How NVIDIA and Pinterest Power Real-Time Visual AI
Pinterest and NVIDIA partnered on Blackwell GPUs to scale multimodal vision-language AI for 640 million users while cutting costs and latency
TL;DR
- Pinterest is collaborating with NVIDIA to scale its multimodal vision-language AI (VLM) for 640 million users.
- The partnership focuses on utilizing NVIDIA Blackwell GPUs to improve recommendation systems and visual discovery.
- Key goals include reducing costs and latency for AI-driven features such as the Pinterest Assistant.
- Pinterest is leveraging NVIDIA's Dynamo/vLLM stack on Blackwell GPUs for performance gains.
- The use of pre-computed embeddings and optimized inference pipelines has drastically reduced AI transaction costs and improved response times.
- This collaboration enables Pinterest to process more visual context in multi-turn conversations, leading to a more seamless user experience.