What is the lowest tick-to-order (or tick-to-trade) latency achievable without the use of FPGAs?

Without FPGAs, the lowest tick-to-order latency can be achieved by a server directly consuming multicast market-data, handling the feed internally, making a trade decision, and emitting an order using a binary protocol. Both the market-data receipt and the emission of the order involve a PCI bus transfer, and will incur a minimum of around one microsecond (1us) each.

The latency of the feed-handler, trade-decision logic, and order-entry packet construction are more difficult to quantify. However, we can estimate them as follows:

  • Depending on the market-data tick received and the complexity of the market microstructure, the work the feed-handler needs to do can vary. In the simplest case, an aggressive low-latency strategy is likely to want to respond to new liquidity being added at the top of the book. Such a tick is likely to be very simple for the feed-handler to process, and the processing time will be dominated by parsing the tick off the wire - of the order of 25ns for SBE (FIX Simple Binary Encoding).
  • Again assuming the simplest possible case, a trading decision could take a negligible amount of time and, to estimate the lowest possible tick-to-order latency, we take the decision latency to be zero.
  • Constructing an order-entry packet in a binary protocol such as OUCH will be even quicker than parsing the tick.
  • With kernel bypass, such as Solarflare OpenOnload or Mellanox VMA, the latency of the TCP packet will easily be under 100ns, assuming the TCP checksumming is offloaded to the NIC.

In summary, the minimum tick-to-order latency achievable on a standard server-based platform will be almost completely due to the PCI latency, namely around 2us.