Google’s open-source, RISC-V-based Coral NPU is integrated into Synaptics SL2610 Edge AI SoCs

Google has very recently introduced Coral NPU full-stack, open-source RISC-V-based platform for always-on AI on low-power edge devices and wearables. The first chip to integrate the Coral NPU is the upcoming Synaptics Astra SL2610 family.

Google Coral NPU

The Coral NPU aims to address the software fragmentation on entry-level AI accelerators that makes them difficult to program. By releasing an open-source NPU and associated source code, Google hopes its design will be adopted by silicon vendors, reduce software fragmentation over time, and help machine learning (ML) developers bring products to market faster.

Building on the works on the Coral platform, the new, open-source Coral NPU is comprised of three main components:

• A scalar core – A lightweight, C-programmable RISC-V core that manages data flow to the back-end cores. It uses a simple “run-to-completion” model for ultra-low power consumption and traditional CPU functions.
• A vector execution unit – A single instruction multiple data (SIMD) co-processor compliant with the RISC-V Vector instruction set (RVV) v1.0.
• A matrix execution unit – An efficient quantized outer product multiply-accumulate (MAC) engine purpose-built to accelerate neural network operations. The matrix core is still being worked on and will only be released on GitHub with the other components later this year.

Besides a hardware reference design for the NPU useful to SoC designers, Google will also release software tools, including the TFLM compiler for TensorFlow, an MLIR compiler, a simulator, and custom kernels. The base design delivers performance of around 512 GOPS while consuming just a few milliwatts for battery-powered devices such as edge devices, hearables, AR glasses, and smartwatches. You’ll find the documentation and tools on Google Developers’ website, and a few more details in the announcement.

Synaptics SL2610 with TorQ and Coral NPUs

The first hardware implementation of Google’s Coral NPU will be found in Synaptics’ SL2610 product line, also featuring the Torq T1 NPU delivering a combined 1 TOPS of AI performance, or twice the reference design performance.

SL2610 highlights:

• Application cores – Single and dual-core Arm Cortex-A55 @ up to 2.0 GHz
• Real-time cores – Arm Cortex-M52 with Helium System Manager (SM) domain, 256KB SRAM
• GPU – Optional Arm Mali-G31 3D GPU
• NPU (Optional) – Transformer-capable Torq and Coral NPU (RISC-V ML Core) for up to 1 TOPS of AI performance
• Memory I/F – Up to 4GB 16-bit DDR4/LPDDR4/DDR3L with inline ECC
• Storage I/F – 2x xSPI and eMMC 5.1 interfaces
• Display and Camera I/F – MIPI-DSI & CSI with 2160p30 and HDR
• Audio
  • 3x TDM/I2S with 16 channels
  • Support for 8 digital mics
  • 1x S/PDIF
  • Hardware audio and camera mute
• Networking
  • Single and dual GbE RGMII with Wake-on-LAN (WoL)
  • 1588 PTP and TSN, 802.1 p/q VLAN tagging
• USB – 2x USB 2.0
• Other peripheral interfaces
  • 2x SDIO 3.0, 4x TWSI/I2C, 1x I3C, x UART
  • 5x SPI, 99x GPIO
  • 12-bit ADC and up to 12x smart PWM modules in SM domain
  • Single/Dual CAN 2.0 A, B interfaces
• Security
  • PSA certified Level 3 (RoT), Level 2 (Product)
  • Secure boot, TRNG, RSA, AWS, SHA, ECC, HASH

Five parts will be available: SL2611, SL2613, SL2615, SL2617, and SL2619, and all will integrate the NPU subsystem, except SL2611.  The company says the Synaptics SL-series of embedded processors are “AI-Native Linux” and Android SoCs optimized for consumer, enterprise, and industrial IoT applications, leveraging hardware accelerators. Target applications include home appliances and automation, consumer goods, industrial control systems, Smart retail, charging infrastructure, healthcare diagnostics systems, casual gaming, and robotics and UAVs.

Synaptics has already developed a range of Astra processors for Edge AI applications, including the Astra-SR (Cortex-M55 + Arm Ethos-U55 NPU) and the SL1620, SL1640, or SL1680 (Cortex-A55 + 7.9 TOPS NPU), so the new SL261x processors appear to offer a mid-range solution. The product page mentions the Astra Machina SL2610 Dev Kit based on the SL2619 SoC, but public details are limited for the “early access” kit. It will be based on a SOM + carrier design with WiFi and Bluetooth, programmable I/Os, and debugging interfaces, and run Yocto Linux. There’s also a separate compiler for the Torq NPU.

The Synaptics Astra SL2610 product line is sampling now, and general availability is planned for Q2 2026. Additional information may also be found on the press release.

Jean-Luc Aufranc (CNXSoft)

Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.

Share this:

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.