This is crucial for FuriosaAI, because we’ve set ourselves the ambitious goal of making AI computing sustainable and accessible to everyone. To do this in a world with soaring demand for AI hardware, we are building chips that are not just highly performant, but also programmable, scalable, easy to deploy, and vastly more power efficient.

Last fall FuriosaAI held its first hackathon (co-hosted with Elice) to assess the strength of our SDK and software stack for our first AI chip, WARBOY, which accelerates inference for more than 100 computer vision models.

A hackathon is a uniquely valuable way to get feedback on products like WARBOY. With no prior experience working with WARBOY or Furiosa’s SDK, participants had 24 hours to build and run a new computer vision application. After the event ended, we conducted a detailed survey of the engineers who took part, in order to learn what went well, where their pain points were, and how we can make it easier to accomplish real world tasks with WARBOY.

As we gear up to launch our second-gen chip, RNGD, later this year, we’ve been able to leverage the insights we’ve gathered from hosting the hackathon and working with customers using WARBOY in production applications.

More than 1,000 people submitted hackathon applications and 40 were selected as semifinalists. They gathered in a conference room downtown Seoul in late November to build their projects and compete for a series of prizes. In preparation, Elice installed 40 WARBOY cards on servers in their data center and Furiosa engineers (and company advisors like Nuno Lopes, a compiler expert and associate professor at Lisbon University of Science and Technology) were on hand to answer questions and provide support.

Hackathon participants started with little or no familiarity with the WARBOY card or the FuriosaAI SDK, and then had to develop, build and debug their project in just 24 hours. The teams included people with a very wide range of experience, including professional AI engineers, academic researchers with PhDs in Computer Science, and college students who have not yet worked professionally in the field.

Despite the pressurized environment of the hackathon, every team was able to get results from their project during the 24-hour event. Prize-winning projects ranged from accident detection on CCTV cameras to vocabulary-building tools for education.

One team used WARBOY to analyze people’s facial expressions and generate personalized color recommendations for them. "We are building a service that extracts features from images of faces. Image detection, and recognition are all implemented by AI,” said the team leader, Gayeon Yu. “So WARBOY's remarkable performance and a specialized architecture in computer vision are helping us greatly in this project."

The grand prize was awarded to a team that created a service to protect people’s privacy by blurring their faces in video in real time. They deployed two different computer vision models running simultaneously on a single WARBOY card: One model to detect the presence of people in the video, and one to blur the faces and bodies of people once they’ve been detected.

We were pleased to see hackathon participants reported high satisfaction and low difficulty with the Furiosa SDK. But we also found ways we can improve it – especially for students and other relatively inexperienced practitioners. These areas include additional resources on quantization and exporting to ONNX, as well as adding more documentation and low-level APIs.

This feedback has been top of mind as we prepare for the launch of RNGD (pronounced “Renegade”), which will be the most efficient data center accelerator for inference with high-performance LLMs and multimodal models.

You can learn more about WARBOY here and sign up here for updates on RNGD as we get closer to launch.