TECHNOLOGY / PiM NPU IP
Bring AI compute
closer to memory.
Integrate PIMIC’s Jetstreme™ Processing-in-Memory architecture and dataflow-based NPU technology into your own silicon platform.
THE ARCHITECTURE
Three pieces. One AI compute platform.
Jetstreme combines Processing-in-Memory, a neural dataflow engine and software tools to help execute supported AI models efficiently.
Jetstreme™ PIM architecture
Place matrix computation closer to stored model weights to reduce unnecessary data movement.
NPU + Dataflow
Coordinate neural-network operations and data movement across the compute architecture.
Software toolchain
Compile models and evaluate compatibility, dataflow and expected performance before deployment.
DEVELOPER ENABLEMENT
From model to implementation.
01 / COMPILE
AI Compiler
Converts supported model formats into the files required by the PIMIC NPU.
02 / VALIDATE
Modal Checker
Checks model compatibility and helps identify issues before deployment.
03 / SIMULATE
Dataflow Simulator
Helps assess dataflow and expected execution behavior before silicon deployment.
Exact model formats, supported operators, performance outputs and tool names should be confirmed against the current developer toolkit.
PiM NPU IP Licensing
Explore PiM for your next design.
Tell us about your target workload, model and silicon platform. PIMIC can discuss architecture fit, evaluation and potential IP integration.