Tech • IA • Crypto
14 articles analyzed by AI / 242 total
Meta has outsourced a critical compute layer of its AI infrastructure to AWS, leveraging AWS's cloud capabilities for scalable AI workload management. This strategic decision allows Meta to efficiently scale AI training and inference workloads without the overhead of maintaining these compute layers in-house, improving operational flexibility.
Meta is exploring innovative technologies to improve the efficiency and power delivery of its AI data centers, targeting better support for large-scale AI workloads. This includes enhancements in hardware utilization and cooling to optimize performance and cost at scale.
Featherless.ai raised $20 million in Series A funding from AMD and Airbus to build open-source AI infrastructure aimed at accelerating accessible and cost-effective AI hardware and software ecosystems. This initiative focuses on creating scalable AI platforms that democratize access to cutting-edge AI tooling.
This study introduces a privacy-preserving federated learning approach combining differential privacy and homomorphic encryption for sensitive healthcare models. It enables collaboration across distributed datasets without sharing raw patient data, facilitating more secure AI model training in regulated environments.
Strait is an ML inference serving system designed to manage task prioritization and reduce interference in GPU-based concurrent deep neural network inference. It improves latency predictability and resource scheduling, enhancing throughput and reducing tail latency in production AI systems.
This work integrates large language model-guided runtime parameter optimization to significantly reduce energy consumption during LLM inference in production pipelines. The approach adapts model behavior dynamically, achieving improved energy efficiency without compromising inference latency.
FedHarmony tackles heterogeneous label correlations in federated multi-label learning, harmonizing label relationships across clients under privacy constraints. This method improves the accuracy and utility of decentralized multi-label AI models without combining raw data, important for distributed AI production systems.