Software & Machine Learning to Realize Analog Compute for Physical AI

Blumind provides the software, simulation, compiler and evaluation stack that you need to deploy your AI model on AMPL™ analog silicon. Six steps take you from a trained model to a validated design: faster time to design-in, an accelerated path to production and lower integration risk.

From trained model to validated silicon in six steps

Faster time to design-in •  Accelerated path to production •  Lower integration risk
Models and datasets: pre-trained audio, vision and sensor models with curated datasets

1. Models & Datasets

Workload ingest

Start with ready-to-deploy multimodal models (model zoo) and curated datasets for audio, vision and sensor workloads. Use a Blumind reference model as your baseline, or bring your own model and data. Tune the model to your product and your use case.

Training flow: noise-aware, quantized training with standard ML frameworks

2. Training Flow

Standard ML frameworks

Train in the ML frameworks that your team already uses. The Blumind training flow adds noise-aware, quantized training. The model learns the behavior of analog compute during training, so it keeps its inference accuracy when it runs on silicon.

AI-driven compiler and mapper: PyTorch or ONNX model compiled and mapped to the analog compute array

3. AI-Driven Compiler & Mapper

Architecture-aware

Import your PyTorch or ONNX model. The compiler optimizes the model graph and translates it to the AMPL™ architecture. The mapper then places each layer on the analog compute array to get the best use of the silicon.

Digital twin: behavioral model that simulates power, performance and latency before silicon

4. Digital Twin

Simulation

A behavioral model of the silicon simulates power, performance and latency before you build hardware. Use the digital twin to compare model options, confirm your power budget and find problems early in your design cycle.

Software development kit: library to provision, monitor and control the silicon

5. Software Development Kit

Easy integration

A comprehensive software library lets your host processor provision, monitor and control the silicon. Load models, start inference and read results from your own application code.

Hardware evaluation kit: reference design board with BM110, microphone and USB to test and validate on silicon

6. Hardware Evaluation Kit

De-risk & accelerate

The evaluation kit is a reference design, use it to test, measure and validate your model on real silicon before you commit to your own board design. (Learn more about BM110).