CNN Edition
~0.5 TOPS
INT8
~3.3 mm² @ 28nm
Vision-first lightweight NPU.
For cost-sensitive edge devices where BOM and power matter.
- Object detection & classification — YOLO-class including YOLOv3-tiny and YOLOv4-tiny
- Face / scene recognition · gesture · keyword spotting
- DDR-less option — SRAM + NOR Flash inference path
- Ultra-low-power · BOM-friendly for emerging markets
- 5-stage RISC-V control core
- AXI Master · APB Slave · Synchronous FIFO interfaces
- Clock gating · external power gating supported
- Built-in Self-Test (BIST) — integration debug + production testing
Generative AI Edition
~1 TOPS
INT8 + FP16
~6.1 mm² @ 28nm
Conversational AI on edge.
CNN + Transformer dual-engine for full speech / LLM / TTS pipeline.
- LLM inference — Qwen 1.7B @ ~140 ms/token (MTP-ready ~47 ms/token, ~3× acceleration)
- Whisper Small speech recognition
- VITS text-to-speech synthesis
- CNN + Transformer dual-engine architecture
- Multi-model runtime switching — ASR → LLM → TTS on one chip
- Custom AI instruction set · 5-stage RISC-V control core
- AXI Master · APB Slave · Synchronous FIFO interfaces
- Built-in Self-Test (BIST) — integration debug + production testing