Tag
Ray 2.56 has been released with improvements to Ray Data, Ray Serve for LLMs, GPU-domain-aware placement groups, and Kubernetes integration.
This paper presents a programmable probabilistic computer with one million p-bits by networking FPGAs, achieving Gibbs sampling at over a trillion flips per second for Ising models while introducing a design rule for scaling beyond single-chip limits.
A personal account of hunting for million-digit prime numbers using a home computer setup.
A discussion explores whether AI training could be decentralized like Bitcoin mining, with participants contributing GPU resources to train open-source models in exchange for tokens, raising questions about verification, fake gradients, and efficiency.
A retrospective on the eight fallacies of distributed computing, originally formulated by Sun Microsystems engineers, examining their continued relevance 21 years later for network operators and developers.
Microsoft AI announces MAI-Thinking-1, a 35B active/1T total MoE reasoning model competitive on STEM and coding tasks, developed using Ray for distributed training and orchestration.
A blog post summarizing a panel at the Dutch-Belgian DataBase Day, which discussed practical data systems problems often ignored by academic research, such as variable-length string handling, unrealistic benchmarks, and networking challenges.
Snowflake now supports job-based batch inference powered by Ray, enabling distributed GPU execution for scaling model inference over millions of unstructured datapoints with a single API call.
A Raspberry Pi Zero 2 W submerged in paraffin oil for a year shows a 5.97% performance improvement over an identical uncooled unit while computing asteroid data for BOINC.
Nvidia is backing Span's initiative to deploy residential mini-data centers that leverage underutilized home electricity to run distributed AI workloads. The concept aims to bypass grid constraints by placing GPU nodes beside houses, though it remains largely unproven in real-world deployments.
An inquiry into the practical value of consumer-grade hardware for AI tasks such as inference, fine-tuning, and synthetic data generation, questioning whether local setups offer genuine contributions beyond privacy.
Anyscale is hosting a hands-on virtual lab session teaching developers how to build and scale data pipelines with Ray, covering video data curation, distributed GPU inference, and CPU/GPU streaming pipelines.
This paper introduces enhancements to the AgentScope platform, featuring an actor-based distributed mechanism and flexible environment support to enable scalable, efficient, and user-friendly very large-scale multi-agent simulations.
LakeSail releases Sail, a Rust-native rewrite of Apache Spark that achieves 8x speed-ups and 94% lower infrastructure costs while maintaining full API compatibility.