theoretical performance difference between the common Wiznet (or IINChip) W5100 used in almost all Arduino Ethernet shields and the component I have selected that uses the W5200 to provide the Ethernet interface. This post demonstrates the real world performance differential with a simple example.
But first, I am happy with the result of the uIP port to the Wiznet platform within freeRTOS. I’ve taken some of the old uIP v0.9 and v1.0 files from many sources, and updated them with the latest snapshot status from Contiki 2.7, to try to bring the last 5 years of experience into the result. Whilst the resulting codebase has not as yet been extensively tested, it seems to work as expected.
This is a simple test, sending 1300 byte PING packets to the MACRAW interface on the IINChip to be handled by uIP. After 100 PINGs the W5200 takes on average 3.804 ms, whilst the W5200 takes on average 22.109 ms for each round trip.
This means the W5200 is nearly 6x faster than the W5100 in real world performance.
Of note, this real world result is achieved whilst over-clocking the W5100 SPI bus out of specification at 5.5MHz (being SCK/4), rather than at 4MHz which is the specification. The W5200 SPI bus can, of course, run up to 30MHz or faster, so its limits are not even being tested by the Goldilocks ATmega1284p MCU.
W5200 SPI bus
The key differential which provides the W5200 its performance advantage is the use of multi-byte burst transfer mode for moving payload data into and out-of its controlling MCU. In theory the entire 32 kByte Address space of the W5200 could be transferred in one transaction. In practice, a full Ethernet frame can be transferred in just over 1 ms.
These shots show how the W5200 SPI multi-byte transfer works in practice.
This screenshot shows an entire received 1300 Byte payload PING frame being transferred in 1.34ms.
The Goldilocks AVR1284p takes 0.29ms to generate the response PING, and then it is transferred back to the W5200 for transmitting on the wire.
This screenshot shows the detail of the transmission of the PING frame to the AVRmega1284p. Note that each Byte takes less than 1 us to transfer.
W5100 SPI bus
The W5100 SPI bus uses a 4 byte transaction to transfer a single payload byte, and it is not capable of a multi-byte burst mode.
This screenshot shows the detail of the transmission between the AVRmega1284p and the W5100. It shows that to transfer 1 payload byte it takes about 0.036 ms (which is 36 us or 36x longer than the equivalent transfer on the W5200).
If you’re planning on building anything that relies on wired Ethernet, then go out of your way to find a Elecrow W5200 Shield or Seeed W5200 Shield. It is about six times faster than the common W5100 in the real world testing, and has many other great features.
Code, as usual, on Sourceforge.