Think Silicon demos GPUs for embedded, wearables

Working with Edge Impulse on ML for wearables and AIoT.

David Harold

The NEOX GA100 is a GPGPU based on RISC-V architecture that combines 3D graphics and AI capabilities in one IP design. It offers flexibility for fabless customers to customize chips for different markets efficiently. The processor has an AI SDK for neural network model deployment, targeting wearables and AI applications. It supports TensorFlow Lite for MCU and requires minimal memory, making it suitable for various systems. Additionally, Think Silicon collaborates with Edge Impulse to integrate the NEOX AI SDK, enabling ML developers to utilize the GA100. 

What do we think? With X-Silicon and Think Silicon in the fray, and the world of RISC-V-based both in the fray, the world of RISC-V-based GPUs is getting more interesting despite targeting relatively simple applications with low API requirements. The aim is to deliver 3D graphics on a lightweight RTOS system with a lot of flexibility in how the end devices are packaged and segmented. Customers using Think Silicon today include Ambiq, STMicro, Qualcomm, and Racyics—and there are plenty of opportunities in the embedded space for such “one-IP” GPU/CPU solutions as RISC-V continues to build support.

Think Silicon NEOX GA100 GPGPU for the MCU segment 

Shown at Embedded World, NEOX GA100 is a GPGPU based on RISC-V that delivers both 3D graphics and ΑΙ in one IP architecture. The processor can be programmed using graphics and AI libraries independently or together to enable fabless customers more opportunities to package and segment chips according to use cases and markets, thereby maximizing silicon area while saving cost.

The IP is delivered with an AI SDK, a collection of open-source and proprietary tools for converting, optimizing, analyzing, and deploying pre-trained and post-trained neural network models. Target segments include sports, wellness, and audio wearables, and gesture and vision-based AI applications. It can perform offline graph optimization and AI inference based on TensorFlow Lite for MCU. The entire 3D/2.5D and VG API requires less than 150kB of memory, making it suitable for use from low-power, bare metal to RTOS- and Linux-based systems. 

Think Silicon also demonstrated its vector graphics GPU flagship, the NEMA pico VG series,  which includes support for the latest version of the open-source graphics software library LVGL (Light and Versatile Graphics Library), 9.0.

GA100 in action,

GA100 in action.

Edge Impulse is collaborating with Think Silicon to integrate its NEOX AI SDK into the Edge Impulse platform.  This collaboration empowers machine learning developers to leverage the GA100 RISC-V GPGPU. By doing so, embedded ML developers can access a comprehensive suite of ML tools for data collection, shaping, model training, and evaluating resource requirements such as inference time, power consumption, and RAM/Flash for their specific AI applications on the GA100.

According to George Sidiropoulos, managing director, of Think Silicon, “The flexibility and programmability of RISC-V is critical in the rapidly changing ML space. Our goal is to ease the developer experience and enable commercially valuable ML workloads to be developed and deployed on NEOX at scale.”

(Source: Think Silicon)