We first showcased the chip at CVPR 2022 and it was a pleasure to return with the commercialized product and share it with computer vision researchers and practitioners, many of whom remembered stopping by our booth last year.

Our first-gen NPU is designed to efficiently accelerate computer vision tasks in both data centers and on-prem and edge deployments. Running YOLOv5s, it delivers peak performance of 330 FPS on a single card, which allows for object detection at 40 FPS on eight concurrent streams.

At CVPR 2023 we ran example applications including object detection/multiple object tracking (yolov8), semantic segmentation (BiSeNetv1), instance segmentation (yolov8-seg), facial recognition (RetinaFace-MobileNet0.25), and pose estimation (yolov7-pose). We also showcased commercial deployment use cases, across applications of optical character recognition, super resolution, and object detection. You can also try the MOT and OCR demos on Hugging Face.

A commercial product, not just a proof of concept

Commercializing our first-gen chip has been crucial to achieving Furiosa’s long-term vision for dramatically expanding access to powerful AI compute.

By shipping our first-gen chip, we’re learning how best to meet business’s real world needs: which models are most important to support, how best to integrate our products into complex deployments, and how our products perform in a wide variety of commercial use cases.

This is especially important as we develop our second-gen chip, which is designed for inference with large language models and multimodal models. These algorithms are more resource-intensive than typical computer vision applications, so it’s essential that we build a hardware and software stack that delivers high performance, ease of use, and dramatically improved power efficiency. Our second-gen chip will launch in 2024.

Trends at CVPR

At our CVPR 2023 booth, we were able to hear feedback from people across the AI community and get first-hand insights into where the field of computer vision is heading next.

The experts we spoke with were excited to integrate large language models (LLMs) in their computer vision projects. We observed interesting new work to extend diffusion into denoising, image editing, and style transfer. NeRFs (Neural Radiance Fields, a method of generating high-quality 3D representations) have moved beyond proof of concept to editing, applications, and training process optimization.

We also heard concerns from both industry and academia about the rising cost of compute. Without hardware that offers lower TCO (total cost of ownership, which includes the cost of electricity to power and cool the chips as well as the upfront price tag), many fear innovation will be stifled and fewer people will have access to new AI-powered services.

Sign up here to be notified first about RNGD availability and product updates. You can also learn more about our first-gen product and schedule a demo here.