For AI agents: a documentation index is available at the root level at /llms.txt and /llms-full.txt. Append /llms.txt to any URL for a page-level index, or .md for the markdown version of any page.
    • Welcome to AIPerf Documentation
  • Getting Started
    • Profiling with AIPerf
    • Comprehensive LLM Benchmarking
    • Migrating from GenAI-Perf
    • GenAI-Perf vs AIPerf CLI Feature Comparison Matrix
  • Tutorials
NVIDIANVIDIA
Developer-friendly docs for your API
Privacy Policy | Your Privacy Choices | Terms of Service | Accessibility | Corporate Policies | Product Security | Contact

Copyright © 2026, NVIDIA Corporation.

LogoLogoDocumentation
On this page
  • Core Subcommands
  • Endpoint Types Support Matrix
  • Endpoint Configuration
  • Input Configuration
  • Output Configuration
  • Tokenizer Configuration
  • Load Generator Configuration
  • Arrival Pattern Configuration
  • Duration-Based Benchmarking
  • Concurrency Control
  • Gradual Ramping
  • User-Centric Timing (KV Cache Benchmarking)
  • Warmup Phase Configuration
  • Session/Conversation Configuration (Multi-turn)
  • Input Sequence Length (ISL) Configuration
  • Output Sequence Length (OSL) Configuration
  • Batch Size Configuration
  • Prefix Prompt Configuration
  • Audio Input Configuration
  • Image Input Configuration
  • Video Input Configuration
  • Multi-Run / Confidence Reporting
  • Parameter Sweeping
  • Adaptive Search / Bayesian Optimization
  • Search Recipes (Preset Experiments)
  • Service Configuration
  • Request Cancellation
  • Additional Features
  • Perf-Analyzer Passthrough Arguments
  • Data Exporters
  • Discrepancies
  • JSON Output
  • Advanced Features Comparison
Getting Started

GenAI-Perf vs AIPerf CLI Feature Comparison Matrix

||View as Markdown|
Previous

Migrating from GenAI-Perf

Next

Profile OpenAI-Compatible Text APIs Using AIPerf

This comparison matrix shows the supported CLI options between GenAI-Perf and AIPerf.

This is a living document and will be updated as new features are added to AIPerf.

Legend:

  • ✅ Fully Supported - Feature available with same/similar functionality
  • ⭐ Enhanced - Feature available in both tools, with broader capabilities or better ergonomics on the marked side
  • 🟡 Partial Support - Feature available but with different parameters or limitations
  • N/A Not Applicable - Feature not applicable
  • ❌ Not Supported - Feature not currently supported

AIPerf is the successor to GenAI-Perf, so most ⭐ marks fall in AIPerf’s column. They flag rows where AIPerf doesn’t merely match the GenAI-Perf surface but expands on it. ❌ vs ✅ rows are already self-explanatory and are left unannotated.

Core Subcommands

SubcommandDescriptionGenAI-PerfAIPerfNotes
analyze-traceAnalyze mooncake trace for prefix statistics❌✅
profileProfile LLMs and GenAI models✅✅AIPerf accepts a YAML config via profile -f config.yaml (CLI flags override)
plotGenerate visualizations from profiling data❌✅Auto-detects multi-run comparison vs single-run analysis; renders Pareto overlays when multiple artifact dirs are passed; supports dashboard mode
analyzeSweep through multiple scenarios✅✅AIPerf folds sweeps into profile via magic-list CLI flags, --variant, or YAML sweep: blocks (grid, zip, Sobol, Latin Hypercube). See Parameter Sweeping
configRun a YAML config end-to-end✅ (separate config subcommand)🟡AIPerf has no aiperf config <yaml> run shortcut — pass -f config.yaml to aiperf profile instead
create-template / config initScaffold a template config✅ (GenAI-Perf: create-template)✅AIPerf: aiperf config init -t <template>; supports --list, --search, --category for discovery
config expandPreview a sweep without running it❌✅Prints every variation the orchestrator would iterate; --full/--index/--format controls verbosity
config validatePre-flight validate a config file❌✅Runs the same load pipeline as profile; non-zero exit on fatal errors, warnings to stderr
pluginsList/inspect registered plugins❌✅aiperf plugins enumerates planners, recipes, exporters, dataset loaders, and more
synthesizeMaterialize a synthetic dataset to disk❌✅Useful for caching dataset generation between repeated sweep cells
process-export-filesMulti-node result aggregation✅N/AAIPerf aggregates results in real-time

Endpoint Types Support Matrix

--endpoint-type

Endpoint TypeDescriptionGenAI-PerfAIPerfNotes
chatStandard chat completion API (OpenAI-compatible)✅✅
completionsText completion API for prompt completion✅✅
embeddingsText embedding generation for similarity/search✅✅
rankingsText ranking/re-ranking for search relevance✅✅ ⭐GenAI-Perf has a single generic rankings endpoint (/v1/ranking, HF-TEI-compatible). AIPerf splits it into dedicated nim_rankings, hf_tei_rankings, and cohere_rankings endpoints.
hf_tei_rankingsHuggingFace TEI re-ranker API🟡✅GenAI-Perf has only generic rankings; AIPerf has a dedicated endpoint at /rerank
nim_rankingsNVIDIA NIM re-ranker API❌✅
cohere_rankingsCohere re-ranker API❌✅
chat_embeddingsChat-style multimodal embeddings (vLLM VLM2Vec)❌✅
embeddings (NIM)NVIDIA NIM embeddings endpoint❌✅AIPerf nim_embeddings; supports text and image inputs
responsesOpenAI Responses API endpoint❌✅Multi-modal (text, image, audio) with streaming
dynamic_grpcDynamic gRPC service calls✅❌
huggingface_generateHuggingFace TGI generate API✅✅/generate and /generate_stream supported
image_generationOpenAI-compatible image generation (/v1/images/generations)❌✅DALL-E-style text-to-image; supports raw export for image extraction
video_generationOpenAI/SGLang text-to-video (/v1/videos)❌✅Async polling; Sora / Wan2.1 / HunyuanVideo compatible; multipart-form requests
image_retrievalImage search and retrieval endpoints✅✅AIPerf serves NIM image retrieval / bounding-box detection at /v1/infer
nvclipNVIDIA CLIP model endpoints✅❌
multimodalMulti-modal (text + image/audio) endpoints✅✅AIPerf uses chat endpoint with multimodal content
generateGeneric text generation endpoints✅❌
kserveKServe model serving endpoints✅❌
templateTemplate-based inference endpoints🟡✅AIPerf supports multimodal and multi-turn templates
tensorrtllm_engineTensorRT-LLM engine direct access✅❌
visionComputer vision model endpoints✅✅AIPerf uses chat endpoint for VLMs
solido_ragSOLIDO RAG endpoint❌✅

Endpoint Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Model Names-m✅✅
Model Selection Strategy--model-selection-strategy
{round_robin,random}		✅✅
Backend Selection--backend
{tensorrtllm,vllm}		✅❌
Custom Endpoint--endpoint✅✅
Endpoint Type--endpoint-type✅✅ ⭐AIPerf supports 15+ endpoint types vs. GenAI-Perf’s 12; see detailed comparison
Server Metrics URL--server-metrics-url❌✅AIPerf uses --server-metrics (enabled by default, auto-collects Prometheus metrics from the inference endpoint at base_url + /metrics). See the note below on GenAI-Perf’s flag name.
Streaming--streaming✅✅
URL-u URL
--url		✅✅
Request Timeout--request-timeout-seconds❌✅
API Key--api-key✅✅ ⭐GenAI-Perf has no dedicated flag — users must pass -H 'Authorization: Bearer ...' manually
Request Content Type--request-content-type
{application/json,multipart/form-data}		❌✅Switch between JSON and multipart-form encoding (required by some video-gen servers)

GenAI-Perf’s --server-metrics-url is misleadingly named. Despite the “server metrics” label, the flag points GenAI-Perf at a Triton / DCGM telemetry endpoint (GPU power, utilization, memory) — it is not a general Prometheus inference-server metrics scraper. AIPerf splits this into two clearly-scoped flags:

  • --server-metrics — Prometheus inference-server metrics from the model endpoint (base_url + /metrics). Enabled by default; pass additional endpoint URLs to scrape extra targets.
  • --gpu-telemetry — GPU telemetry collection. Supports both the DCGM exporter HTTP endpoint (default; localhost:9400 + localhost:9401) and the local pynvml library (pass pynvml). Custom DCGM exporter URLs and a dashboard realtime view are also accepted.

If you’re porting a GenAI-Perf invocation, --server-metrics-url http://node:9400 maps to AIPerf’s --gpu-telemetry http://node:9400, not to --server-metrics.

Input Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Extra Inputs--extra-inputs✅✅
Custom Headers--header -H✅✅
YAML Config File-f --config✅ (separate config subcommand)✅AIPerf passes YAML to profile -f; CLI flags override file values
Input File--input-file✅✅
Inline Records in YAMLdataset.records: (YAML only)❌✅Embed dataset rows directly in the YAML config; >500 records emits a warning
Dataset Entries--num-dataset-entries --num-prompts✅✅GenAI-Perf and AIPerf both accept this flag. In AIPerf it is collapsed with --num-sessions / --conversation-num / --num-conversations into a single conversation count; GenAI-Perf keeps --num-dataset-entries and --num-sessions distinct (see Session Configuration).
Public Dataset--public-dataset❌✅sharegpt, aimo, mmstar, vision_arena, llava_onevision, speed_bench_* (50+ subsets), librispeech, voxpopuli, gigaspeech, ami, spgispeech, instruct_coder, blazedit_5k, blazedit_10k, …
HuggingFace Subset Override--hf-subset❌✅Override the HF subset/config for HF-backed public datasets
Custom Dataset Type--custom-dataset-type
{single_turn,multi_turn,random_pool,mooncake_trace,bailian_trace,burst_gpt_trace,sagemaker_data_capture}		❌✅GenAI-Perf infers dataset type from input file format
Dataset Sampling Strategy--dataset-sampling-strategy
{sequential,random,shuffle}		❌✅Controls how entries are drawn during benchmarking
Fixed Schedule--fixed-schedule✅✅
Fixed Schedule Auto Offset--fixed-schedule-auto-offset❌✅
Fixed Schedule Start/End Offset--fixed-schedule-start-offset
--fixed-schedule-end-offset		❌✅
Random Seed--random-seed✅✅
GRPC Method--grpc-method✅❌

Output Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Artifact Directory--artifact-dir✅✅
Checkpoint Directory--checkpoint-dir✅❌
Generate Plots--generate-plots✅✅ ⭐AIPerf replaces inline plot generation with the dedicated aiperf plot subcommand: dashboard mode, Pareto overlays across runs, configurable plot envelope, auto-plot hook
Auto-Plot After Profile--auto-plot --no-auto-plot❌✅Auto-runs aiperf plot on the artifact dir after the benchmark completes; honored by recipe defaults
Plot Required (Strict)--plot-required❌✅Treat auto-plot failures as fatal (non-zero exit)
Export Level--export-level --profile-export-level
{summary,records,raw}		❌✅Controls whether per-record and raw request/response files are emitted alongside the summary
Time-Sliced Metrics--slice-duration❌✅Window the benchmark timeline into fixed slices and compute metrics per slice
Enable Checkpointing--enable-checkpointing✅❌
Profile Export File--profile-export-file✅✅AIPerf works as a prefix for the profile export file names.

Tokenizer Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Tokenizer--tokenizer✅✅
Tokenizer Revision--tokenizer-revision✅✅
Tokenizer Trust Remote Code--tokenizer-trust-remote-code✅✅

Load Generator Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Concurrency--concurrency✅✅
Request Rate--request-rate✅✅
Request Count--request-count
--num-requests		✅✅
Request Rate w/ Max Concurrency--request-rate with --concurrency❌✅Dual control of rate and concurrency ceiling
Measurement Interval--measurement-interval -p✅N/ANot applicable to AIPerf
Stability Percentage--stability-percentage -s✅N/ANot applicable to AIPerf

Arrival Pattern Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Arrival Pattern--arrival-pattern
{constant,poisson,gamma}		❌✅Controls inter-arrival time distribution
Arrival Smoothness--arrival-smoothness
--vllm-burstiness		❌✅Gamma distribution shape: <1=bursty, 1=Poisson, >1=smooth

Duration-Based Benchmarking

FeatureCLI OptionGenAI-PerfAIPerfNotes
Benchmark Duration--benchmark-duration❌✅Stop after N seconds
Benchmark Grace Period--benchmark-grace-period❌✅Wait for in-flight requests after duration (default: 30s, supports inf)

Concurrency Control

FeatureCLI OptionGenAI-PerfAIPerfNotes
Session Concurrency--concurrency✅✅Max concurrent sessions
Prefill Concurrency--prefill-concurrency❌✅Limit concurrent prefill operations (requires --streaming)

Gradual Ramping

FeatureCLI OptionGenAI-PerfAIPerfNotes
Concurrency Ramp--concurrency-ramp-duration❌✅Ramp concurrency from 1 to target over N seconds
Prefill Concurrency Ramp--prefill-concurrency-ramp-duration❌✅Ramp prefill concurrency over N seconds
Request Rate Ramp--request-rate-ramp-duration❌✅Ramp request rate over N seconds

User-Centric Timing (KV Cache Benchmarking)

FeatureCLI OptionGenAI-PerfAIPerfNotes
User-Centric Rate--user-centric-rate❌✅Per-user rate limiting with consistent turn gaps
Number of Users--num-users❌✅Number of simulated users (required with --user-centric-rate)
Shared System Prompt--shared-system-prompt-length❌✅System prompt shared across all users (KV cache prefix)
User Context Prompt--user-context-prompt-length❌✅Per-user unique context padding

Warmup Phase Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Warmup Request Count--warmup-request-count✅✅
Warmup Duration--warmup-duration❌✅Duration-based warmup stop condition
Warmup Session Count--num-warmup-sessions❌✅Session-based warmup stop condition
Warmup Concurrency--warmup-concurrency❌✅Override concurrency during warmup
Warmup Prefill Concurrency--warmup-prefill-concurrency❌✅Override prefill concurrency during warmup
Warmup Request Rate--warmup-request-rate❌✅Override request rate during warmup
Warmup Arrival Pattern--warmup-arrival-pattern❌✅Override arrival pattern during warmup
Warmup Grace Period--warmup-grace-period❌✅Grace period for warmup responses
Warmup Concurrency Ramp--warmup-concurrency-ramp-duration❌✅Ramp warmup concurrency
Warmup Prefill Ramp--warmup-prefill-concurrency-ramp-duration❌✅Ramp warmup prefill concurrency
Warmup Rate Ramp--warmup-request-rate-ramp-duration❌✅Ramp warmup request rate

Session/Conversation Configuration (Multi-turn)

FeatureCLI OptionGenAI-PerfAIPerfNotes
Number of Sessions--num-sessions✅✅
Session Concurrency--session-concurrency✅✅Use --concurrency for AIPerf
Session Delay Ratio--session-delay-ratio✅✅
Session Turn Delay Mean--session-turn-delay-mean✅✅
Session Turn Delay Stddev--session-turn-delay-stddev✅✅
Session Turns Mean--session-turns-mean✅✅
Session Turns Stddev--session-turns-stddev✅✅

Input Sequence Length (ISL) Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Input Tokens Mean--synthetic-input-tokens-mean
--isl		✅✅
Input Tokens Stddev--synthetic-input-tokens-stddev✅✅
Input Tokens Block Size--prompt-input-tokens-block-size
--isl-block-size		❌✅Used for mooncake_trace hash_id blocks

Output Sequence Length (OSL) Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Output Tokens Mean--output-tokens-mean
--osl		✅✅
Output Tokens Stddev--output-tokens-stddev✅✅
Output Tokens Mean Deterministic--output-tokens-mean-deterministic✅❌Only applicable to Triton

Batch Size Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Text Batch Size--batch-size-text
--batch-size -b		✅✅
Audio Batch Size--batch-size-audio✅✅
Image Batch Size--batch-size-image✅✅

Prefix Prompt Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Number of Prefix Prompts--num-prefix-prompts✅✅
Prefix Prompt Length--prefix-prompt-length✅✅

Audio Input Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Audio Length Mean--audio-length-mean✅✅
Audio Length Stddev--audio-length-stddev✅✅
Audio Format--audio-format
{wav,mp3,random}		🟡✅GenAI-Perf supports {wav, mp3} only; AIPerf adds random
Audio Depths--audio-depths✅✅
Audio Sample Rates--audio-sample-rates✅✅
Audio Number of Channels--audio-num-channels✅✅GenAI-Perf accepts {1, 2}; AIPerf accepts {0, 1, 2} (0 disables)

Image Input Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Image Width Mean--image-width-mean✅✅
Image Width Stddev--image-width-stddev✅✅
Image Height Mean--image-height-mean✅✅
Image Height Stddev--image-height-stddev✅✅
Image Format--image-format
{png,jpeg,random}		🟡✅GenAI-Perf supports {png, jpeg} only; AIPerf adds random

Video Input Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Video Batch Size--video-batch-size --batch-size-video❌✅Set to 0 to disable video inputs
Video Duration--video-duration❌✅Seconds per clip; requires FFmpeg
Video FPS--video-fps❌✅Frames per second
Video Width/Height--video-width --video-height❌✅Resolution in pixels (both or neither)
Video Synth Type--video-synth-type
{moving_shapes,grid_clock,noise}		❌✅Synthetic content generator
Video Format--video-format
{webm,mp4}		❌✅Container format
Video Codec--video-codec❌✅Any FFmpeg-supported codec (libvpx-vp9, libx264, h264_nvenc, …)
Embedded Audio Track--video-audio-num-channels --video-audio-sample-rate --video-audio-codec --video-audio-depth❌✅Optional audio mux for video clips
Download Video Content--download-video-content❌✅Include video download time in request latency

Multi-Run / Confidence Reporting

AIPerf can repeat the same benchmark N times and report mean / std / confidence-interval / coefficient-of-variation across runs, optionally stopping early once a target metric stabilizes.

FeatureCLI OptionGenAI-PerfAIPerfNotes
Number of Profile Runs--num-profile-runs❌✅1-10 runs; >1 enables aggregate statistics
Profile Run Cooldown--profile-run-cooldown-seconds❌✅Stabilization gap between runs
Confidence Level--confidence-level❌✅0.90 / 0.95 (default) / 0.99 CI width
Disable Warmup After First--profile-run-disable-warmup-after-first❌✅First run warms, rest measure steady state
Consistent Seed Across Runs--set-consistent-seed❌✅Auto-pin --random-seed=42 for valid statistics
Vary Seed Per Trial--vary-seed-per-trial❌✅Capture input-noise + runtime-noise variance
Adaptive Convergence Stopping--convergence-metric --convergence-stat --convergence-threshold --convergence-mode❌✅Stop early when CI width, CV, or KS-distribution stabilizes

Parameter Sweeping

AIPerf folds GenAI-Perf’s analyze subcommand into profile via three composable mechanisms: magic-list CLI flags, --variant scenarios, and YAML sweep: blocks. Multi-cell sweeps stream a per-cell results table; QMC sweeps additionally write a sampling_design.json for reproducibility.

FeatureCLI OptionGenAI-PerfAIPerfNotes
Magic-List CLI Flags--concurrency 1,10,100 --request-rate 50,100,200 --isl 128,512,2048 --osl ... --isl-stddev ... --osl-stddev ... --conversation-turn-mean ...🟡✅Any CLI flag in the allowlist accepts a comma list and triggers a sweep
Grid Sweep--sweep-type grid (default)✅ (via analyze)✅Cartesian product of all magic-list flags
Zip Sweep--sweep-type zip❌✅Lockstep element-wise pairing; YAML form: sweep: {type: zip}
Scenario Sweep--variant --sweep-variant❌✅Repeatable [name:] key=value, ... per occurrence; emits a ScenarioSweep. GenAI-Perf’s analyze sweeps one stimulus at a time ({batch_size, concurrency, num_dataset_entries, input_sequence_length, request_rate}), so multi-parameter scenarios are not expressible.
Quasi-Monte-Carlo (Sobol)YAML sweep: {type: sobol}❌✅Low-discrepancy quasi-random sampling over continuous + integer dimensions
Latin Hypercube SamplingYAML sweep: {type: latin_hypercube}❌✅Stratified sampling alternative to Sobol
Sweep Variation Cooldown--parameter-sweep-cooldown-seconds❌✅Inter-variation pause
Same Seed Across Variations--parameter-sweep-same-seed❌✅Correlated comparisons vs. independent draws
Sweep Order--parameter-sweep-mode
{repeated,independent}		❌✅Outer loop = trials or variations
Live Sweep Table(auto) / --no-sweep-table❌✅Per-cell streaming results table; auto-suppressed for non-TTY, dashboard UI, or single-cell sweeps
Sampling Design Artifactsweep_aggregate/sampling_design.json❌✅Emitted only for Sobol / Latin Hypercube (QMC) sweeps
YAML Config Drivendataset: phases: sweep: multi_run: blocks✅ (separate config cmd)✅Single AIPerf YAML drives sweep + multi-run + plot envelope; aiperf config expand previews variations without running

Adaptive Search / Bayesian Optimization

AIPerf ships a native Bayesian-optimization search planner (Optuna + BoTorch preset, Hvarfner-DSP Matern-5/2 kernel), with native multi-objective Pareto support (qLogNEHVI), outcome constraints, posterior-regret stopping, and a curated set of preset “search recipes”. GenAI-Perf does not offer adaptive search.

FeatureCLI OptionGenAI-PerfAIPerfNotes
Search Space--search-space 'path:lo,hi[:kind]'❌✅Repeatable; CLI grammar supports int and real dimensions. YAML sweep.search_space[] also supports prior: log-uniform.
Search Metric--search-metric❌✅Tag from RunResult.summary_metrics
Search Stat--search-stat
{avg,p50,p90,p95,p99}		❌✅
Search Direction--search-direction❌✅Maximize / minimize
Search Iterations--search-max-iterations --search-initial-points --search-random-seed❌✅Sobol seed phase + GP fit + stopping
Search Planner Plugin--search-planner
{bayesian,monotonic_sla,smooth_isotonic,optuna}		❌✅bayesian is curated Optuna+BoTorch preset; third-party planners registerable
Optuna Sampler--optuna-sampler
{tpe,gp,botorch}		❌✅--search-planner=optuna expert mode
Optuna Acquisition--optuna-acquisition
{logei,qlogei,qnei,qlognei,qehvi,qnehvi,qlognehvi}		❌✅Modern noisy-EI defaults; multi-objective variants gated on len(objectives) > 1
Posterior-Regret Stopping--optuna-terminator
{regret,emmr,none}		❌✅RegretBoundEvaluator (Makarova 2022) / EMMR (Ishibashi 2023)
Percentile Pooling--search-percentile-pooling
{mean,pooled}		❌✅Pool raw samples across trials for tail-correct percentile objectives
SLA Filter--search-sla 'metric:stat:op:threshold'❌✅Repeatable; outcome-constraint or hard filter
Multi-Objective Paretoobjectives: [...] (YAML)❌✅qNEHVI / qLogNEHVI; emits Pareto front in search_history.json
Pareto Overlay Renderingaiperf plot <dir1> <dir2> ...❌✅Multi-directory invocation triggers the Pareto overlay handler

Search Recipes (Preset Experiments)

FeatureCLI OptionGenAI-PerfAIPerfNotes
Named Recipe--search-recipe❌✅Expands to a search-space + SLA filter + post-process pipeline
Pareto Sweep--search-recipe pareto-sweep --isl-osl-pairs '128/128,512/256,...'❌✅Multi-shape throughput/latency Pareto with paired ISL/OSL workloads
Max Throughput under TTFT SLA--search-recipe max-throughput-ttft-sla --ttft-sla-ms❌✅Log-uniform concurrency prior with TTFT constraint
Max Throughput under ITL SLA--search-recipe max-throughput-itl-sla --itl-sla-ms❌✅Streaming required
Max Concurrency under SLA--search-recipe max-concurrency-under-sla --tpot-sla-ms --e2e-sla-ms --error-rate-sla --search-style {smooth_isotonic,monotonic,bo,optuna,grid}❌✅Selectable 1D SLA-saturation strategy
Max Goodput under SLO--search-recipe max-goodput-under-slo --slo-attainment-fraction❌✅DistServe-style per-request SLO attainment (default 0.95)
Concurrency Ramp / Degradation Knee--search-recipe concurrency-ramp --degradation-threshold --degradation-metric-tag --degradation-stat --concurrency-min --concurrency-max --concurrency-steps❌✅Reports first concurrency where stat exceeds baseline × (1 + threshold)
Prefill TTFT Curve--search-recipe prefill-ttft-curve --isl-min --isl-max --isl-steps❌✅Log-spaced ISL ramp for prefill characterization
Decode ITL Curve--search-recipe decode-itl-curve --osl-min --osl-max --osl-steps❌✅Log-spaced OSL ramp for decode characterization

Service Configuration

FeatureCLI OptionGenAI-PerfAIPerfNotes
Record Processor Service Count--record-processor-service-count
--record-processors		❌✅
Maximum Workers--workers-max
--max-workers		❌✅
ZMQ Host--zmq-host❌✅
ZMQ IPC Path--zmq-ipc-path❌✅

Request Cancellation

FeatureCLI OptionGenAI-PerfAIPerfNotes
Request Cancellation Rate--request-cancellation-rate❌✅Percentage of requests to cancel (0-100)
Request Cancellation Delay--request-cancellation-delay❌✅Seconds to wait before cancelling

Additional Features

FeatureCLI OptionGenAI-PerfAIPerfNotes
Goodput Constraints--goodput -g✅✅
Verbose-v --verbose✅✅
Extra Verbose-vv✅✅
Log Level--log-level❌✅{TRACE,DEBUG,INFO,NOTICE,WARNING,SUCCESS,ERROR,CRITICAL} (Loguru; case-insensitive in practice)
UI Type--ui-type --ui
{dashboard,simple,none}		❌✅
Help-h --help✅✅

Perf-Analyzer Passthrough Arguments

GenAI-Perf supports passing through arguments to the Perf-Analyzer CLI. AIPerf does not support this, as it does not use Perf-Analyzer under the hood.

FeatureCLI OptionGenAI-PerfAIPerfNotes
Perf-Analyzer Passthrough Arguments--✅N/AOnly applicable to GenAI-Perf

Data Exporters

FeatureGenAI-PerfAIPerfNotes
Console output✅✅
JSON output✅✅See discrepancies below
CSV output✅✅
API Error Summary❌✅
profile_export.json✅✅Use --export-level raw in AIPerf to get raw input/output payloads
Per-Record Metrics❌✅
inputs.json✅✅AIPerf format is slightly different

Discrepancies

JSON Output

  • Fields in the input_config section may differ between GenAI-Perf and AIPerf.

Advanced Features Comparison

FeatureGenAI-PerfAIPerfNotes
Multi-modal support✅✅
GPU Telemetry✅✅ ⭐AIPerf supports dual backends: DCGM exporter HTTP endpoints (default; localhost:9400 + localhost:9401, custom URLs accepted) and local pynvml. GenAI-Perf is DCGM-only via --server-metrics-url.
Streaming API support✅✅
Multi-turn conversations✅✅Full multi-turn benchmarking with session tracking
Payload scheduling✅✅Fixed schedule workloads
Distributed testing✅✅ ⭐GenAI-Perf has post-hoc multi-node result aggregation via process-export-files. AIPerf runs a single federated benchmark across nodes via ZMQ-TCP service-to-service communication.
Custom endpoints✅✅ ⭐AIPerf ships 15+ endpoint types incl. responses, chat_embeddings, nim_embeddings, nim_rankings, cohere_rankings, image_generation, video_generation, image_retrieval, solido_rag. See Endpoint Types Support Matrix.
Synthetic data generation✅✅
Bring Your Own Data (BYOD)✅✅Custom dataset support
Audio input support✅✅Both tools synthesize audio inputs (WAV/MP3). Neither computes audio-specific metrics (e.g. WER, audio-token-rate)
Vision metrics✅✅Image-specific performance metrics
Image generation benchmarking❌✅Text-to-image with raw export for image extraction
Video input benchmarking❌✅Synthetic video generation (FFmpeg) for VLM endpoints with configurable codec, resolution, FPS, audio
Live Metrics❌✅Live metrics display
Dashboard UI❌✅Dashboard UI
Reasoning token parsing❌✅Parsing of reasoning tokens
Arrival pattern control❌✅Constant, Poisson, Gamma distributions with tunable burstiness
Prefill concurrency limiting❌✅Fine-grained prefill queueing control for TTFT behavior
Gradual ramping❌✅Smooth ramp-up for concurrency and rate
Duration-based benchmarking❌✅Time-based stop conditions with grace periods
User-centric timing❌✅Per-user rate limiting for KV cache benchmarking
Configurable warmup phase✅✅ ⭐GenAI-Perf has only --warmup-request-count. AIPerf adds duration-based, session-based, and full per-phase overrides (rate, concurrency, prefill, arrival pattern, grace period, ramping).
HTTP trace metrics❌✅Detailed HTTP lifecycle timing (DNS, TCP, TLS, TTFB)
Request cancellation❌✅Test timeout behavior and service resilience
Timeslice metrics❌✅Per-timeslice metric breakdown
Interactive plot dashboard❌✅Web-based exploration with dynamic metric selection and filtering
Multi-run comparison plots❌✅Auto-detected Pareto curves and throughput analysis
YAML-first configuration✅✅ ⭐GenAI-Perf has a separate config subcommand to run a YAML file. AIPerf threads a single YAML through dataset, phases, sweep, multi-run, and plot envelope with a config init / expand / validate lifecycle.
Inline datasets in YAML❌✅Embed dataset records directly in the config file (no sidecar .jsonl)
Plot config envelope❌✅YAML carries the visualization spec; auto-plot materializes .aiperf-plot-config.yaml for reproducibility
Parameter sweeps✅ (analyze subcommand)✅ ⭐GenAI-Perf’s analyze sweeps one stimulus at a time. AIPerf folds sweeps into profile with magic-list CLI flags, multi-key scenario --variant, and YAML sweep: blocks (grid / zip / Sobol / Latin Hypercube).
Confidence reporting❌✅Multi-run aggregation with CI / CV / KS-distribution convergence stopping
Bayesian optimization❌✅Native Optuna+BoTorch search planner with Hvarfner-DSP kernel
Multi-objective Pareto search❌✅qLogNEHVI acquisition, hypervolume stopping, Pareto front in search_history.json
Search recipes (preset experiments)❌✅pareto-sweep, max-throughput-under-SLA, max-concurrency-under-SLA, max-goodput-under-SLO, concurrency-ramp, prefill-ttft-curve, decode-itl-curve
Live sweep table with Pareto marking❌✅Per-cell streaming sweep table inline-marks Pareto-dominant cells (★) for recipes that declare pareto_axes