Continuous Collision Detection (CCD) vs Discrete Collision at Hyper Velocity: Fundamental Mechanics and Deep Mathematical Modeling

[Gabriel Young]

I’ve been benchmarking collision detection overhead on fast-moving projectiles in my sim builds, and the tunneling effect keeps wrecking my discrete timestep results. When a projectile exceeds roughly one object diameter per frame, my swept-sphere queries are clearly losing accuracy, but I’m not fully convinced the math holds up. I’ve been reading about conservative advancement and continuous SAT, but I want to hear from people who’ve actually implemented swept-volume calculations in a real engine versus just theoretical models.

For those working with hyper-velocity objects, do you rely on a hybrid approach where CCD only activates above a critical speed threshold, or do you run continuous detection globally? How are you handling the intersection time computation—root-finding on the SAT distance functions, or are you approximating with swept GJK? And practically, what’s your tolerance for false positives when the swept volume over-approximates the geometry? I’m particularly interested in how you balance the O(n) cost of continuous broadphase against the accuracy gains. Anyone got solid references on the underlying mathematical proofs, or is this mostly been trial and error from your own physics code?

🎮 Battlestation & Rig: Ryzen 5 7600X | RTX 4070 Super Dual | 32GB DDR5 6000 CL30 | FormD T1 SFF | ASUS ROG Swift OLED PG27AQDM 27″ 240Hz 0.03ms | Logitech G Cloud + Razer DeathAdder V3 Pro + Endgame Gear EM-C Pad | Beyerdynamic DT 770 Pro (80 Ohm) + Focusrite Scarlett Solo Gen 4

📍 Melbourne (Australia)

"1% low FPS flat as a runway before queuing high-elo ranked."

[Isaac Ward]

The core problem you’re wrestling with is that discrete collision detection samples the world only at frame boundaries, so any object whose swept volume during a single timestep exceeds the thickness of its target collider simply tunnels through — the classic “bullet through paper” failure.

🎮 Battlestation & Rig: i5-13600K | RTX 4070 Super | 32GB DDR5 6000 CL32 | DeepCool LT720 | Alienware AW2725DF 27″ 360Hz QD-OLED | SteelSeries Apex Pro TKL Wireless + Zowie EC2-CW + Zowie G-SR-SE Bi | Audio-Technica ATH-R70x + Schiit Magni/Modi Stack

📍 Chicago, IL (USA)

"1% low FPS flat as a runway before queuing high-elo ranked."

[Cole Brooks]

Great topic. The core issue with discrete detection at hyper velocity is simply the sampling gap—your engine evaluates overlap only at t and t+Δt, so any object moving faster than its own thickness per frame (think a fast bullet or a high-velocity projectile) slips entirely through a thin collider

"────────────────────────────────────────"

"Stay humble in victory, gracious in defeat."

[James Harris]

Good topic, and since we’re talking hyper velocity here, the core problem everyone underestimates is that discrete detection is fundamentally a sampling problem. In DCD you’re only checking overlap at frame boundaries, so once an object’s sub-frame displacement exceeds the thickness of the collision mesh it’s traveling toward, it simply vanishes through it—classic tunneling. The math behind it is brutal: for a projectile moving at, say, 2000 m/s you’re talking sub-millimeter penetration on a 1mm wall within a single 16ms frame. CCD fixes this by sweeping the shape through time

"i5-13600K | RTX 4070 Super | 32GB B-Die 3800 CL14 tight timings"

[Author]

Posts