Beginner5 min read33 of 52

Latency

Time-to-first-token and total generation time — the key performance metrics for production LLM systems.

Read first:Inference

SPEC: Latency

Definition

[Definition] Latency is the time elapsed between submitting a prompt to an LLM and receiving the output. It encompasses network transmission, server-side queuing, model computation (prefill + decode), and response delivery. Latency is a critical quality-of-service metric for LLM applications.

Latency Components

Total Latency = Network (client→server) + Queue wait time + Prefill time (process input tokens) + Decode time (generate output tokens × N) + Network (server→client)

Key Latency Metrics

Time to First Token (TTFT)

  • Time from sending request to receiving the first output token
  • Dominated by: network RTT + queue + prefill computation
  • Critical for perceived responsiveness in chat UIs
  • Streaming enables users to see output while generation continues

Time per Output Token (TPOT) / Tokens per Second (TPS)

  • How fast the model generates each subsequent token
  • TPOT = milliseconds per token; TPS = tokens per second (inverse)
  • Dominated by: memory bandwidth (reading KV cache + weights)
  • Typical: 20–150 TPS for hosted APIs

End-to-End Latency

  • Total time from request to complete response
  • = TTFT + (output_tokens × TPOT)
  • Relevant for batch jobs and non-streaming use cases

Time Between Tokens (TBT)

  • Consistency of generation speed across tokens
  • High variance → choppy streaming experience

Typical Latency Ranges (2024)

Model / ServiceTTFTTPS
GPT-4o~0.5–1s40–80
Claude 3.5 Sonnet~0.5–1.5s50–100
GPT-3.5 Turbo~0.3–0.5s80–150
Local LLaMA 3 8B (GPU)~0.1–0.3s50–200
Local LLaMA 3 70B (GPU)~0.3–0.8s15–50

Factors Affecting Latency

Input Side

FactorEffect on Latency
Prompt length (tokens)↑ Prefill time linearly
Context window usage↑ Prefill time
KV cache miss (no caching)↑ TTFT significantly

Model Side

FactorEffect on Latency
Model size (parameters)↑ Both TTFT and TPOT
Number of layers↑ TPOT
Attention head count↑ Memory bandwidth requirements
Quantization level↓ TPOT significantly

Output Side

FactorEffect on Latency
Output length (tokens)↑ Total latency linearly
max_tokens settingDoesn't affect speed, only when to stop

Infrastructure Side

FactorEffect on Latency
GPU memory bandwidth↓ TPOT when higher
GPU count (tensor parallel)↓ TPOT
Server load / queue depth↑ TTFT when high
Geographic region↑ Network RTT when far
Prompt caching hit↓ TTFT significantly

Latency vs. Throughput Trade-off

OptimizationLatency EffectThroughput Effect
Small batch size↓ (good for latency)↓ (bad for throughput)
Large batch size↑ (bad for latency)↑ (good for throughput)
Speculative decoding↓ Latency≈ Same or slightly better
Quantization (INT8/INT4)↓ Latency↑ Throughput

Latency Reduction Techniques

Speculative Decoding

  • A small "draft" model generates K tokens quickly
  • The large "target" model verifies all K tokens in one forward pass
  • Accepted tokens are kept; rejected tokens fall back to the target's output
  • Net effect: 2–3× speedup at same quality

Prompt Caching

  • Cache the KV states for repeated system prompts / documents
  • Cache hit → prefill cost drops by ~90%
  • Supported by: Claude (Anthropic), GPT-4o (OpenAI prompt caching)

Quantization

  • INT8 or INT4 weights → smaller memory footprint → faster memory reads
  • Modest quality loss, significant speed gain

Flash Attention

  • Reorders attention computation to be I/O efficient
  • Reduces memory bandwidth requirements for attention
  • 2–4× faster attention computation

Streaming

[Key Insight] Not faster — just makes latency feel lower. First tokens appear immediately; users read while rest generates.

Smaller Models

  • Use the smallest model that meets quality requirements
  • GPT-3.5 vs. GPT-4: ~5× cost, ~3× faster

Latency SLAs (Service Level Agreements)

Common real-world targets by use case:

Use CaseTTFT TargetTPOT Target
Interactive chat< 500ms< 30ms
Autocomplete / copilot< 200ms< 20ms
Async document processing< 5sAny
Background batch jobsNo TTFT targetThroughput-focused

Measuring and Monitoring Latency

  • LangSmith, Helicone, Braintrust — LLM observability platforms
  • Custom logging: timestamp at request, first chunk, last chunk
  • p50 / p95 / p99 percentiles — don't rely on averages
  • Test under realistic load (latency degrades under high concurrency)

Related Concepts

  • Inference, Token, Context Window, Streaming, KV Cache, Throughput, Quantization, Prompt Caching