WAVE-N v2: Chips&Media’s Custom NPU Retains 16-bit FP for Superior Efficiency at High TOPS

November 26, 2025 -- Chips&Media is announcing that the next-generation customized NPU, WAVE-N v2, is ready for release, delivering higher computational power along with improved area and power efficiency. WAVE-N v2 is also designed to support a wider range of applications, such as object detection and segmentation, in addition to visual quality enhancement tasks like super-resolution and noise reduction, which were supported by the previous generation.

Key Highlights

• Extreme efficiency, achieving up to 80% MAC utilization for modern CNN computation
• Highly optimized for real-time video processing across a variety of markets
• 16-bit floating-point support, eliminating risks associated with quantization and fixed-point arithmetic errors
• Improved area efficiency, surpassing even INT8 fixed-point implementations at high TOPS
• Superior power efficiency, preventing thermal issues by requiring zero or minimal DRAM bandwidth
• Supports a wide range of applications, including object detection, segmentation, and visual quality enhancement such as super-resolution, HDR, noise reduction, and more

While it introduces many updates and improvements, WAVE-N v2—successor to WAVE-N—continues to follow the same philosophy: a custom NPU designed for imaging applications requiring high quality and high resolution across markets such as IoT, automotive, and edge devices including mobile, compute, and XR.

WAVE-N v2 retains exclusive support for 16-bit floating-point computation to avoid the time-consuming optimization typically required to mitigate quantization and fixed-point arithmetic issues. It provides higher computational capability (TOPS) using 2K and 4K MACs within a single-core architecture. With improved area efficiency at higher MAC counts, the physical silicon area does not scale proportionally with the increased MACs. As a result, at higher TOPS using 4K MACs, the estimated WAVE-N v2 area is even smaller than that of INT8 fixed-point designs with the same MAC count.

Another key feature of WAVE-N v2 is its line-by-line processing approach for CNN computation. Unlike frame-by-frame processing, line-by-line processing can be handled entirely within the NPU’s internal SRAM, avoiding DRAM interactions that typically introduce latency and power issues. This approach eliminates these bottlenecks at the root, enabling real-time processing for high-definition and high-resolution imaging applications.

Like its predecessor, WAVE-N v2 is a fully programmable processing core, now supporting a broader set of imaging applications. In addition to visual quality enhancement tasks like super-resolution and noise reduction, applications such as object detection and segmentation also achieve significantly higher performance on WAVE-N v2.

Chips&Media continues to collaborate with partners delivering advanced, state-of-the-art AI networks to maximize the benefits of the WAVE-N customized NPU IP. These partnerships, along with demonstration showcases, will be unveiled to the public at CES 2026 in Las Vegas.

About Chips&Media,Inc.

Chips&Media is a multimedia IP leader offering HW video codec IP, image processing NPU, lossless/lossy frame buffer compression IP & others. Since 2003, Chips&Media has exclusively focused on the HW video codec IP industry. It has powered over 3,000,000,000 units in the market, adopted by more than 150 semiconductor companies. Chips&Media's video codec IP supports all major and legacy formats from 1080p up to 8K 120fps, including HEVC/H.265, AVC/H.264, AV1, VP9, APV, VVC, JPEG & others. The recent introduction of our image processing NPU IP is transforming Chips&Media into a premium multimedia IP provider. It is custom NPU designed to work for imaging application at high-quality and high-resolution in edge devices. For more information, please visit our website at www.chipsnmedia.com