AiOnIc processor

HW / AIONIC PROCESSOR

The chip behind the demos.

AiOnIc chip photograph from Acube’s Our Services section — AiOnIc · photograph from Acube / Our Services

AiOnIc is the edge-AI LSI designed by ArchiTek. The guide follows the engines used by our deployed kernels:

PSS

Link commands and select the next engine. A chain can contain many operations within one execution submission.

DMAP

Move and rearrange tiles between memory and VREG operands.

MMAP

Multiply prepared matrix tiles and reuse operands held in VREG.

Run programmable arithmetic, normalization, Softmax and precision corrections where selected.

SW / TOOLS THAT MAKE IT RUN

Prepare. Link. Validate.

Convert the trained model

Freeze the matching checkpoint, tokenizer and processor. Extract weights, choose tile dimensions and prepare FP16 operands.

prepare_vision_encoder.py · build_whisper.py

Build a resident program

Lay out parameters and scratch, emit DMAP/MMAP/KP commands, relocate addresses and link the PSS chain.

resident_attention.py · resident_batches.py

Check the original values

Compare tile outputs and greedy tokens with ONNX or the original FP32 model. Retain failed trials and measure actual submissions.

tools/prepare_reference.py · profile_instructions.py

Run the application

Resident C++ owners connect model execution to camera capture, Japanese questions, caption display and USB audio.

resident_caption.cpp · run_voice.sh

Acube’s graphical development tools

AGTSAiOnIc runs as a Linux server. AGTCAiOnIc provides a graphical client on Windows, Mac and Linux for assembling algorithms with connected plugins and sending them to the server.

AGTCAiOnIc graphical tools and plugin interface

Acube’s GUI development project →