My role
I proposed the project, designed the architecture, and led a six-person team across RTL, host trading software, testing, and hardware validation.
Inside the system
- A cut-through Ethernet path in SystemVerilog on LAN8720 PHYs over RMII — receive, transmit, CRC32 frame check, and IPv4/UDP checksums — with no vendor MAC core on the trading path.
- The trader decides on the last price byte and starts transmitting two clocks later, while the market-data frame is still coming in. An optional speculative mode starts the reply on the first byte and kills it with a bad frame check if the trade is off.
- An exchange that replays market data and timestamps every send and reply at 20 ns resolution.
- A two-trader version with a best-bid/ask matching engine speaking fixed-width FIX 4.4 across four asynchronous clock domains.
- Verified against Python reference models: generated test vectors drive the testbenches, and the C++ strategy matched the model byte-for-byte over 46,392 messages.
13.2 µsmedian round trip, 40 ns spread, 0 timeouts
23.8 µsreaction to a price move vs. ~2 ms for the laptop
15 / 15stale laptop quotes picked off in a market-making race
13.2 µs is the speculative-transmit build; without speculation it ran 20.9 µs. The C++ laptop trader’s p99 was 297 µs. In the race the FPGA quotes wider by design, so the laptop’s quote is the one left resting.