The Coralboard SL is available now (Fig. 1, see image above). It’s powered by Synaptics’ SL2619 AI SoC with an NPU based on Google’s Coral core architecture. The Astra SL2619 has a pair of 2-GHz Arm Cortex-A55 cores along with a 200-MHz Cortex-M52 with Helium support for real time chores (Fig. 2). The Synaptics Torq/Coral NPU delivers 1 TOPS.

2. The Astra SL2610 features a 1-TOPS NPU and multicore CPU and GPU. (Electronic Design)

Internally the NPU incorporates RISC-V cores. The board adds 2 GB of DDR4 DRAM, CSI/DSI imaging ports, QWICC and mikroBUS connectors, and an M.2 expansion slot for Wi-Fi and Bluetooth connectivity.

The Cortex-A55 cores can run operating systems like Yocto Linux while the Cortex-M52 is able to run an RTOS like FreeRTOS or Zephyr. With its NPU, the SL2619 can handle large language models (LLMs) locally.

The new Coralboard SR runs Synaptics’ SR250 (Fig. 3). It’s designed for even lower power operation by replacing the more powerful dual-core Cortex-A55 with an 800-MHz Cortex-M85. The NPU is also streamlined, delivering 400 GOPS. The Coralboard SR has the same set of interfaces, including NOR flash memory and a MicroSD card slot.

Both platforms target vision and audio AI applications. They contain the same hardware root-of-trust security subsystem.

The SR200 family includes a number of different platforms, such as the SR210 designed for very low power, always-on voice and sensing. The NPU delivers 200 GOPS. The SR230 adds vision, display, and optional audio support with 300 GOPS of performance. And the SR250 offers 400 GOPS of performance plus dual vision and display paths.

All of the platforms’ NPUs utilize Synaptics’ Torq AI engine supported by Synaptics’ SDK. The compiler is based on the open-source MLIR (Multi-Level Intermediate Representation) subproject that’s part of the LLVM project, host to the popular compiler collection.

“The MLIR project provides a modular and extensible intermediate representation (IR) framework intended to facilitate constructing domain-specific compilers and improve compiling for heterogeneous computing platforms.” Wikipedia)

The MLIR project supports a variety of AI compiler ecosystems, including runtimes like TensorFlow.