Infrastructure / Edge AI hardware

MemryX MX3

An edge AI accelerator from MemryX (USA) built on a dataflow architecture, where each MX3 chip provides ~6 TOPS and an M.2 module packs four chips for up to 24 TOPS at roughly 6-8W with passive cooling. The M.2 (2280, M-Key) module uses PCIe Gen3 and supports 4/8/16-bit weights plus BFloat16, with no model-zoo lock-in. Sold at retail (~$149) via Amazon, Mouser, and distributors, it won a 2025 Edge AI & Vision Alliance Product of the Year award and is supported by the MemryX SDK and open source tooling.

Verified live 2026-06-22 via primary sources. Proprietary dataflow silicon with public datasheets, retail availability, and several open source runtime/driver repos, though the compiler core is closed.

Openness

3 high confidence

3.0

schematics: none
toolchain: partial (open runtime (MxAccl) + utils + driver mirror, closed NeuralCompiler)
datasheets: public
blobs: required (proprietary silicon)
retail: open_market (Amazon/Mouser ~$149)

Proprietary dataflow silicon with public datasheets, retail availability, and several open source runtime/driver repos, though the compiler core is closed.

https://developer.memryx.com/specs/M.2_datasheet.html recorded 2026-06-22
MX3 M.2 datasheet: chip count, TOPS, data formats, power
https://github.com/memryx/MxAccl recorded 2026-06-22
open source MemryX C++ runtime library

Adoption

3 medium confidence

3.0

Low-cost retail availability (Amazon, Mouser), independent reviews, a 2025 industry award, and an active GitHub org point to growing but emerging traction.

https://www.tomshardware.com/tech-industry/artificial-intelligence/memryx-launches-usd149-mx3-m-2-ai-accelerator-module-capable-of-24-tops-compute-power recorded 2026-06-22
retail launch at $149 with broad distribution

Capability

3 high confidence

3.0

Moderate throughput with flexible precision, but on-chip parameter capacity caps deployable model size.

https://developer.memryx.com/specs/M.2_datasheet.html recorded 2026-06-22
per-chip/module TOPS, supported data formats, parameter capacity

Unchanged since 2026-08-01 (last edited, not re-checked)