Ampere open-sources Panthera, its cycle-accurate CPU performance model
Ampere's Architecture & Performance team has released Panthera on GitHub. Panthera is the cycle-accurate software performance model the team built to study prospective CPU features, evaluate microarchitectural tradeoffs and guide the development of Ampere's high-performance cloud processors. A User Guide in the repository explains how to build and use it and its associated tooling.
Ampere's stated reason for opening it up: industrial-grade performance models calibrated against production silicon have historically been closely guarded trade secrets, and the company believes the computer systems and architecture community would benefit from a modern, commercial-caliber modeling infrastructure. It argues that academic architecture research and graduate teaching have long leaned on academic simulators which, while foundational, often diverge from the realities and trade-offs of commercial server processors. Panthera, by contrast, was developed for designing CPU products. Because AmpereOne is deployed across major cloud providers, researchers, educators and practitioners can run workloads on both the model and real production silicon, then correlate the two. Ampere says this lets the community validate ideas against realistic microarchitectural baselines and separate the architectural "wheat from the chaff."
The README says Panthera was built from scratch by CPU architects with decades of experience across the semiconductor industry, and that its modeling abstractions reflect practical lessons from building high-performance CPUs. It is shared alongside companion tools such as catscan, with the aim of fostering collaboration across industry and academia.
The README also frames the release with a metaphor. Low-background steel is steel forged before atmospheric nuclear testing, free of background radiation and prized as a baseline for sensitive instruments. In that spirit, Ampere calls Panthera "low-background steel for computer systems": an end-to-end, industrial-grade performance model conceived, designed and hand-crafted entirely by human architects before the era of generative AI. The repository represents over 8 years of work and 6000 individual git commits by 45 people. Ampere hopes the human-authored codebase will serve as a trusted, high-watermark reference for human engineers and future automated hardware-modeling tools alike.
Panthera is released as a public reference implementation under a permissive open-source license. Ampere provides no commercial support or active feature roadmaps for the public repository, but welcomes the community to explore the code, adapt it for research and fork it for new architectural exploration.
The model covers several generations of Ampere processors. Configuration files use code names for the SoC products and their custom cores, and the README includes a decoder between marketing and engineering names. The AmpereOne (Siryn) configuration contains Ampere's original custom core, Furiosa, designed for high performance and low-power cloud computing; microarchitecture notes describe it in more detail. A product floorplan gallery maps AmpereOne, AmpereOne X, AmpereOne MX and AmpereOne Aurora to their SoC and core code names. Each core builds on the previous micro-architecture, and the model mirrors that with "delta" configurations that build on the previous generation's.
Key facts
- Ampere's Architecture & Performance team published Panthera, its cycle-accurate CPU performance model, under a permissive open-source license.
- The repository represents over 8 years of work and 6000 git commits by 45 people.
- Ampere says such industrial-grade models have historically been closely guarded trade secrets; researchers can compare the model against real AmpereOne silicon in major clouds.
- The model covers several generations: AmpereOne (core code-named Furiosa), AmpereOne X, AmpereOne MX and AmpereOne Aurora, via 'delta' configurations.
- No commercial support or active feature roadmap is offered; forks and research adaptations are welcome.
Why it matters
Ampere says performance models calibrated against production silicon have historically stayed trade secrets, and that academic simulators often diverge from commercial server processors. Panthera is the model its own architects use to design cloud CPUs, now public. Ampere also pitches it as "low-background steel": a codebase written entirely by humans before generative AI, offered as a reference for engineers and future automated hardware-modeling tools.
Who it affects
Ampere names researchers, educators and practitioners in computer architecture as the audience, along with graduate education that has relied on academic simulators. Anyone with access to AmpereOne instances in the public cloud can run workloads on both the model and the hardware.
How to use it
The repository is on GitHub at github.com/AmpereComputing/panthera, with a User Guide covering building and using Panthera and its tooling. Companion tools such as catscan are shared alongside it. It is released under a permissive open-source license, and Ampere invites exploring the code, adapting it for research and forking it. Configuration files use code names, and the README includes a decoder to marketing names.
How solid is it
The material is Ampere's own README, so every claim is the company's self-description. The repository's size is stated as over 8 years of work and 6000 commits by 45 people. The README gives no accuracy figures for the model against real silicon, and no benchmark results.
Risks and caveats
Ampere provides no commercial support or active feature roadmaps for the public repository. The README does not name the specific license beyond "permissive", and gives no accuracy tolerance versus AmpereOne hardware. The "entirely human-authored" description is Ampere's own framing, and the source does not say how it was verified.
“Panthera represents low-background steel for computer systems.”
— Ampere, Panthera README