Using vectors in Java

Using vectors in Java or how I learned to stop worrying and love the dark silicon apocalypse

Computation is increasingly constrained by power. With each advancement in the manufacturing process, a decreasing percentage of the CPU can operate at full capacity, leading to the emergence of the term 'dark silicon'. This trend necessitates techniques that utilize chip area to optimize power efficiency through specialized accelerators.

The presentation will outline key concepts that led to the dark silicon such as Moore’s law and breakdown of Dennard scaling, followed by an overview of current and upcoming CPU accelerators. The focus will then shift to vector units and the specifics of vector programming. Attendees will be introduced to registers, a range of vector operations, and methods to develop branchless algorithms such as sorting networks. The session will conclude with an overview of the new Java Vector API and how it was already picked up by projects to do AI inference (Llama 2) and vector search (AstraDB and Cassandra).

Tomasz Kowalczewski

Veteran of many failed paradigms, I worked in the industry long enough to remember when microservices was a new thing. My interests and talks focused on reactive programming, garbage collection, and performance engineering. I worked on high traffic systems (airline and ad industries), databases and system monitoring.

Right now working at VoltDB, a database that originated from a project by Michael Stonebraker and Andy Pavlo. Its in-memory architecture with efficient, shared nothing, thread per core design provides a great testing ground for various theories and architectural concepts.