FPGA Cards Get New Life as Cut-Rate AI Accelerators

Startup Fortune recently featured a project on GitHub where a developer, going by the handle Nero7991, managed to run the Qwen3.5-9B transformer model entirely inside FPGA fabric. The hardware used for this project is an SQRL FK33, a Xilinx XCVU33P FPGA card that was originally sold to mine cryptocurrency. The card retailed for over $1,000 when new, but can now be found on eBay for around $280. This project is significant because it shows that even with the current shortage of Nvidia GPUs, there are alternative solutions available, albeit ones that are slower and more DIY-oriented.

The project uses INT4 quantization to shrink the model down to 4-bit weights, which makes 8GB of HBM2 memory enough to hold the model's parameters. The 27B target requires a two-module setup, but the results are still impressive given the hardware used. The project was able to achieve 2.5 tokens per second on a two-FK33 pipeline split of the 9B model, which is slower than what a modern Nvidia GPU can do. However, it's a demonstration that the constraint of GPU supply is not fixed and can be worked around with enough patience and technical expertise.

The same mining boom that drove GPU prices into the stratosphere for gamers in 2017 and 2021 left behind a graveyard of specialized chips that never found a second life. However, with the rise of AI and the need for cheap and available silicon, these boards are getting a second career as cut-rate AI accelerators. This development may not threaten Nvidia's pricing power anytime soon, but it's an interesting example of how innovation can come from unexpected places.