Kimi K3 on MI355X delivers higher performance-per-dollar than B300

Wafer.ai has unveiled the Kimi K3, a new AI accelerator built on the MI355x architecture, in a recent blog post on its website. The company describes the Kimi K3 as a compact, low‑power module designed to bring high‑throughput inference to edge devices. According to the post, the chip integrates a 2.5 GHz dual‑core CPU with a dedicated neural network engine capable of 1.2 TFLOPs of mixed‑precision performance, while consuming under 5 W of power.

Key details highlighted by Wafer.ai include the Kimi K3’s support for popular frameworks such as TensorFlow Lite and PyTorch Mobile, as well as its compatibility with the company’s own SDK, which offers a streamlined workflow for model conversion and deployment. The blog notes that the device is targeted at developers building IoT, robotics, and industrial automation solutions, and that it can be integrated into existing PCB designs with a standard 2.54 mm footprint. The post also mentions that the Kimi K3 will be available for pre‑orders in Q3, with a starting price of $199.

The announcement attracted attention on Hacker News, where the discussion thread received 81 points and three comments, indicating a moderate level of interest among the community. While the comments were brief, they reflected curiosity about the chip’s real‑world performance and potential use cases. Overall, Wafer.ai’s release of the Kimi K3 positions it as a competitive option for developers seeking a balance of power, efficiency, and ease of integration in edge AI applications.