SynapseAI / SynapseAI Core (Habana's Gaudi compiler and runtime)
intel-gaudi-hpu, first-gen-gaudi, gaudi2
Fields
id
engines/xpu-synapseaiRepo
HabanaAI/SynapseAI_Core
Languages
C++C
License
no machine-detectable license (NOASSERTION on the GitHub API)
First release
2021
Design goal
compile-graphs-for-gaudigraph-based-executionresearch/reference-only
Scheduling
none
Backends
intel-gaudi-hpufirst-gen-gaudigaudi2
Notable features
This is the layer that makes Gaudi a graph machine. The production SynapseAI compiler is CLOSED SOURCE and shipped only inside the Gaudi software stack installer; SynapseAI Core is the only open-source artifact of the API, and the README itself labels it a 'reference implementation'The layer stack below the compiler is fully open and is the real architectural content: habanalabs kernel driver, hl-thunk user-space library, and the TPC LLVM compiler (HabanaAI/tpc_llvm) that emits code for the Tensor Processing Core. ELFTools in this repo parses ELF sections to extract metadata generated by that TPC LLVM compiler - i.e. the toolchain is open even where the compiler is notSynapse Backend is what executes a compiled graph on Gaudi, going through habanalabs and hl-thunk rather than through any CUDA-like runtimeTPC Kernels DB: sample TPC kernels with host-side code, which is where the scalar-vs-MME split from the Gaudi accelerator records becomes visible in code - MME does matrix work, TPC does elementwise and reduction work, and they are different engines on the same dieHardware-first in the most literal sense available in this repo: the README's own limitations section says synchronization is through the HOST, event APIs are no-ops, there are no built-in operations (reshape and split must be user-resolved or written as TPC kernels), and execution is limited to SINGLE-NODE GRAPHSExecutable without hardware. Two tests ship and both run on the reference implementation: div_test demonstrates building a single-node graph whose node is a divide operator implemented with a TPC kernel on the TPC engines; memcpy_test demonstrates the DMA engine moving a tensor in and out with no TPC kernel at all. Two engines, two mechanisms, named explicitlyGAUDI-SPECIFIC MINIMUM VERSIONS: habanalabs driver added Gaudi in kernel 5.7, but a SECURED device requires kernel 5.15+; plus CMake >= 3.5.1 and GCC >= 7.546 stars, 9 forks, 2 open issues, created 2021-07-25, last push 2025-02-03, ARCHIVED=true on the GitHub API
Hardware caveats
ABANDONED BY ITS OWN VENDOR, and this is the most important fact in the record. The README opens with 'PROJECT NOT UNDER ACTIVE MANAGEMENT', states Intel has ceased development, maintenance, bug fixes, new releases and updates, and that Intel no longer accepts patches. The GitHub API reports the repo as archived. Do not plan around this repo; plan around the fact that it tells you the production compiler is closedIt is a degraded reference, not the shipping compiler, and its own README says so in three separate ways: many APIs are 'not implemented for brevity', operations are synchronous through the host so synStreamWaitEvent and synEventCreate are no-ops, it implements no operations of its own, and it is limited to single-node graphs. Single-node-only is the crucial one - a graph machine that only runs single-node graphs is not the production execution modelThe README states plainly that its divergence from closed-source SynapseAI is why the user must write their own kernels for anything non-trivial, so it teaches you the API surface and the toolchain layout, not the behaviour you will getIt targets FIRST-GEN Gaudi and Gaudi2 and was last touched 2025-02-03, so even as a reference it predates Gaudi 3's on-die SRAM and two-die topology as recorded in [[intel-gaudi2]] and [[intel-gaudi3]]Gated pre-requisites: you need the driver, hl-thunk and the TPC LLVM compiler built before this repo will build at all2 open issues and 46 stars across five years - this is a code dump, not a community
Best for
Understanding what Gaudi's graph-based execution model actually consists of, from the layer names down (driver, hl-thunk, TPC LLVM, Synapse Backend) rather than from marketingLearning the TPC/MME engine split in the only form where both engines are visible in runnable codeStudying MME-vs-TPC nomenclature and the DMA engine as distinct hardware pathsAn honest-negative reference: a good example of a vendor open-sourcing the toolchain skeleton while keeping the compiler closed, and of what that costs the user
Avoid for
Any production use - the repo says in its own first line that it will never be maintained againLearning Gaudi 3 behaviour; this reference predates itUsing the sync-through-host and single-node-graph semantics as a model of production SynapseAI - they are the reference implementation's limitations, not the compiler's design
notes
Set status deprecated deliberately: the GitHub API reports archived=true and the README states Intel no longer accepts patches. The repo already carries [[xpu-gaudi-static-shapes]] and [[xpu-gaudi-hpu-graphs]] for the constraint mechanics, so this record focuses on the compiler/runtime layer and its openness boundary rather than repeating them. Repo metadata from the GitHub API on 2026-10-04. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2025-02-03 (20.0 months before this record), latest release v1.1.0 published 2022-11-27, 3 contributors (distinct committers including anonymous), open_issues_count 2 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), ARCHIVED (read-only), 46 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.
Notes
Set status deprecated deliberately: the GitHub API reports archived=true and the README states Intel no longer accepts patches. The repo already carries [[xpu-gaudi-static-shapes]] and [[xpu-gaudi-hpu-graphs]] for the constraint mechanics, so this record focuses on the compiler/runtime layer and its openness boundary rather than repeating them. Repo metadata from the GitHub API on 2026-10-04. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2025-02-03 (20.0 months before this record), latest release v1.1.0 published 2022-11-27, 3 contributors (distinct committers including anonymous), open_issues_count 2 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), ARCHIVED (read-only), 46 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.
Cross-references
This record references no other records — it is a leaf. It is cited by 1.