Frame Budget Pipeline Timing (2D)

CPU and GPU lanes against the v-sync deadline. Model constants are illustrative, not benchmarks.

Switch to 3D

Presets

Scene load

Modelled machine

Independent numerical checks

The science

A real-time renderer must finish every frame inside the display interval T = 1000 / refresh: 16.67 ms at 60 Hz, 13.89 ms at 72, 11.11 ms at 90, 8.33 ms at 120. CPU and GPU work in a pipeline, so the frame cost is roughly max(t_cpu, t_gpu). The CPU side grows with the number of draw calls; the GPU side grows with triangles processed and pixels shaded. Instancing draws many copies of one mesh in one call, level of detail swaps in a cheaper mesh, shadows add a second pass over the casters, and the render scale changes the pixel count by its square.

t_cpu = base + calls x c_call, t_gpu = base + tris/rate_tri + pixels x overdraw/rate_px, v-sync (double buffered): shown = T x ceil(t/T), latency = t_sensor + (depth + 1) x shown.

The 2D view

The timing diagram draws consecutive frames: a frame starts when the previous one is swapped, its CPU and GPU work run in parallel lanes (pipelined), and the swap waits for the next v-sync tick. A frame longer than T misses the deadline and is shown at a later tick, so the displayed rate falls in steps (60, 30, 20 fps at 60 Hz). The map shows the displayed rate for every crate count and render scale: its plateaus are exactly Hz/k. The latency waterfall splits motion-to-photon time into tracker delay, waiting for the next pose sample, pipeline frames and scan-out; around 20 ms is a commonly quoted comfort target for head-mounted displays.

Honest limits

The model constants (6 to 8 microseconds per call, GPU classes of 1, 3 and 8 billion triangles per second, 1.6 overdraw) are illustrative round numbers, not benchmarks of any product; real drivers, tilers, triple buffering, variable refresh and asynchronous reprojection all behave differently. This 2D page has no live renderer, so it has no measured column (the 3D page reads renderer.info). Use it to learn the arithmetic and the relationships, not to predict a specific headset or game.

Frame cost against crate count
V-sync staircase: frame cost to shown rate
Motion-to-photon latency waterfall
Made with 🟠 Blender