AnyVision-AI accelerator IP — sub-1 W CNN inference, silicon-proven on GF 22FDX
AnyVision-AI delivers 2.4 TOPS at 880 mW for INT8 CNN inference, with a fully synthesisable footprint and a Keras / ONNX toolchain. First customer SoC has now sampled.
Author
Antikode
AnyVision-AI is our second-generation neural accelerator IP, targeted at battery-powered edge devices that need on-sensor CNN inference without offloading to a host. The headline number is 2.4 TOPS sustained at 880 mW on GlobalFoundries 22FDX (typical, 0.8 V, 25 °C), or roughly 2.7 TOPS/W at the SoC pin.
The architecture is a 16x16 INT8 systolic MAC array with a configurable on-chip weight cache, structured sparsity support (2:4), and a hardware ReLU/quantize block. INT4 mode trades precision for ~1.7x throughput on suitably-trained networks. The toolchain takes Keras, PyTorch, or ONNX models, runs the Antikode quantization-aware fine-tune, and emits a binary that the on-die loader can stream from external flash at boot.
First-customer SoC sampled in late February — a smart-camera SoC for an Indonesian OEM, running MobileNetV3-Small at 30 fps for person/vehicle classification while staying under a 1.2 W power envelope including the image sensor interface.
Targeted use cases
Smart cameras, doorbell and intercom systems, agricultural monitoring (pest and disease detection from a battery-powered field unit), retail anti-shoplift, and basic driver-monitoring for aftermarket automotive. AnyVision-AI is not aimed at LLM inference or large transformer workloads — it is purpose-built for the sub-50 MB CNN class where power and silicon area matter more than peak throughput.
AnyVision-AI is available now under per-tape-out and per-unit royalty licensing. Reference SoC integration with the Antikode RV-32IM host and a standard MIPI CSI-2 input is included in the evaluation package. Datasheet and quantization-aware training guide available under NDA.
Keep reading
Related stories · Product
ProductAntikode RV-32IM Core v1.2 — silicon-proven on TSMC 22ULL assaa
Our flagship 32-bit RISC-V core hits 480 MHz worst-case at 0.72 V on TSMC 22ULL with a 2.1 CoreMark/MHz score. v1.2 ships with hardened debug, JTAG-PMP, and an updated UVM testbench.
Read articleEngineeringTest new article
Test Antikode Antikode Sub Ops is a Jakarta-based fabless design house. We license silicon-proven embedded RISC-V cores and AI accelerator IP, and we deliver custom RTL and FPGA-to-ASIC services on TSMC 28nm and GF 22FDX. Founded 2018, 47 IPs in production, 220M+ units shipped
Read articleFPGA-to-ASIC migration: the lessons we keep relearning
Most FPGA prototypes are not ASIC-ready, no matter what the vendor synthesis report says. After 12 tape-outs, here is the short list of things that always bite — and how we now catch them in week one.
Read article