AWS Pushes Graviton into Redshift Analytics Stack

extract

AWS Pushes Graviton into Redshift Analytics Stack New Redshift instances combine warehouse and data lake analytics as AWS pushes custom silicon higher into AI-era data infrastructure. Amazon Web Services this week launched new AWS Graviton-powered RG instances for Amazon Redshift. The release combines warehouse and data lake analytics in a single query engine as cloud providers try to cut AI-era infrastructure costs and reduce data movement across sprawling analytics environments. AWS said Redshift RG instances deliver up to 2.2x faster data warehouse performance than previous-generation RA3 systems at a 30% lower cost per vCPU. The company also said the integrated data lake query engine delivers up to 2.4x faster Apache Iceberg performance and 1.5x faster Apache Parquet performance compared to RA3. The launch extends AWS’s broader effort to push Graviton custom Arm processors beyond general-purpose compute and into higher-level analytics and AI infrastructure services. AWS said the integrated engine lets Redshift query warehouse and data lake data from the same compute layer without relying on separate Redshift Spectrum infrastructure or per-terabyte scan charges. That matters as enterprises increasingly try to unify fragmented analytics stacks built around cloud warehouses, object storage, and open table formats such as Apache Iceberg. AWS positioned the new instances for “analytics and agentic AI workloads” that require lower-latency access to large datasets. AWS Pulls Custom Silicon Higher into the Analytics Layer Matt Kimball, vice president and principal analyst for datacenter technologies at Moor Insights & Strategy, said the launch is more than a routine instance refresh. “AWS is addressing significant challenges that are of concern to every enterprise – performance, TCO, and complexity,” Kimball told Data Center Knowledge. Kimball said the performance and pricing gains AWS disclosed make the platform shift difficult to dismiss as incremental tuning. “When looking at numbers like 2.2x faster for data warehouse workloads and 2.4x for Apache Iceberg alongside a 30% lower price per vCPU, it’s hard to see this as incremental in that cost/performance metric,” he said. He also pointed to AWS eliminating separate Spectrum scanning infrastructure and related fees as a meaningful operational change for large analytics deployments. “Zooming out a little, AWS does seem to be pulling its silicon up the stack,” Kimball said. “Not just into infra like EC2, but into the data and analytics layer itself.” He noted that the move points toward a longer-term architectural direction where hyperscalers increasingly fuse custom silicon with managed analytics services. Lakehouse Pressure Builds The rollout also reflects growing competition around “lakehouse” architectures that blur traditional boundaries between structured data warehouses and large-scale data lakes. AWS said RG instances process Iceberg and Parquet data directly on cluster nodes via a built-in vectorized query engine, rather than routing scans through a separate fleet. The company also highlighted local NVMe caching designed to keep frequently accessed datasets closer to compute resources. Kimball said the integrated lake query engine addresses both operational complexity and the growing cost of moving and duplicating enterprise data. “As an enterprise IT leader, I want to reduce how much I have to duplicate and move data,” Kimball said. “And I want to avoid having all of these fragmented data pipelines.” He said those inefficiencies become more damaging as enterprises deploy retrieval systems and agentic AI workloads that depend on rapid access to structured and semi-structured data. “Keeping analytics closer to where data already lives, while maintaining consistent governance, is becoming a core design principle,” Kimball said. For enterprises already standardized on Amazon S3 and open formats like Iceberg and Parquet, tighter integration between warehouses and lakes lowers consolidation barriers and simplifies analytics operations, he added. AI Infrastructure Economics Spread Beyond GPUs The Redshift launch also underscores how AI infrastructure pressure is spreading beyond GPU clusters into the surrounding analytics ecosystem. As enterprises build retrieval pipelines, inference systems, and agentic AI environments, demand for lower-latency access to large datasets has intensified. That is forcing cloud providers to reduce query overhead, collapse analytics layers, and limit costly transfers between warehouse and lake environments. Kimball said hyperscaler custom silicon is becoming an increasingly important competitive differentiator across analytics and AI infrastructure. “Undoubtedly, yes,” he said. “AWS has been leading on this front – Graviton, Trainium, Inferentia, Nitro.” He argued custom silicon increasingly enables architectural optimizations that extend beyond raw infrastructure economics. “With Graviton, this isn’t just about cost,” Kimball said. “It enables the vectorized execution engine and NVMe caching that makes the integrated lake query even viable at the levels AWS claims.” Kimball also warned that tighter coupling between custom silicon and managed cloud services creates deeper long-term dependencies for enterprise customers. “The tighter the integration between silicon and platforms or services, the better the efficiency gains across performance, cost, and operations,” he said. “But the counter to this is that this also creates deeper dependencies.” AWS said existing RA3 clusters can migrate to RG instances using Snapshot & Restore, Elastic Resize, or Classic Resize. The new instances are initially available in multiple US regions, including Northern Virginia, Ohio, Oregon, and Northern California.