Oracle's Exadata platform has always been synonymous with extreme database performance. But with the release of Exadata System Software 24ai and 25ai, alongside the debut of Oracle Exadata X11M in January 2025, Oracle has taken a decisive step into the AI era. Exadata is no longer just a high-performance OLTP and analytics machine — it is now purpose-built to accelerate AI vector search, in-database machine learning, and mixed enterprise workloads, all on a single converged platform.
This post breaks down the key new features across software, infrastructure, high availability, monitoring, and security — and explains why they matter.
What Is Exadata AI Storage?
At its core, Exadata AI Storage refers to Oracle's strategy of pushing intelligence deeper into the storage layer. Rather than offloading AI computation to separate, purpose-built vector databases or GPUs, Oracle brings AI operations — particularly AI Vector Search — directly to the storage servers themselves. This means vector index builds, similarity searches, and distance calculations happen closer to where data lives, dramatically reducing the data movement that kills performance in distributed architectures.
The result? Key vector search operations running up to 30x faster with Exadata System Software 24ai, and further accelerated on X11M with in-memory vector index (HNSW) scans running up to 43% faster on database servers and up to 55% faster on storage servers compared to the previous generation.
Key New Software Features
1. AI Smart Scan
AI Smart Scan is an extension of Exadata's legendary Smart Scan technology, now purpose-built for AI workloads. It offloads compute-intensive AI Vector Search operations — including vector index builds and similarity queries — directly to Exadata's intelligent storage servers. This eliminates the need to ship raw data up to the database tier for processing.
Critically, AI Smart Scan enables thousands of concurrent AI vector searches in multi-user environments. This is a significant differentiator for enterprise RAG (Retrieval Augmented Generation) pipelines and AI applications that need to serve many users simultaneously, not just batch processes.
With the latest release (Exadata System Software 25ai / 25.1), Adaptive Top-K Filtering further extends this: each storage server maintains a running Top-K result set, reducing data returned to the database servers by up to 4.7x. Similarly, VECTOR_DISTANCE() calculations are now projected from storage, delivering up to 4.6x faster distance-based queries.
2. Exadata RDMA Memory (XRMEM)
XRMEM replaces the persistent memory (Intel Optane PMem) that earlier generations used, adapting to changes in the memory vendor landscape. It is built on DDR5 DRAM and accessed via RDMA (Remote Direct Memory Access) over Converged Ethernet (RoCE), bypassing the OS, I/O, and network software stacks entirely.
The practical benefit: ultra-low read latency as low as 14 microseconds — a 21% improvement over prior generations — with scan throughput of up to 500 GB/s from XRMEM alone. Each Exadata X11M Extreme Flash Storage Server contains 1.25 TB of XRMEM, which sits as an acceleration tier in front of the Smart Flash Cache.
XRMEM is particularly impactful for OLTP workloads that require sub-20-microsecond response times, and it now also accelerates AI vector index reads transparently.
3. On-Storage Processing
This is the foundational principle that makes all the above possible. Exadata storage servers are not passive disk arrays — they are intelligent compute nodes in their own right. SQL filtering, column projection, decompression, encryption/decryption, bloom filters, and now AI vector operations all execute on the storage servers, not the database servers.
This dramatically reduces the volume of data sent over the internal network and processed by database-tier CPUs. For analytics workloads, this "push-down" processing model is why Exadata consistently delivers 10–100x better throughput than general-purpose storage.
4. In-Memory Columnar Speeds for JSON Queries
Exadata's In-Memory Columnar Cache on storage servers (also called Columnar Cache) stores a columnar representation of row-oriented data directly in flash cache. When queries access this data, they get the performance of columnar analytics without requiring data to be reformatted or migrated.
With Oracle Database 23ai — which is required to unlock the full Exadata Exascale feature set — JSON documents stored natively benefit from this columnar acceleration. Oracle Database 23ai's JSON Relational Duality views, which expose the same data as both JSON and relational tables simultaneously, can be queried at columnar memory speeds on Exadata, collapsing the performance gap between document and relational workloads.
5. Transparent Cross-Tier Scan
Exadata's multi-tier storage hierarchy — XRMEM → Smart Flash Cache → disk — is managed automatically and transparently. When a Smart Scan or AI Smart Scan runs, Exadata intelligently sources data from whichever tier contains it, combining reads from memory, flash, and disk in parallel.
This means administrators and developers never need to manually partition hot vs. cold data or tune tier placement. The system continuously tracks access patterns and moves data to the appropriate tier based on usage, keeping the hottest data in XRMEM and the next hottest in flash. The database never sees these tiers explicitly — it simply issues a SQL or vector query and receives results.
6. Caching Enhancements
Several caching improvements ship with the latest Exadata System Software releases:
-
Automatic KEEP Object Load into Exadata Flash Cache: Objects tagged with the
KEEPstorage clause in Oracle Database are automatically and proactively loaded into Exadata Smart Flash Cache — even before they are first accessed — ensuring zero cold-start latency for critical tables and indexes. - Write Back Flash Cache: Database block writes are cached in flash, eliminating disk I/O bottlenecks for large OLTP and batch workloads.
- Cell-to-Cell Rebalance preserving XRMEM and Flash Cache: When data rebalances across storage servers (due to maintenance or failure), both the XRMEM and flash cache contents are also rebalanced, preserving performance levels rather than causing a cold-cache performance dip.
7. Columnar Smart Scan at Memory Speed
When Oracle Database In-Memory is enabled, Exadata automatically stores data in columnar format within Flash Cache and XRMEM if it will improve query performance. This brings columnar analytics performance — historically associated only with in-database memory (DRAM on database servers) — to storage-resident data, enabling analytics at memory speeds even when datasets exceed what fits in database-server DRAM.
A single Exadata X11M rack can deliver up to 100 GB/s of flash throughput and 500 GB/s from XRMEM for the hottest data, far exceeding what traditional storage arrays can achieve even with flash added.
8. Exadata Cache Observability
Exadata System Software 24.1 introduced ecstat (Exadata Storage Cache Statistics), a real-time utility that provides per-storage-server statistics on Smart Flash Cache usage, XRMEM hits, and I/O performance. This was a long-standing gap — DBAs previously had to rely on AWR snapshots to understand cache behavior.
In Exadata System Software 25.2, this was extended with CellSQLStat, which provides real-time, per-storage-server insights into active Smart Scan operations: CPU and memory usage, Storage Index and Columnar Cache I/O savings, flash and XRMEM hit rates, scan rates, and more. Both ecstat and CellSQLStat data are automatically included in ExaWatcher collections, making them available for historical analysis.
Infrastructure Improvements
Increased Number of Virtual Machines
With Exadata Exascale (the new intelligent storage architecture introduced in 2024 and available on X11M), the limit on Virtual Machine clusters per database server increases dramatically. Traditional Exadata with ASM supported 4, 8, or 12 VMs per database server. Exascale raises this ceiling to 50 VMs per database server, enabling far greater consolidation of Oracle Database workloads on a single Exadata system without sacrificing isolation or performance.
Exascale also centralizes VM storage in the shared Exascale storage pool rather than on individual database servers, increasing flexibility and simplifying management.
Secure Boot for KVM Virtual Machines
Exadata System Software now supports Secure Boot for KVM guest VMs, ensuring that only cryptographically signed and trusted OS images can boot on Exadata database servers. This closes a significant attack vector in virtualized deployments and aligns Exadata's on-premises security posture with cloud-native security standards. It complements existing features like Trusted Partitions for Oracle Linux Virtualization.
High Availability and Network Resilience
Improved RoCE Network Resilience: ExaPortMon
Every Exadata database and storage server connects to the internal network via dual 100 GbE RoCE ports for an aggregate of 200 Gbps. If a RoCE leaf switch port becomes stalled — appearing online but unable to pass traffic — it can cause cluster instability without triggering a clean failure.
ExaPortMon (introduced in Exadata System Software 24ai) solves this. It continuously monitors both RoCE ports on every server. When it detects a stalled port, it automatically migrates the IP address to the healthy port, keeping network traffic flowing and preventing outages. When the stalled port recovers, ExaPortMon automatically returns the IP address to its original port. No manual intervention required.
Enhanced RoCE Network Security: Exadata Secure RDMA Fabric Isolation
Exadata Secure RDMA Fabric Isolation (Secure Fabric) provides strict network isolation between VM clusters sharing the same physical Exadata infrastructure. It prevents database servers in one VM cluster from communicating with those in another over the RoCE fabric, eliminating lateral movement risk in consolidated and multi-tenant deployments.
Starting with Exadata System Software 25.1, Secure Fabric is automatically selected by default for all new on-premises deployments with X8M and newer hardware — bringing on-premises deployments into alignment with cloud deployments, which have always used Secure Fabric.
Monitoring and Management
AWR & SQL Monitor Enhancements
Oracle Automatic Workload Repository (AWR) on Exadata is enhanced with Exadata-specific storage server metrics alongside standard Oracle wait event data. AWR now collects and reports on XRMEM, Flash Cache, and HDD device performance, enabling DBAs to correlate database wait events with storage-tier behavior in a single report.
SQL Monitor is similarly enhanced, providing end-to-end visibility into query execution that includes storage offload statistics, Smart Scan I/O savings, and flash cache hit rates — all tied to the specific SQL statement being analyzed.
JSON API for Management Server
Exadata's Management Server (MS) now exposes a JSON REST API, enabling programmatic access to Exadata management and monitoring functions. This is a significant modernization of Exadata's management interface, making it easier to integrate Exadata health metrics, alerts, and configuration into modern observability stacks (Grafana, custom dashboards, CI/CD pipelines) without relying solely on traditional CLI tools like cellcli or Enterprise Manager.
Security Enhancements
SNMP v3 Security
Exadata System Software 24.1 introduced mandatory SNMP security improvements across all storage servers and database servers. The key changes:
- SNMP v3 is now the recommended and encouraged standard, supporting SHA-256, SHA-384, and SHA-512 authentication protocols for strong authentication and encryption.
- All SNMP subscriber definitions now require the connection type to be explicitly specified — administrators can no longer leave it ambiguous.
- SNMP v1 remains available but triggers an explicit warning discouraging its use.
- Default community strings like
publicandprivateare actively discouraged by the system, prompting administrators to set strong, unique values.
This tightens Exadata's management plane security, closing a common vulnerability in enterprise infrastructure where SNMP v1 with default community strings is still widely used.
Why This All Matters: The Convergence Thesis
The strategic bet Oracle is making with Exadata AI Storage is one of convergence over fragmentation. The enterprise AI market has seen an explosion of purpose-built vector databases (Pinecone, Weaviate, Qdrant, Milvus, Chroma), and many organizations are building RAG pipelines that shuffle data between separate systems: a relational database for operational data, a vector database for embeddings, an object store for documents.
Exadata offers a fundamentally different architecture: bring the AI to the data, not the data to the AI. With Oracle AI Database 23ai (now succeeded by Oracle AI Database 26ai as the long-term support release), all of this runs in a single converged engine — relational queries, vector similarity search, JSON document queries, graph traversals, and full-text search — executed as optimized SQL on Exadata hardware. Advanced AI features including AI Vector Search are included at no additional charge.
And with Exadata Exascale reducing the entry cost for Exadata Database Service by up to 95% and enabling organizations to start with as little as 300 GB of storage, the platform is no longer exclusively for Fortune 500 database estates. It is increasingly accessible to organizations of any size that need to build AI applications on enterprise-grade, governed, transactionally consistent data.
Summary Table
| Category | Key Feature | Benefit |
|---|---|---|
| AI Workloads | AI Smart Scan + Adaptive Top-K | Up to 30x faster vector search; 4.7x less data to DB servers |
| Memory | XRMEM (DDR5 + RDMA) | 14µs read latency; 500 GB/s scan throughput |
| Caching | Auto KEEP Load, Write Back, Columnar Cache | Zero cold-start for critical objects; analytics at memory speed |
| Observability | ecstat + CellSQLStat | Real-time per-cell Smart Scan and cache monitoring |
| Infrastructure | Exascale VM limit increase | Up to 50 VMs per DB server (up from 12) |
| Security | Secure Fabric default on-prem | Automatic lateral-movement isolation for VM clusters |
| Network HA | ExaPortMon | Auto-failover between RoCE ports; no manual intervention |
| Security | SNMP v3 enforcement | SHA-512 auth; eliminates default community string risk |
| Management | JSON API for Management Server | Programmatic integration with modern observability stacks |
Exadata AI Storage represents Oracle's clearest articulation yet of its "converged data" strategy: a single platform that handles OLTP, analytics, and AI workloads without requiring organizations to build and manage a fragmented ecosystem of specialized tools. With Exadata System Software 25ai, the X11M generation, and the Exascale architecture now generally available across OCI, multicloud (AWS, Azure, Google Cloud), and on-premises, there has never been a better time to evaluate what Exadata can do for your AI application stack.
The numbers speak for themselves — but the architecture is the real story.
Have you worked with Exadata AI Storage or Oracle Database 23ai/26ai AI Vector Search? Share your experience in the comments.
Top comments (0)