AI Potluck
Infrastructure / Edge AI hardware

Axelera Metis AIPU

Axelera AI

Edge AI inference accelerator from Axelera AI in the Netherlands, built on a quad-core digital in-memory-compute architecture that performs matrix-vector multiplication in place. A single quad-core Metis delivers on the order of 214 TOPS at about 15 TOPS per watt, sold as M.2 modules, PCIe cards and an integrated compute board, and paired with the Voyager SDK covering more than a hundred models for computer-vision workloads.

The openness axis weighs a publicly released SDK and retail availability against board design files that are not published. Verified 2026-08-13 via the Axelera product site and the voyager-sdk repository.

Openness

3 medium confidence
3.0
schematics
none
toolchain
open (Voyager SDK on public GitHub)
datasheets
brief (product pages/briefs public)
blobs
required (proprietary AIPU silicon + firmware)
retail
open_market (webstore + partners)

Proprietary in-memory-compute silicon with a publicly available Voyager SDK on GitHub and retail purchasing, but board design files are not open.

  • https://www.axelera.ai/ recorded 2026-08-13

    Metis AIPU product pages and a Shop, "D-IMC technology performs matrix-vector multiplications in-place, delivering 50+ TOPs per core" and "At 15 TOPs/W" - marketing briefs and a store, with no board design files offered anywhere in the document set

  • https://github.com/axelera-ai-hub/voyager-sdk recorded 2026-08-13

    public voyager-sdk repository under Axelera's own org, "This is a production-ready release of Voyager SDK" at v1.8, with release notes and a compatibility matrix in-tree

Adoption

3 medium confidence
3.0

Commercially available via webstore and partners with a public, versioned SDK, but a younger ecosystem than Hailo/Coral. The Voyager SDK is at v1.8 with release notes and a release-compatibility matrix in-tree, and Axelera announces an ASRock Industrial collaboration integrating Metis AIPU cards into edge platforms.

Capability

4 medium confidence
4.0

High raw inference throughput via in-memory compute, currently oriented toward CNN and computer-vision pipelines. The site no longer prints the ~214 TOPS aggregate the recorded value quotes; it states "50+ TOPs per core" and "15 TOPs/W" for the quad-core Metis, and Voyager still claims "100+ supported models". Four cores at 50+ TOPS is the same order as that aggregate so the band holds, but the specific 214 figure is no longer citable at source.

  • https://www.axelera.ai/ recorded 2026-08-13

    "delivering 50+ TOPs per core at FP32-equivalent accuracy with zero data movement", "At 15 TOPs/W", and "100+ supported models and a single toolchain" for Voyager

Verified 2026-08-13