The intersection of retro gaming and AI-assisted engineering has produced a remarkable piece of hardware: an FPGA-based video scaler designed with the help of Claude. Hosted on GitHub by user bryfur, the 'tang-av-scaler' project converts 480i component video from sixth-generation consoles like the PlayStation 2, Xbox, and GameCube into crisp 1080p60 HDMI output. The design achieves an impressive latency of approximately 2 milliseconds, making it a viable option for competitive retro gaming where input lag is critical.

Technical Architecture and Performance

The board is built around a Sipeed Tang Mega 138K System-on-Module, which houses a Gowin GW5AST-138 FPGA. This isn't just a simple upscaler; it implements a sophisticated processing pipeline that includes motion-adaptive deinterlacing, noise reduction, and SMAA anti-aliasing. The scaling logic leverages AMDโ€™s FidelityFX Super Resolution 1 (FSR 1) algorithms, specifically the Edge Adaptive Upsampling (EASU) and Robust Contrast Adaptive Sharpening (RCAS), to enhance image quality beyond simple bilinear interpolation. The project utilizes 16% of the FPGA's logic and 75% of its block RAM, demonstrating efficient resource management for a complex video processing task.

AI-Assisted Design and Control

What sets this project apart in the current LLM news cycle is the role of Claude in the design process. The repository indicates that the board, FPGA design, and firmware were all developed with AI assistance, showcasing how large language models are increasingly becoming partners in hardware engineering. Control is handled via a Seeed XIAO ESP32-C5 module, which provides Wi-Fi connectivity and a web interface for real-time settings adjustment. This allows users to tweak parameters like sharpening strength and noise reduction without needing physical buttons or complex menu systems on the TV.

Current Limitations and Future Roadmap

As of November 12, 2026, the board (revision C-B) is functional and has been tested with a PlayStation 2. However, the Xbox and GameCube have not yet been verified on the hardware, though the signal format is identical. The project also faces some known issues, such as a double-load of the FPGA image at power-on that adds about 20 seconds to startup time. Furthermore, while 4K output is theoretically possible, the current scaler path is limited to 1080p, and 2160p60 links currently exhibit scrambling issues on displays.

Key Takeaways

  • The project achieves ~2ms latency from console capture to HDMI output, competitive with dedicated commercial scalers.
  • Claude was used to assist in designing the hardware, FPGA logic, and firmware, highlighting AI's growing role in embedded systems.
  • The implementation uses advanced algorithms like SMAA and FSR 1 for upscaling, rather than basic nearest-neighbor or bilinear methods.
  • The board is open-source, with KiCad files, RTL code, and firmware available on GitHub for community review and modification.

The Bottom Line

This project proves that AI assistants like Claude are no longer just for writing code snippets; they are becoming integral to complex hardware-software co-design. For retro enthusiasts, this offers a high-performance, customizable alternative to expensive commercial scalers, provided one is comfortable with open-source hardware quirks.