We provide everything required to go from neural network model to SPU
Quickly evaluate and deploy a solution with access to our repository of pre-trained, pre-optimized neural network models. You may also customize or fine-tune these models for a specific application and hardware configuration.
Take your own trained or untrained model and optimize it for the SPU using FMOT tooling. FMOT provides Python utilities for cutting edge techniques like Quantization Aware Training, Heterogeneous Mixed Precision, and Sparsity Regularization to squeeze out every last drop of efficiency. When you're finished designing, training, fine-tuning, and testing your model, FemtoMapper will translate it into code that runs efficiently on the SPU.
Once the model is ready, FemtoBehave accurately estimates power, performance, area, and other key metrics for your system before deploying to an FPGA, IP block, or ASIC. The FemtoBehave library allows you to get these metrics working in standard Python. This simulation runs about 10x faster than an unsynthesized RTL simulation, enabling rapid iteration of model and hardware configurations for a quick and agile system design process. We also provide tools to generate a PDK-informed, logic-consistent simulation of the embedded SPU for cycle-accurate performance.
Rapidly prototype and test real-time performance with our FPGA prototype platform. Map your model to an FPGA prototype of the SPU hardware design to assess real-world performance before deploying to ASIC or SoC.
After simulating and verifying the performance of your system, generate an IP block that is ready for ASIC or SoC tapeout. We also provide pre-designed IP and embedded chips for a more turnkey solution.
Evaluate prospective deployments with Femtosense technology at low risk. We provide software tools and support for partners to benchmark models targeting Femtosense hardware. We can also take a partner-defined task and design, train, and benchmark reference models.
Prototype real-time deployments with Femtosense development kits to test our complete technology stack. We provide the materials for you to prototype a physical or virtual simulation in-house, but we also help you through the process. We can even provide the working prototype for you for a more hands-off process.
Deploy the full technology stack according to partner specifications. We provide embedded ASICs that are ready for deployment. Alternatively, we can provide IP blocks if you want to do the tapeout yourself or integrate it in an SoC. The model tools, software, and ongoing support are all included in case you want to reprogram your SPU or tweak how it's integrated in your system.