ORACLE AI WORLD / LIVE EXPERIENCE
Oracle 26AI · CuVS · GPU INDEX OFFLOAD

Less waiting. More discovery.

One document. Two index builds. See the difference GPU acceleration makes.

CONNECTING TO LIVE SYSTEM
Oracle 26aiPrivate AI ServicesNVIDIA cuVS
01

Your document

PRIVATE SESSION
OR TRY OUR DOCUMENT

Your document becomes searchable first. Then the CPU and GPU builds start together.

DOCUMENT CHECKPOINTS0 / 6
Upload receivedWaiting
Collection createdWaiting
Ingestion submittedWaiting
Vectors storedWaiting
HNSW confirmedWaiting
Search verifiedWaiting
Build pathAwaiting evidence
Source checkAwaiting document
RunNew session
02

The acceleration, measured.

AWAITING DOCUMENT
YOUR UPLOAD + 250K SCALE DATA

Same data.
Different speed.

Your upload + 250K scale vectors.
Fresh GPU and CPU indexes, verified by search.

—×
MEASURED ACCELERATIONYour result appears after both builds finish.
Oracle + NVIDIA cuVSGPU INDEX OFFLOAD
READY TO RUN
—s
ADD UPLOAD —BUILD INDEX —VERIFY SEARCH —
Oracle Database 26AICPU INDEX BUILD
READY TO RUN
—s
ADD UPLOAD —BUILD INDEX —VERIFY SEARCH —
0 secondsShared elapsed-time scale
—SECONDS SAVED

Measured from prepared data to indexed search.
Document extraction and embedding happen first.

03

Build on GPU. Search in Oracle.

AWAITING BUILD
OracleAI Database 26aiVectors · native HNSW · searchSystem of record
VECTORS→
←GRAPH
PRIVATE AI SERVICES
NVIDIAcuVS · H100
Accelerated graph constructionCompute offload
Oracle owns the index and serves every search.

GPU activity

LIVE DEVICE SAMPLES
OBSERVED PEAK—% GPU
PEAK MEMORY—GiB
DEVICE NOW—%
Activity appears as the GPU is sampled.100%
Memory —Awaiting device samples

Observed samples stay visible after the build finishes.

MEASUREMENT TRANSPARENCY

What this run measures

After the document is searchable, both prepared scale-data lanes receive the uploaded vectors. The GPU and CPU lanes are released together on separate connections and share Oracle resources. Each lane must retrieve an uploaded vector through its index.

DOCUMENT READINESS—
UPLOADED VECTORS—
Live index parameters and measurement details appear here when a run begins.

The displayed acceleration compares prepared data to indexed search. Upload, extraction, and embedding are measured separately. Device samples describe observed activity, and may miss brief GPU peaks.
A 2–3 MINUTE CONVERSATION

Let the experiment tell the story.

  1. 00:00
    Start with their data.

    “How long does new information take to become searchable in your organization?” Upload a document, or use the Blackwell on OCI guide.

  2. 00:20
    Show the document becoming searchable.

    Point to the observed checkpoints. Ask a question while the matched scale comparison runs.

  3. 00:40
    Explain the handoff.

    Oracle sends vector graph construction to Private AI Services and NVIDIA cuVS. The completed index returns to Oracle, where searches run.

  4. 01:15
    Read this run’s result.

    Compare the CPU and GPU lanes that started together, using the same upload plus 250K scale vectors. Device activity remains visible after the GPU finishes.

  5. 02:00
    Connect it to their workload.

    “How many vectors are you indexing, and how often does the data change?” Use the result to discuss their scale and freshness needs.