Ikva ML Accelerator IP Core

Industry’s first 8-bit/1-bit deep learning accelerator IP core and software framework

Our system is going under maintenance starting December 9, 2024 at 8:30 AM Pacific and ending December 9, 2024 at 10:30 AM Pacific. During this window, the website may not be reachable. For immediate assistance, please contact techsupport@latticesemi.com.

相关产品

Full stack product: Ikva is more than just an IP core for ML. We offer a training framework, inference software, data preprocessing framework, and pretrained models, optimized for highest performance.

High-efficiency, low power scalable solution. Depending on the application, the Ikva IP core can be configured to find the perfect balance between performance and resource utilization.

Software-defined models. There is no need to reconfigure the hardware when the model is changed. Multiple models can be inferenced without changing the bitstream.

Features

  • Support for 8-bit and 1-bit convolutions with kernels of arbitrary size, depthwise convolution, skip connections, fully connected layers as well as high-precision (32-bit) batch normalization.
  • Industry-standard AXI4 interface.
  • Control firmware ready for RISC-V and Arm Cortex-M processor cores.
  • Inference firmware compatible with TensorFlow Lite for Microcontrollers and Larq.
  • It's possible to reuse the IP core block RAM as a system memory in SoC.

Block Diagram

Ordering Information