# InferenceX

> Open-source platform that continuously benchmarks LLM inference performance across AI accelerators and serving frameworks.

InferenceX (formerly InferenceMAX) is an open-source inference performance research platform from SemiAnalysis. It continuously benchmarks popular open-source inference frameworks on real accelerator hardware and publishes the results on a public dashboard. The repository is licensed under Apache 2.0.

## What It Is

InferenceX is a benchmark suite and live dashboard for LLM inference serving. It tracks how software stacks such as SGLang, vLLM, TensorRT-LLM, CUDA and ROCm perform across GPUs and other accelerators, so results reflect the latest software releases rather than a fixed point in time. The repository is organized into projects: end-to-end inference serving benchmarks, networking and collective communication benchmarks (experimental beta), operator and kernel-level benchmarks (experimental beta), and open-source system-level power modelling.

## How Results Are Produced

The site states that every data point on the dashboard comes from a public GitHub Actions workflow run on the target hardware, with logs and artifacts publicly viewable. Model, framework, precision and parallelism settings are committed to the repo as shell scripts. The full benchmark database is published as a weekly public GitHub Release. The site says it adds 1,000+ new benchmark datapoints per week on average.

## AgentX and Hardware Coverage

AgentX is a benchmark for long-context, multi-turn agentic inference, with results for models such as Kimi K3, DeepSeek V4.1 Flash, GLM 5.3 and MiniMax M3. The README lists officially supported hardware including Vera Rubin NVL72, GB300 NVL72, GB200 NVL72, B200, B300, H100, H200, MI300X, MI325X, MI355X, TPUv7x Ironwood and RTX PRO 6000 Server, with others marked as coming soon. The dashboard is complemented by calculators and comparison pages such as TCO and performance-per-dollar views.

## Official Results Notice

The README states that only the SemiAnalysisAI/InferenceX repository contains official InferenceX results, and that forks and other repositories must be labelled unofficial. The project also documents a standalone path to run the AgentX harness against an existing server without CI or Slurm.

## Features
- Continuous end-to-end LLM inference serving benchmarks
- AgentX long-context multi-turn agentic benchmark
- Public GitHub Actions runs with open logs and artifacts
- Open recipes committed as shell scripts
- Weekly benchmark database snapshots published as GitHub Releases
- Live public performance dashboard
- Networking and collective communication benchmarks (experimental beta)
- Operator and kernel-level benchmarks (experimental beta)
- System-level power modelling
- TCO and token revenue calculators
- Standalone AgentX harness for benchmarking existing servers

## Integrations
GitHub Actions, SGLang, vLLM, TensorRT-LLM, CUDA, ROCm, PyTorch

## Platforms
LINUX, WEB, API, CLI

## Pricing
Open Source

## Links
- Website: https://inferencex.com/
- Documentation: https://deepwiki.com/SemiAnalysisAI/InferenceX
- Repository: https://github.com/SemiAnalysisAI/InferenceX
- EveryDev.ai: https://www.everydev.ai/tools/inferencex