Oracle’s AI coverage tends to gravitate toward two things: Database 23ai’s vector search and Oracle’s multicloud database deals. There’s an older, less-hyped piece of Oracle engineering that both of those stories quietly depend on: Exadata, the hardware-software system Oracle has sold for over 15 years, and Autonomous Database, the managed service built on top of it. Exadata is the reason Oracle can claim database performance numbers that a generic cloud VM running Oracle Database on commodity storage can’t match — not through a faster CPU, but through moving part of the database engine itself into the storage layer. Here’s what that actually means, verified against Oracle’s own architecture documentation.
At a Glance
| Thing | What it is |
|---|---|
| Exadata | Engineered system pairing Oracle Database software with purpose-built storage servers and an RDMA network fabric |
| Key mechanism | Smart Scan: pushes SQL filtering and column projection down to storage servers, so only matching rows/columns travel back to the database tier |
| Autonomous Database | Oracle’s managed cloud database service, built on Exadata infrastructure, that automates tuning, patching, scaling, and failover |
| Deployment options | Serverless (OCI public cloud, pay-per-use), Dedicated Infrastructure, Exadata Cloud@Customer (your data center), plus multicloud on AWS/Azure/Google Cloud |
| Pricing model | Serverless is usage-based with a free tier and 30-day trial; Dedicated and Cloud@Customer are quote-based; Oracle publishes no list price for Exadata hardware itself |
| Best for | Oracle Database shops running large OLTP/analytics workloads who want storage-layer acceleration without re-architecting queries |
Short Verdict
Exadata’s core idea — run part of the database engine on the storage hardware itself, and connect everything with RDMA instead of a generic network — is genuinely distinctive among relational database platforms, and it’s been refined since the first Exadata machine shipped in 2008. Autonomous Database wraps that hardware in automation that removes most of the day-to-day DBA work: patching, indexing, backups, and scaling happen without a maintenance window. The catch is the one every Oracle-engineered-system story has: Oracle’s own pages make broad performance claims (“dramatically better performance,” “extreme performance”) without publishing benchmark numbers or prices, so anything beyond the documented mechanics needs a sales conversation to confirm.
What Exadata Actually Is
Exadata is Oracle’s term for a system built from four pieces working together: database servers that run Oracle Database, storage servers that hold the data, a low-latency RDMA network fabric connecting them, and Exadata System Software that runs on both tiers. The point of the design is to target Oracle Real Application Clusters’ (RAC) shared-everything model — many database servers sharing the same storage — without making that sharing the bottleneck.
The detail that separates Exadata from “Oracle Database running on a SAN” is that the storage servers aren’t dumb disks. Each one runs a slice of Oracle Database’s own code and can process query logic locally — filtering rows, projecting columns, even applying some predicates — before anything crosses the network back to a database server. Oracle’s architecture page describes the storage servers’ CPUs as accelerating database-intensive work “similar to how GPUs accelerate graphics,” working alongside the database servers’ CPUs rather than replacing them.
Smart Scan: How the Offload Actually Works
The mechanism behind that offload is called Smart Scan (Oracle also calls it SQL Offload), and it’s the single most load-bearing idea in the Exadata design. Per Oracle’s documentation, Smart Scan kicks in automatically whenever the query optimizer picks a full table scan, a fast full index scan, or a full bitmap index scan — the scan types that would otherwise mean “read everything, then filter it after it arrives.”
Instead, the database server sends the storage servers the column projections and SQL predicates for the query. Each storage server applies those filters locally, in parallel, as it reads from flash or disk, and returns only the rows and columns that actually matched. Oracle’s own example: a query for North American sales orders of at least $500 is evaluated entirely on the storage tier, so only the matching customers’ names and addresses — not the full order table — travel back across the network. Oracle states this can reduce the data volume crossing the wire by orders of magnitude for large scans, and that the optimization still applies to compressed and encrypted tables.
Exadata System Software: What Runs Where
| Component | Runs on | Job |
|---|---|---|
| CELLSRV | Storage servers | Core service handling Smart Scan offload, I/O Resource Manager, and flash/XRMEM tiering |
| IORM | Storage servers | I/O Resource Manager — stops one database from monopolizing shared storage bandwidth |
| MS / RS | Storage servers | Management and automatic service restart |
| CELLOFLSRV | Storage servers | Version-specific offload functions |
| libcell | Database servers | Lets Oracle Database talk to storage using Oracle’s iDB protocol over RDMA |
The networking layer underneath all of this is RDMA (remote direct memory access) over a converged Ethernet fabric (RoCE), carrying a purpose-built Oracle protocol called iDB. RDMA lets one server read or write another server’s memory without routing through either machine’s operating system or network stack — no extra copying, no kernel involvement. The same fabric also carries RAC’s instance-to-instance traffic. Oracle also layers automatic, workload-aware caching on top: data moves between hard disk, flash, and XRMEM (Exadata RDMA Memory) based on access patterns, and because every storage server is reachable from every database server, that cache benefit is shared across the whole cluster rather than pinned to one machine.
Oracle’s newest storage architecture, Exadata Exascale (available from release 24.1), goes a step further by decoupling the database and Grid Infrastructure layers from the storage servers entirely, so customers can scale compute and storage independently — though it can also run alongside the older Automatic Storage Management (ASM) model, and Oracle says a full Exascale cluster needs at least five storage servers to run its own grid services rather than borrowing cycles from the database servers.
From Exadata to Autonomous Database
Exadata is infrastructure; Autonomous Database (Oracle now calls it Autonomous AI Database) is the managed service built on top of it that automates nearly everything a DBA would otherwise do by hand. Oracle’s own product page groups the automation into three themes, even though it doesn’t literally use the old “self-driving / self-securing / self-repairing” marketing labels anymore:
- Automated operation: provisioning, backups, monitoring, alerting, and disaster recovery run without manual intervention; automatic indexing and partitioning tune the schema for the actual workload; compute and storage scale up to 3x base-provisioned resources with no downtime.
- Automated security: patches apply without downtime, Transparent Data Encryption covers data at rest, in motion, and in backups, and Oracle Data Safe plus Database Vault add threat detection, role-based access, and always-on auditing.
- Automated resilience: Oracle Autonomous Data Guard and Application Continuity recover in-flight transactions during outages, backed by the hardware redundancy and RDMA networking Exadata already provides.
Oracle’s default SLA for Autonomous Database is 99.95% uptime without a standby, rising to a claimed 99.995% with Autonomous Data Guard enabled. The service ships in several deployment shapes:
| Deployment | Where it runs | Pricing model |
|---|---|---|
| Serverless | OCI public cloud, fully managed | Pay-per-use; free tier and 30-day trial available |
| Dedicated Infrastructure | OCI public cloud, single-tenant | Quote-based; more control over patch timing |
| Exadata Cloud@Customer | Your own data center | Quote-based; for data residency/latency requirements |
| Multicloud | AWS, Azure, Google Cloud (via Oracle Database@AWS/Azure/GCP) | Quote-based |
Oracle also claims that consolidating multiple databases into shared, elastic Autonomous Database resource pools can cut compute costs by up to 87% — a figure from Oracle’s own marketing that, like the Exadata performance superlatives, comes with no stated benchmark methodology, so it’s worth treating as a best-case scenario rather than a typical result.
Exadata vs. Running Oracle on Generic Cloud Storage
| Dimension | Exadata | Oracle DB on generic block storage |
|---|---|---|
| Where filtering happens | Partly in storage (Smart Scan), reducing data moved to compute | Entirely at the database server; storage just returns raw blocks |
| Network path | RDMA (iDB over RoCE), bypassing OS network stack | Standard storage network protocol (iSCSI, NFS, etc.) |
| I/O isolation across tenants | Enforced by IORM at the storage layer | Depends on the underlying storage platform’s own QoS features |
| Caching | Automatic, workload-aware, tiered across flash/XRMEM/disk, shared cluster-wide | Typically a single storage-side cache tier, not database-aware |
| Who manages patching/tuning | Largely automated by Autonomous Database | Manual, unless the customer builds equivalent automation |
Trade-offs
- Smart Scan only triggers on specific scan types (full table, fast full index, full bitmap index) — index range scans and other access paths don’t get the same offload benefit.
- Oracle’s architecture and pricing pages make strong comparative claims (“dramatically better performance… than other platforms”) without publishing the benchmarks, hardware specs, or prices behind them — you need a direct quote to know what you’re actually paying.
- Exadata’s value is specifically about offloading SQL processing to storage; it doesn’t replace good query design, and non-Exadata workloads (or queries that can’t use Smart Scan’s supported scan types) won’t see the same gains.
- Autonomous Database’s automation is a trade of control for convenience — DBAs who want to hand-tune indexing or control exact patch windows have less to do, by design.
Who It’s a Good Fit For
Organizations already committed to Oracle Database, running large OLTP or analytics workloads where full scans are common, get the clearest benefit from Exadata’s storage-offload model — and Autonomous Database is the easiest way to consume that infrastructure without staffing a dedicated DBA team for patching and tuning. Regulated industries with data-residency requirements have a specific reason to look at Exadata Cloud@Customer, which puts the same hardware and automation inside their own data center.
Bottom Line
Exadata’s hardware-software co-design — storage servers that run database logic, RDMA networking, and workload-aware caching, all coordinated by Exadata System Software — is a genuinely different architecture from “database plus generic storage,” and it predates Oracle’s current AI push by well over a decade. Autonomous Database is the layer that turns that infrastructure into something you consume as a managed service instead of a rack you babysit. The mechanics are well documented; the performance superlatives and the actual bill are not, so budget a real conversation with Oracle before assuming either one.
FAQ
Is Exadata the same thing as Autonomous Database?
No. Exadata is the underlying hardware-software system (database servers, storage servers, RDMA networking). Autonomous Database is a managed cloud service that runs on Exadata infrastructure and automates administration on top of it.
Does Smart Scan work with every query?
No. Oracle’s documentation says it activates when the optimizer chooses a full table scan, fast full index scan, or full bitmap index scan — not every access path qualifies.
Can I run Autonomous Database outside Oracle Cloud?
Yes. Exadata Cloud@Customer brings the same infrastructure into your own data center, and Oracle also offers Autonomous Database through its multicloud deals on AWS, Azure, and Google Cloud.
How much does Exadata cost?
Oracle doesn’t publish a list price for Exadata hardware or for Dedicated/Cloud@Customer deployments on its architecture or product pages — those require a quote. Serverless Autonomous Database is pay-per-use with a published free tier and 30-day trial.
Sources
- Oracle — Exadata Database Machine Architecture
- Oracle — Exadata Smart Scan
- Oracle — Autonomous AI Database
Discover more from Software Online Review
Subscribe to get the latest posts sent to your email.
