Achieving Extreme Efficiency through Specialized GPU Kernel Generation
Summary
Automated GPU kernel generation produced Qwen 3.5 122B kernels that are 1.8–5.2× faster than the top implementations available in vLLM. Achieving these efficiency gains requires pairing unconstrained agent exploration with a strict outer verification system to catch measurement bugs and validate real-world speedups.
Summary generated by brickster.ai. For the full article, follow the source link above.
More from Databricks Blog
MemCurrent: Designing Persistent Memory for Production Agents
MemCurrent is a persistent memory architecture for production agents, now powering Agent Bricks Memory and Genie, that combines agent-written memory with an offline "dreaming" pipeline that consolidates, deduplicates, and resolves conflicts in a structured knowledge base. On Databricks' internal benchmark built from real Genie conversations, it triples the agent's pass rate versus no memory while using 31% fewer tokens and running 26% faster, outperforming Mem0, PropMem, and OpenClaw memory.
Introducing Agent Memory on Databricks
Agent Memory is now in Beta on Agent Bricks, giving agents a Lakebase-backed store for facts, preferences, and workflows, built on Databricks Research's MemCurrent architecture, which combines plugins, on-demand tools, and an offline "dreaming" pipeline. In Databricks' internal benchmarks on enterprise tasks, it tripled agent pass rate while using 31% fewer tokens, and the Agent Bricks CLI can deploy an agent with memory and sessions in minutes.
How Hamlet uses Databricks Lakebase Postgres to test changes against production data
Hamlet gives every pull request its own Lakebase Postgres branch of production data, so engineers can reproduce and test fixes against the exact data that caused a bug without touching production. AI coding agents run the pipelines and check their work on each branch before a person approves the merge, and Hamlet is now exploring Lakebase Search, branching for object storage, and a home for longer agentic jobs.
How HP Is Building an AI-Ready CDP with Databricks CustomerLake
HP reduced audience-readiness timelines from months to two hours and cut projected CDP total cost of ownership by up to 40% by implementing Databricks CustomerLake directly on its existing foundation. The deployment unified media, brand, and intent data into a governed Marketing Lakehouse, scaling topic ingestion to approximately 20,000 while slashing data-sharing turnaround from three months to 48 hours.
CustomerLake enters Beta: Leading enterprises going live with the Agentic CDP
CustomerLake's Agentic CDP is now in Beta with Campaign Agents and in Private Preview with Profile Agents, bringing audience building, activation, personalization, and Customer 360 into the governed Databricks Data + AI platform. Early customers report cutting audience turnaround time by 90%, becoming audience-ready in 2 hours instead of 2–3 months, a potential 40% reduction in CDP total cost of ownership, and a 10-point lift in conversion rates.
What makes a catalog open? Unity Catalog Advances the Open Lakehouse
Unity Catalog is the first catalog to support the Apache Iceberg REST catalog spec for REGISTER and UNREGISTER, enabling seamless table movement across catalogs without data duplication. By keeping managed tables in customer-owned storage, it provides full data portability while delivering advanced catalog features directly on tables you own.