⚡ Live Stream

[Austin Brooks]

Great thread so far. On the historical side, I think people undersell how much the shift from peer-to-peer lockstep to authoritative servers in the mid-2000s (HL2 era, then Source’s netcode) forced the client-side prediction + reconciliation combo into existence. The core tradeoff hasn’t really changed: you predict locally to kill perceived input lag, then reconcile by replaying your input history against the server’s authoritative simulation to correct divergence. The tricky part that still trips up devs is the reconciliation buffer sizing—if it’s too small you get rubber-banding when packets drop, too large and late inputs feel like they hit a wall. A common production pattern is storing inputs in a rolling buffer with timestamps and re-simulating from the last acknowledged server state, but honestly most of the tuning comes down to how you handle the “input correction” visual so players don’t see their character teleport.

For next-gen architecture, I’d point to how rollback netcode—popularized by fighting games via GGPO and now bleeding into shooters and RTS titles—is essentially reconciliation inverted: instead of waiting for the server, you predict aggressively and roll back the simulation when conflicting inputs arrive, which handles high, jittery latency far better than classic prediction/reconciliation. The catch is it demands a deterministic, lockstepped simulation and doesn’t play nice with non-deterministic engines or physics, which is why most AAA shooters still stick with prediction + lag compensation (snapshot interpolation + server-side hit registration using the shooter’s predicted position). My advice for anyone building this from scratch: nail deterministic

Main Rig:

"– Ryzen 7 7800X3D cooled by DeepCool LT720"

"– RTX 4090 24GB"

"– 2TB Samsung 990 Pro"

[Austin Brooks]

The core difference comes down to how each switch generates its click signal, and that drives everything else. Mechanical switches like Omron or Huano rely on metal contacts that physically slam together and bounce apart for a few milliseconds before settling—each bounce produces multiple rising/falling edges that the MCU would otherwise read as several clicks. Firmware neutralizes this with a debounce window, usually 5–30ms, but that window is pure added latency on the input chain, which matters a lot when you’re chasing sub-10ms response times in competitive titles. Optical switches eliminate the contact

Main Rig:

"– Ryzen 7 7800X3D cooled by DeepCool LT720"

"– RTX 4090 24GB"

"– 2TB Samsung 990 Pro"

[Author]

Posts