Accelerating Large Language Model Inference with NVIDIA in the Cloud
PerplexityNVIDIA TensorRT-LLM · NVIDIA A100 Tensor Core GPUs · NVIDIA H100 Tensor Core GPUs +2
Perplexity built pplx-api, an API for developers to integrate open-source LLMs with fast inference, served on Amazon EC2 P4d instances powered by NVIDIA A100 Tensor Core GPUs and accelerated with NVIDIA TensorRT-LLM (with a planned move to Amazon P5 instances with NVIDIA H100 GPUs). pplx-api achieves up to 3.1X lower latency and up to 4.3X lower first-token latency versus other deployment platforms, and switching external inference-serving API references to pplx-api lowered costs 4X, saving $600,000 per year. Using NVIDIA H100 GPUs and FP8 precision on Amazon P5 instances cuts latency in half and boosts throughput by 200 percent versus A100 GPUs in the same configuration. Perplexity also uses AWS's Kubernetes integration to scale elastically beyond hundreds of GPUs.
Overview
Perplexity built pplx-api, an API for developers to integrate open-source LLMs with fast inference, served on Amazon EC2 P4d instances powered by NVIDIA A100 Tensor Core GPUs and accelerated with NVIDIA TensorRT-LLM (with a planned move to Amazon P5 instances with NVIDIA H100 GPUs). pplx-api achieves up to 3.1X lower latency and up to 4.3X lower first-token latency versus other deployment platforms, and switching external inference-serving API references to pplx-api lowered costs 4X, saving $600,000 per year. Using NVIDIA H100 GPUs and FP8 precision on Amazon P5 instances cuts latency in half and boosts throughput by 200 percent versus A100 GPUs in the same configuration. Perplexity also uses AWS's Kubernetes integration to scale elastically beyond hundreds of GPUs.
The challenge
Delivering fast and efficient LLM inference is critical for real-time applications. As a startup, Perplexity faced escalating costs associated with LLM inference to support its rapid growth, while needing to maintain strict service-level agreement requirements and adapt quickly to an explosively growing ecosystem of community LLMs.
The solution
Perplexity built pplx-api, an API for developers to integrate open-source LLMs with fast inference, served on Amazon EC2 P4d instances powered by NVIDIA A100 Tensor Core GPUs and accelerated with NVIDIA TensorRT-LLM, with a planned full transition to Amazon P5 instances powered by NVIDIA H100 Tensor Core GPUs. Perplexity uses AWS's integration with Kubernetes to scale elastically beyond hundreds of GPUs.
Reported business value
pplx-api achieves up to 3.1X lower latency and up to 4.3X lower first-token latency relative to other deployment platforms. Switching external inference-serving API references to pplx-api lowered costs 4X, saving $600,000 per year. Using NVIDIA H100 GPUs and FP8 precision on Amazon P5 instances cuts latency in half and boosts throughput by 200 percent compared to NVIDIA A100 GPUs in the same configuration.
Sources
Open any source and check the claim yourself — that is the point of the register.
This record was researched and written with AI assistance, and its claims were checked against the sources above. (EU AI Act art. 50 transparency notice.)
Other technology & software entries in the register.
HP crafts marketing campaigns that resonate with customers using Databricks and Uniphore
HP centralized first-party customer data on the Databricks Data + AI Platform with Delta Lake and Unity Catalog, and connected it to Uniphore's HybridCompute for federated query pushdown, cutting campaign setup from 2 weeks to 2 hours and processing 400 million records in seconds.
Building a safer and more sustainable world
Novade, a construction management software company, partnered with Databricks to modernize its analytics and ML infrastructure, moving from a homegrown Apache Airflow stack to the Databricks Data + AI Platform with Delta Lake, Unity Catalog and MLflow. The move reduced total cost of ownership by 60% and supported a 100% increase in new clients, while enabling ML models for predicting incident risk affecting worker safety and project delivery schedules.
SafeGraph optimizes geospatial data processing with Databricks and Delta Sharing
SafeGraph used the Databricks Data + AI Platform, Delta Lake, Delta Sharing and MLflow to process petabytes of geospatial data and feed predictive models, achieving 2x-10x faster spatial querying, indexing and partitioning and a 50% reduction in peak memory consumption compared to other platforms, while reducing data access time for partners from months to minutes via Delta Sharing.
Transforming Weather Forecasting with Lakeflow Jobs
AccuWeather migrated from on-premises infrastructure to Databricks and Lakeflow Jobs, working with Datadog for observability, to unify diverse weather data formats and orchestrate 4,500+ weekly jobs. Lakeflow Jobs coordinates the ingestion of multiple weather models, triggers machine learning processes that weight and blend different forecasts, and manages complex job dependencies for reinforcement training workflows used in AccuWeather's proprietary forecasting engine. AccuWeather reports 3x faster dataset development (three months to one month per dataset), a 50% reduction in unactionable alerts, and 50% cost savings on serverless job usage.
AccuWeatherLakeflow Jobs · Delta Sharing · DatadogWas this helpful?
Your feedback helps us improve our use case database