Gati: Hardware Accelerated DNNs on FPGAs (github.com)

Gati has been introduced as a hardware-accelerated Deep Neural Network (DNN) inference engine tailored for FPGA platforms, offering a powerful solution for machine learning workloads. Developed entirely in Verilog, Gati emphasizes portability across different FPGA vendors and supports a wide range of neural network architectures, including CNNs. Its architecture is optimized for resource-constrained devices, featuring a scalable design that accommodates larger FPGAs while maximizing throughput and minimizing resource utilization. The significance of Gati lies in its ability to address common challenges faced in deploying large neural network models, particularly those with high memory demands. By leveraging efficient DRAM utilization, optimized memory access patterns, and a streaming dataflow architecture, Gati ensures effective management of feature maps and weights. Key technical capabilities include support for quantized inference, a variety of DNN operations, and a straightforward ONNX model deployment workflow. As Gati continues to evolve, it promises to broaden compatibility and enhance performance, making it a valuable addition to the AI/ML community focused on FPGA-based acceleration.