In a visually-driven digital landscape, companies must capture attention quickly with engaging content. Imgix, a leader in online visual media, serves over 8 billion images and videos daily for renowned brands like Bugatti and Spotify. To meet the growing demand for high-fidelity media, Imgix has upgraded its infrastructure to a GPU-based environment on Google Cloud's AI Hypercomputer.
By transitioning to G4 VMs equipped with NVIDIA RTX PRO 6000 Blackwell GPUs, Imgix has significantly improved its processing capabilities, achieving a 50% reduction in latency and a sixfold increase in throughput without altering its core application code.
Addressing the Need for Instant Visuals
To effectively engage users, businesses require fast-loading, rich content that can be delivered to millions of devices simultaneously. Real-time transformations, such as resizing and applying effects, demand substantial computational power. Imgix's "just-in-time" approach allows for instant image processing upon request, eliminating the need for pre-rendered variations.
Leveraging Google Cloud's Infrastructure
Utilizing G4 VMs, Imgix benefits from a robust architecture that includes eight NVIDIA GPUs and two AMD CPUs, enabling efficient management of server tasks. This setup enhances throughput by up to 168% compared to standard configurations, allowing Imgix to handle multiple requests in parallel.
Imgix's Dynamic Processing Pipeline
Imgix's architecture supports over 150 visual filters, dynamically processing requests through a four-stage pipeline:
- Ingestion: Assets are retrieved and routed to a 2.5 petabyte storage cache on Google Cloud Storage, ensuring reliable access.
- Decoding: High-performance libraries decode assets into raw data, utilizing the G4 VM's parallelism for efficiency.
- Transformation: A custom Vulkan compute shader stack processes transformations as parallel math problems, enabling numerous simultaneous operations.
- Encoding and Delivery: Transformed images are re-encoded using hardware-accelerated tools and delivered via a global CDN.
Advanced Video and Image Analytics
Imgix employs NVIDIA's CUDA libraries for video analytics, enabling real-time object tracking and automated content analysis. For static images, the nvJPEG library offloads JPEG decoding to the GPU, preventing CPU bottlenecks and allowing immediate transformations.
Performance Gains Achieved
With the transition to G4 VMs, Imgix has realized significant performance improvements:
- 50% reduction in latency: Median latency decreased from 100 milliseconds to 50 milliseconds.
- Increased throughput: The G4 VMs can now manage up to six times the workload of previous nodes.
- Seamless migration: The transition was supported by updating Terraform scripts without code changes.
Alfonso Acosta, Head of Engineering at Imgix, emphasized that building on Google Cloud's AI Hypercomputer not only optimizes current workloads but also future-proofs the platform for advanced generative AI capabilities.
Orchestration for Scalability
To handle billions of daily requests, Imgix has developed a sophisticated orchestration model:
- Management: Google Cloud Run oversees session and account layers.
- Core Processing: Instance groups managed by Google Compute Engine host G4 VMs.
- Dynamic Scaling: Autoscaling is based on custom metrics for maximum efficiency.
- Self-Healing: A monitoring system automatically restarts GPU instances as needed.
- Optimization: NVIDIA Nsight Systems is used to identify and resolve performance bottlenecks.
Looking Ahead
Despite the impressive performance enhancements, Imgix continues to expand its AI capabilities, aiming to provide features like generative fill and background replacement. By leveraging Google’s AI Hypercomputer, Imgix is well-positioned to deliver real-time, production-ready AI editing for increasingly complex visual experiences.
