Perplexity AI Leverages NVIDIA Inference Stack to Handle 435 Million Monthly Queries
Terrill Dicki Dec 06, 2024 04:17
Perplexity AI utilizes NVIDIA's inference stack, including H100 Tensor Core GPUs and Triton Inference Server, to manage over 435 million search queries monthly, optimizing performance and reducing costs.
Perplexity AI, a leading AI-powered search engine, is successfully managing over 435 million search queries each month, thanks to NVIDIA's advanced inference stack. The platform has integrated NVIDIA H100 Tensor Core GPUs, Triton Inference Server, and TensorRT-LLM to efficiently deploy large language models (LLMs), according to NVIDIA's official blog.
Serving Multiple AI Models
To meet diverse user demands, Perplexity AI operates over 20 AI models simultaneously, including variations of the open-source Llama 3.1 models. Each user request is matched with the most suitable model using smaller classifier models that determine user intent. These models are deployed across GPU pods, each managed by an NVIDIA Triton Inference Server, ensuring efficiency under strict service-level agreements (SLAs).
The pods are hosted within a Kubernetes cluster, featuring an in-house front-end scheduler that directs traffic based on load and usage. This ensures consistent SLA adherence, optimizing performance and resource utilization.
Optimizing Performance and Costs
Perplexity AI employs a comprehensive A/B testing strategy to define SLAs for varied use cases. This process aims to maximize GPU utilization while maintaining target SLAs, optimizing inference serving costs. Smaller models focus on minimizing latency, while larger, user-facing models like Llama 8B, 70B, and 405B undergo detailed performance analysis to balance costs and user experience.
Performance is further enhanced by parallelizing model deployment across multiple GPUs, increasing tensor parallelism to achieve lower serving costs for latency-sensitive requests. This strategic approach has enabled Perplexity to save approximately $1 million annually by hosting models on cloud-based NVIDIA GPUs, surpassing third-party LLM API service costs.
Innovative Techniques for Enhanced Throughput
Perplexity AI is collaborating with NVIDIA to implement 'disaggregating serving,' a method that separates inference phases onto different GPUs, significantly boosting throughput while adhering to SLAs. This flexibility allows Perplexity to utilize various NVIDIA GPU products to optimize performance and cost-efficiency.
Further improvements are anticipated with the upcoming NVIDIA Blackwell platform, promising substantial performance gains through technological innovations, including a second-generation Transformer Engine and advanced NVLink capabilities.
Perplexity's strategic use of NVIDIA's inference stack underscores the potential for AI-powered platforms to manage vast query volumes efficiently, delivering high-quality user experiences while maintaining cost-effectiveness.
Image source: Shutterstock