Diffusion vs Autoregressive LLM Denoising Simulator

Celeris-1 Non-Autoregressive Engine
Speedup: 15.7x (Sub-200ms p50)

P50 LATENCY SPEEDUP RATIO

15.7x Faster

Parallel Discrete Diffusion vs Sequential Autoregressive

120ms (Diffusion) vs 1880ms (Autoregressive)

Autoregressive Generation (O(N) Sequential)
Status: Complete
TTFT
15.0 ms
Time-per-Token
29.1 ms
Total Latency
1880 ms
Discrete Diffusion Denoising (O(K) Parallel)
Status: Complete
Steps (K)
8 steps
Time per Step
15.0 ms
Total Latency
120 ms
Diffusion Step Inspector & Confidence Propagation
Step 8 / 8

Drag the scrubber below to inspect how parallel discrete diffusion progressively unmasks tokens and refines confidence across denoising iterations.

Step 8: All discrete token positions converged with full confidence. Final sequence output unlocked.
Scaling Latency vs Sequence Length
— Sequential O(N) — Diffusion O(K)
Enjoy this tool? Build your own with Super