Amazon Web Services officially launched Amazon Redshift RG instances on May 12, 2026, delivering a new generation of Graviton-powered provisioned clusters to its cloud data warehouse. The announcement marks the first time AWS has extended its custom Arm-based silicon to Redshift’s provisioned compute, promising up to 2.2 times the performance of existing RA3 instances at no additional on-demand cost.

What Are Amazon Redshift RG Instances?

Redshift RG—short for Redshift Graviton—is a new instance family designed from the ground up for AWS’s Graviton3 processors. Unlike the x86-based RA3 nodes that have dominated the Redshift lineup for years, RG runs on 64-bit Arm Neoverse V1 cores, each cluster node packing up to 64 physical cores with high memory bandwidth and custom networking acceleration. The architecture is tailored to handle the sort of large-scale joins, aggregations, and window functions that typify enterprise analytics workloads.

AWS offers two initial RG node sizes: rg4xlarge (with 4 vCPUs, 32 GiB of memory, and 15 TB of attached NVMe SSD cache) and rg16xlarge (with 16 vCPUs, 128 GiB of memory, and 60 TB of SSD cache). Both run Amazon Redshift’s distributed query engine unchanged from RA3, so existing SQL, stored procedures, and ETL scripts work without modification—provided that any compiled UDFs or extensions are recompiled for the Arm instruction set.

Performance Leap: Graviton3 Under the Hood

The headline figure—2.2× performance over RA3—comes from AWS’s internal runs of the TPC-DS decision-support benchmark. RG instances complete typical star-schema queries and complex aggregations in less than half the time, even when both instance types are configured with identical node counts. Much of the gain comes from Graviton3’s wider vector processing units, improved branch prediction, and lower memory latency. For workloads heavy on scan and filter operations, Redshift RG also benefits from AWS’s AQUA (Advanced Query Accelerator) hardware offload cards, which pre-filter data at the storage layer before it hits the CPU.

This isn’t just synthetic benchmark bravado. Early adopters testing on the AWS Big Data discussion forums report that their real-world ETL pipelines—many built around nightly batch jobs pulling from S3 data lakes—saw similar speedups, with some jobs finishing in 45% of the original runtime. One user noted a 1.9× improvement on a particularly gnarly 22-table join used for daily financial consolidation.

Cost Impact: Same Price, Half the Nodes

Amazon has kept on-demand pricing for RG nodes nearly identical to the RA3 equivalents. In the US East (Northern Virginia) region, for instance, rg4xlarge nodes list at $1.086 per hour, while rg16xlarge list at $4.346 per hour—mirroring their RA3 counterparts. The difference lies in throughput: because a single RG node can handle more queries in the same timeframe, you can often downsize your cluster by 50% and still meet SLA targets.

Consider a typical RA3 cluster of four ra3.4xlarge nodes that processes 200 concurrent user queries during peak hours. At the same workload, a cluster with only two rg4xlarge nodes could deliver comparable latency, effectively halving the monthly bill from roughly $6,260 to $3,130. For larger deployments, the savings scale dramatically. AWS has also confirmed that all existing RA3 reserved instances (RIs) can be converted to RG commitments at no penalty, preserving any negotiated discounts through the remainder of the contract term.

To smooth the transition, AWS is offering a free 30-day extended trial for any production RG cluster launched before July 15, 2026, with at least four nodes. The trial includes full migration assessment via the Redshift Console.

Unified Lake Queries: Data Warehouse and Data Lake, Finally Together

The most strategic shift in Redshift RG is how it handles external data. Redshift has long supported querying S3-based data lakes through Redshift Spectrum, but that capability felt bolted on—queries on external tables were slower and subject to separate capacity limits. With RG, AWS has moved lake query processing directly into the cluster’s compute layer. The query planner now treats external partitions identically to native tables, splitting workload across the same pool of Graviton cores and removing the need to copy data into Redshift-managed storage first.

For data engineers, this means you can now create views that join internal dimension tables (hosted on Redshift’s managed storage) with massive fact tables sitting in S3 as Parquet sets, without manually syncing partitions or worrying about Spectrum’s node count constraints. The result is a true “zero-ETL” architecture for many reporting scenarios. Coupled with AWS Lake Formation’s fine-grained access controls, organizations can enforce column-level security across both warehouse and lake data from a single glass pane.

Windows-based BI tools such as Power BI, SQL Server Reporting Services, and even legacy Excel pivot tables connect to Redshift via the standard ODBC or ADO.NET drivers. Under RG, queries that reference S3-backed external tables now return results faster and more consistently, cutting dashboard refresh times and reducing “query timeout” incidents during Monday-morning reporting sprints.

Implications for Windows-Centric Shops

Many enterprise analytics pipelines still run on Windows Server virtual machines, scheduled through PowerShell scripts or custom .NET services that push data into Redshift. For these workloads, the move to Graviton is transparent on the client side—the wire protocol remains ODBC-compatible, and AWS has updated its .NET driver to handle Arm-specific response optimizations without requiring any client-side recompilation.

Windows administrators also benefit from Redshift RG’s improved Workload Management (Auto WLM). The service’s query prioritization engine now takes full advantage of Graviton3’s larger per-core caches, reducing queue wait times for short interactive queries even when heavy ETL jobs are consuming cluster resources. In internal AWS tests, mixed-workload environments saw a 40% drop in spot-query latency compared to identically configured RA3 clusters running the same WLM rules.

Migration Path: From RA3 to RG

Switching an existing RA3 cluster to RG can be done in-place from the AWS Management Console or CLI. Simply modify the cluster’s node type, choose an RG instance, and apply the change. Redshift will handle the node replacement during a maintenance window or immediate reboot, depending on your settings.

Pre-Migration Checklist

  • Compiled UDFs: Any Python UDFs that rely on native libraries (such as NumPy compiled for x86) need to be recompiled for Arm. AWS ships a compatibility scanner in the Redshift Console that flags non-portable UDFs and links to updated packages from the Redshift Marketplace.
  • Reserved Instance Conversions: Use the AWS Reserved Instance Marketplace tool to trade RA3 RIs for RG equivalents. AWS has waived the conversion fee until December 31, 2026.
  • Concurrency Scaling Warm-up: Concurrency scaling clusters spun up on RG can take 90–120 seconds longer to reach full throughput than on RA3, a known behavior attributed to Arm-specific just-in-time compilation in the query engine. AWS states it is working on a firmware update to narrow that gap.
  • Testing: Leverage the built-in Query Profiler to compare RA3 vs. RG runs on a snapshot of your production workload before cutting over.

What’s Different from Redshift Serverless?

Redshift RG is a provisioned offering, meaning you choose and pay for a fixed number of nodes. In contrast, Amazon Redshift Serverless automatically scales capacity up and down, charging only for the data scanned. However, AWS has made sure the two services share a common metadata catalog and query interface. A common pattern emerging from early RG adopters is to use a small RG cluster for predictable, steady-state loads (where reserved pricing delivers the biggest discount) and burst to Serverless for unpredictable spikes, all within the same Lake Formation-managed data perimeter.

The Road Ahead for Amazon Redshift

Redshift RG represents the fullest expression of AWS’s “custom silicon everywhere” philosophy. It builds on the RA3 architecture’s separation of compute and storage, adds a processor tuned specifically for analytic queries, and pulls data lake querying out of its silo. While AWS will continue to support RA3 instances for the foreseeable future, the innovation pipeline is clearly Graviton-first.

For Windows-focused teams, this means a chance to modernize their analytics stack without a painful migration to a new platform. The same .NET applications, SQL Server Reporting Services instances, and Power BI datasets will simply run faster and cost less—provided you take the time to revalidate any client-side UDFs and accept the slightly longer spin-up times for concurrency scaling.

As the data warehouse industry continues its long march toward lakehouse architectures, Redshift RG shows that mature, provisioned services can still deliver knockout price-performance. The only question now is how quickly competitors like Snowflake and Google BigQuery will counter with their own Arm- or hardware-accelerated offerings.