basicmi/AI-Chip: A Curated Index of AI Accelerators, Not a Benchmark Suite
A list of ICs and IPs for AI, Machine Learning and Deep Learning.
At a glance
- What is it?
- AI-Chip is a maintainer-edited catalogue of AI and machine learning chips, IP blocks and startups, published as a README and a GitHub Pages site. It is useful for orientation and vendor discovery, but it carries no licence file, no releases and no measured performance data of its own.
- Who is it for?
- Adopt AI-Chip as a starting map if you need to find vendor names, group them by category and follow the links onward to primary sources, and treat every entry as a pointer rather than a verified specification. Do not use it as a procurement shortlist, a performance comparison or a citation source for a technical decision, because the repository ships no licence file, no releases and no benchmark data of its own.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 62 days ago.
- What is it written in?
- Mainly PHP, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What AI-Chip Actually Is
AI-Chip is a list. The README describes it as a set of ICs and IPs for AI, machine learning and deep learning, and the repository is organised around that single artefact: one long README plus an image set under resource/ and a published page at basicmi.github.io/AI-Chip/. The GitHub description repeats the same scope. There is no library, no CLI, no daemon and no generated dataset. The PHP label on the repository is a classification detail, not a signal that you install anything; nothing in the supplied material describes PHP code that a user runs.
The audience is narrow and specific. It is written for engineers and analysts who need to know which companies build AI accelerators, which IP vendors license NPU blocks, and which startups exist in the category. That is a discovery problem, not a benchmarking problem. If you already know the vendor you want and need throughput numbers, register pressure or memory bandwidth, this repository does not answer your question and does not claim to.
How the Catalogue Is Organised
The structure is a four-way split, stated explicitly in the Shortcut table: IC Vendors, Tech Giants and HPC Vendors, IP Vendors, and Startups. IC Vendors lists Intel, Qualcomm, Nvidia, Samsung, AMD, IBM and Marvell. Tech Giants and HPC Vendors covers Google, Amazon AWS, Microsoft, Apple, Alibaba Group, Tencent Cloud, Baidu, Fujitsu, Nokia, Facebook and Tesla. IP Vendors names ARM, Synopsys, Imagination, CEVA, Cadence and VeriSilicon. The Startups row is the longest and runs from Cerebras and Graphcore through to newer names such as d-Matrix, Neureality, Quadric, EdgeQ and Innatera Nanosystems.
That taxonomy is the real product. It separates companies that sell silicon from companies that sell licensable IP from companies that consume silicon internally at cloud scale, which is a distinction that matters when you are trying to work out whether a given accelerator is something you can buy, something you can license into your own design, or something you can only rent. The Startup list also mixes compute architectures with sensing and analog approaches, since Aspinity and Syntiant sit alongside Cerebras and Groq in the same row. The grouping is editorial rather than technical, and the README does not explain the criteria for placing a company in one bucket rather than another.
Navigation, Anchors and the Landscape Image
The README uses in-page anchors throughout. Each company name in the Shortcut table links to a matching anchor, for example #Cerebras, #Groq, #d-matrix, #SambaNova and #Nvidia, and the Latest updates block is a running list of the same anchors. This is the mechanism that makes a very long document usable: the Shortcut table is the index, the anchors are the destinations, and the Latest updates list at the top is a changelog of which entries were touched most recently.
Two images carry the visual summary. resource/AI-chips.png appears at the top, and resource/AI_Chip_Landscape_v0p7.png is embedded lower down. The version suffix v0p7 in the filename is the only version marker visible in the material, and it sits in an image path rather than in a release or a tag. There is also a QR code image for the editor's WeChat account, and the README links to a WeChat blog column called StarryHeavensAbove. The editorial voice is therefore personal and single-maintainer, with a Chinese-language publication channel running in parallel to the English README.
Getting to the Content: Clone, Read, or Use the Pages Site
There is nothing to build. The two practical routes are the rendered homepage at https://basicmi.github.io/AI-Chip/ and the raw README on the master branch. For a local copy, the standard clone applies:
git clone https://github.com/basicmi/AI-Chip.git
Because the default branch is master, a shallow fetch of that branch is enough to get the README and the resource/ directory:
git clone --depth 1 --branch master https://github.com/basicmi/AI-Chip.git
There are no config keys, no environment variables and no package manifest described in the supplied material. You do not run composer install, because nothing in the README indicates a PHP application to install, despite the repository being labelled PHP. The repo name in the image URLs still points at basicmi/Deep-Learning-Processor-List, which suggests the project was renamed or moved at some point and the asset links were not rewritten. If you mirror the README, keep that in mind: the image paths resolve through the older repository name.
No Licence File, No Releases, No Versioning Contract
The licence is listed as unknown, and no release has been retrieved. Those two facts shape how the repository can be used. Without a licence file, the default position under most copyright regimes is that no reuse rights are granted beyond what the platform's terms allow, so copying the catalogue into an internal wiki, a slide deck or a commercial tool is a question you would need to resolve with the maintainer rather than assume. This is a description of the situation, not legal advice; if redistribution matters to you, ask the editor directly.
The absence of releases also means there is no version to pin. The only version-like string in the material is v0p7 inside an image filename, and it does not correspond to a tag or a changelog entry you can check out. The Latest updates list functions as a human-readable change log, but it records additions of companies and news links, not corrections or removals. If an entry becomes wrong, for example a startup is acquired or a product is discontinued, the README has no visible mechanism for marking it stale. Treat every row as a claim with an unknown timestamp.
Where the List Falls Short
The most obvious gap is that the catalogue carries no specifications. There are no process nodes, no TOPS figures, no memory capacities, no TDP values and no interface details in the material supplied. Entries are names with links, and the substantive information lives on the vendor pages they point to. That is a defensible design for a link index, but it means the README cannot be used to compare two accelerators. You cannot answer whether one part is faster than another from this repository alone.
A second limitation is the update cadence. The Latest updates block is a flat list of additions, and several entries repeat the same company, for example Cerebras, Nvidia, Groq and SambaNova each appear more than once, alongside repeated additions of the MLPerf results link from MLCommons. The list is not dated in the material, so you cannot tell how recent any individual entry is. The last push timestamp is the only freshness signal available, and it tells you when the repository was touched, not which rows changed.
A third issue is scope drift. The Startup section mixes digital accelerators, analog in-memory compute, neuromorphic parts and optical approaches in one alphabetical run, and the README does not tag entries by architecture. If you are looking specifically for inference at the edge, you have to open each link and read the vendor's own description.
How It Differs From a Benchmark-Driven Source
The natural alternative for performance questions is the MLPerf results published by MLCommons, which the README itself links to under the AIChipBenchmarks anchor. The difference in approach is fundamental. MLCommons publishes measured results from a defined benchmark suite with rules about what counts as a valid submission, so you get numbers you can compare across systems. AI-Chip publishes names and links with no measurement layer at all.
The trade-off runs the other way too. A benchmark suite only covers hardware that someone chose to submit, and it says nothing about a startup that has announced a part but not benchmarked it. The AI-Chip list is broader and earlier: it includes companies such as Neureality, Innatera Nanosystems, EdgeQ and Quadric that may never appear in a benchmark table. Use the list to discover what exists, then use the benchmark source or the vendor's own datasheet to find out how it performs. Neither source replaces the other, and the README's own link to MLPerf suggests the editor sees them as complementary.
Who Should Keep This Bookmarked
Use AI-Chip if your job involves scanning the accelerator market: competitive analysis, sourcing conversations, or a literature review where you need to know that a vendor exists before you can search for its papers. The four-way Shortcut table is genuinely useful for that, because it separates merchant silicon from licensable IP from captive cloud silicon in a way that a plain alphabetical list would not. The landscape image gives a one-glance taxonomy that is faster to skim than the README itself.
Do not use it as the basis for a purchase decision, a tape-out decision or a technical comparison. The material contains no performance data, no licence and no release history, and the entries are undated. The concrete next step before trusting any row is to follow its anchor to the vendor link and verify the part against the vendor's own documentation, then check whether the MLCommons results page linked under AIChipBenchmarks covers that hardware. If a name in the Startup list cannot be confirmed on the vendor's own site, treat it as unverified until it can.
Editorial conclusion
Adopt AI-Chip as a starting map if you need to find vendor names, group them by category and follow the links onward to primary sources, and treat every entry as a pointer rather than a verified specification. Do not use it as a procurement shortlist, a performance comparison or a citation source for a technical decision, because the repository ships no licence file, no releases and no benchmark data of its own. Before you rely on any row, open the vendor page it links to and confirm the part number, the process node and the availability date against that vendor's own documentation.
Community notes