Token Latency Sim

Token Latency Sim —

—

Feel the latency before you buy the box

Type anything. The reply is generated at the hardware speed you pick — streaming or all-at-once. Move the tokens/sec slider, then hit replay to compare the exact same answer.

default: RTX 3090 · 95 tok/s

Simulation settings

Hardware presets

Card = tokens/sec + $/tok/s. Tap to load. The slider stays yours — presets only set the starting point.

Setup	tok/s	600 tok	$/tok/s

Green $/tok/s = cheapest speed on the list. tok/s are typical published community ranges (±30%) for the model class shown — a 120B dense model will be far slower than an 8B one on identical hardware. Prices are Sep 2026 street; verify before buying.

First-token delay

TTFT700 ms

Prompt prefill + scheduling before the first token appears. Real GPU prefill is usually 200–800 ms; reading a big prompt or a cold context is worse.

Realism

Jitter Instantaneous rate wobbles ±50%, average stays on the slider

Cold start First message also loads weights from disk

Show live counters Running elapsed / tokens while it streams

Reset

How to read it

Faster than you can read (100+ tok/s) — feels like a hosted API.
Snappy (60–100) — no waiting perceptible on short answers.
Comfortable (30–60) — fine unless you're generating 2,000 tokens.
Noticeable (15–30) — you wait, but it reads as "thinking".
Slow (8–15) — you'll browse another tab.
Painful (<8) — unusable for chat, fine for batch.

The number that actually decides it is the 600-token column: a normal answer. If that's over ~20 s you'll feel it every single turn.