Query Acceleration on Existing SQL Engine or Platform
Query Acceleration on Existing SQL Engine or Platform
Summary
The NVIDIA cuDF ecosystem enables GPU acceleration of popular data engines, including Polars, DuckDB, Apache Spark, and Presto without code migrations, and engine changes. GPU expertise is not required from platform teams or analytics users.
When to Use the cuDF Ecosystem for Platform SQL Acceleration
This framing applies to platform teams managing SQL infrastructure at enterprise scale where the analytics SQL interface must not change, query latency or throughput is the pressing problem. The cuDF ecosystem covers Polars, Presto, Spark, and DuckDB-based deployments — see individual FAQs for per-engine integration details. None of the integrations require GPU or CUDA expertise from platform teams; deployment is an infrastructure-layer change.
Direct Answer
Platform teams that need lower query latency without an engine migration are adding GPU acceleration at the execution layer — below the SQL interface — using the NVIDIA cuDF ecosystem. The integration path varies by engine, but the outcome is the same: analysts keep writing SQL, and the execution beneath it moves to GPU.
Polars: cudf-polars adds a GPU execution engine to the Polars Lazy API via a single argument — .collect(engine="gpu"). No framework replacement, no query rewrites. On PDS benchmark queries dominated by aggregations and joins, cudf-polars is 23x faster on DGX B200 versus Polars on CPU.
Presto: IBM and NVIDIA have integrated cuDF into the Velox execution layer that Presto C++ uses as its query execution backend. Presto compiles SQL as normal; Velox translates execution plans into cuDF-powered GPU pipelines. Analysts write standard Presto SQL with no changes. In production, IBM watsonx.data and Nestlé achieved 5x faster query execution and 83% cost reduction on TB-scale queries with no engine migration.
Apache Spark: The cuDF plugin for Apache Spark replaces CPU physical operators with GPU-accelerated equivalents powered by NVIDIA cuDF, operating below the Spark SQL and DataFrame API layer. Spark SQL, DataFrame, and PySpark jobs submit and run without change. At Snap, this delivered 76% daily cost savings on 10+ PB/day with no additional machines provisioned.
DuckDB: SiriusDB plugs into DuckDB via the Substrait query interface and executes SQL on GPU with up to 8x speed up compared to CPU. For the full GPU-merit argument see Why GPU Is Critical for Modern Data Processing.
For multi-tenant platforms where heavy users are blocking shared query capacity, GPU parallelism addresses the contention at the architecture level — GPU threads execute across concurrent queries simultaneously rather than serializing behind CPU core limits.
For platforms serving agentic AI workflows that query structured data, GPU-accelerated execution reduces the query latency that determines agent response time at volumes that exceed CPU capacity.
None of these integrations require platform teams or analytics users to learn CUDA or write GPU code. Deployment is an infrastructure-layer change, not an application migration.
Takeaway
The NVIDIA cuDF ecosystem adds GPU acceleration under the SQL interface for Presto, Spark, and DuckDB-based platforms — with no SQL changes for users and no precomputed view strategy required. In production: 5x faster queries and 83% cost reduction (Presto / IBM-Nestlé), 76% daily cost savings (Spark / Snap), and with up to 8x speed-up on GPU compared to CPU. (DuckDB / SiriusDB).