# TTFT (time to first token)

> VI: https://5ac.vn/wiki/ttft/ · EN: https://5ac.vn/en/wiki/ttft/ · lang: en · spec_as_of 2026-10-22 (spec no formal spec) · cluster C17

The wait from submitting a request to the first generated token — what decides whether chat feels responsive. 5ac's GB10 benchmark: 8.3 seconds on a warmed 16K prompt.

---
## Definition

TTFT (Time To First Token) is the delay between submitting a request to an inference engine and receiving its first output token. For a chat user it is one half of a single experience: TTFT decides how long you watch a blinking cursor, while the generation rate in tokens per second decides how fast the answer unrolls. An assistant with a high tokens-per-second rating but a long TTFT still feels sluggish next to a slower generator that starts almost immediately.

## Mechanism

The pre-token stretch is a sum of waits: queueing, processing the whole prompt, then computing the first token. Splitting prompt processing (prefill) and text generation (decode) into two phases — including running them on separate pools — is the subject of the [Prefill/decode disaggregation](/en/wiki/prefill-decode-disaggregation/) entry, and the place of TTFT in the metric set a serving stack must report belongs to [Inference engine](/en/wiki/inference-engine/). This entry holds one definition only: TTFT ends at the first token, not at the end of the answer.

5ac's measured figures come from the DSpark vs MTP benchmark on DGX Spark, run with the repeated "n=5" method on one set of NVFP4 weights: "a TTFT of about 8.3 seconds for a warmed 16K prompt", reported beside decode speeds of "17.2–20.5 tok/s (thinking) and 21.6–22.7 tok/s (non-thinking)". The word "warmed" matters: the 8.3-second figure is not a cluster measured from a cold start, and the article publishes no cold number. Repetition discipline like n=5 is a benchmarking-method question covered under [Benchmark suite](/en/wiki/benchmark-suite/).

## Application context

TTFT matters most where the first question drags context behind it: agents that paste tool results into the prompt, bots that answer after reading documents. As prompt length grows, the wait before the first token usually grows with it — which is why comparisons between stacks are only honest when TTFT and tokens-per-second are reported as a pair, measured the same way. The GB10 benchmark's two numbers make that concrete: an 8.3-second wait, then text streaming at 17–23 tok/s depending on thinking mode.

## Sources fetched

- [DSpark vs MTP on DGX Spark — practical benchmark](/en/blog/dspark-vs-mtp-benchmark-gb10/) — TTFT of about 8.3 seconds for a warmed 16K prompt, decode 17.2–20.5 tok/s (thinking) and 21.6–22.7 tok/s (non-thinking), n=5 runs on NVFP4 weights — accessed 2026-10-22

## Related

- [Inference engine](/en/wiki/inference-engine/)
- [Benchmark suite](/en/wiki/benchmark-suite/)
- [Prefill/decode disaggregation](/en/wiki/prefill-decode-disaggregation/)