AI / ML

AMD has acquired AI chip startup Taalas to integrate model-specific integrated circuits directly into silicon. Early demonstrations indicate these custom chips can achieve inference speeds of up to 17,000 tokens per second. This move aims to significantly boost performance for specific AI workloads by hardcoding model architectures.

• AMD targets inference speed gains by integrating Taalas' custom silicon technology.
• Early demos show model-specific chips reaching 17,000 tokens per second throughput.
• Hardcoding models into hardware reduces latency compared to general-purpose accelerators.
• Acquisition signals AMD's push into specialized AI inference hardware solutions.

Amazon RDS for Oracle now supports Oracle Machine Learning for SQL (OML4SQL), enabling practitioners to build, train, and score models entirely within the SQL engine. The new AutoML feature automatically selects the optimal algorithm without requiring data movement or external ML platforms. This capability allows credit risk and similar workloads to be processed directly on the database layer.

• Eliminates data movement to external ML platforms by keeping training and scoring in-database.
• AutoML automatically selects the best algorithm, reducing manual model tuning effort.
• Enables SQL-only workflows for machine learning on managed RDS for Oracle instances.

Amazon DynamoDB now supports native vector similarity search, removing the need to maintain a separate vector database for applications already using DynamoDB for operational data. The feature allows developers to integrate vector search directly into their existing DynamoDB tables, as demonstrated with a semantic search application over research paper abstracts using Amazon Bedrock embeddings. The blog post also details how vector search operations are metered and billed within the DynamoDB service.

• No longer need a separate vector database for semantic search alongside DynamoDB.
• Vector search is now a native feature within DynamoDB tables.
• Metering for vector search operations is distinct and detailed in the post.
• Can integrate with Amazon Bedrock embeddings for building semantic search apps.