Planner-Driven GPU Execution Inside PostgreSQL
CYBERTEC PGEE GPU Edition adds GPU-accelerated execution to CYBERTEC PostgreSQL Enterprise Edition (PGEE) for selected analytical queries. It is not a separate database engine and does not replace PostgreSQL: applications keep using the PostgreSQL SQL interface, and the query planner decides when a GPU path makes sense and when the CPU should handle the work.
This GPU-aware behaviour is the core of the design. Queries that do not benefit from GPU execution, or whose correctness cannot be proven on the GPU path, stay on the CPU. The implementation is built on NVIDIA GPUs and CUDA, and includes mechanisms for GPU memory management and for handling PostgreSQL's process model.
The architecture combines the following components:
| Component | Role |
| CYBERTEC PGEE | Provides the PostgreSQL enterprise database foundation |
| PGEE GPU Edition | Adds GPU execution paths for supported queries |
| NVIDIA GPUs | Execute suitable data-intensive operations in parallel |
| PostgreSQL query planner | Evaluates execution options and selects an appropriate path |
| Existing applications and SQL | Continue to use the PostgreSQL environment |
The intent is to let organizations evaluate GPU acceleration as an extension of their existing PostgreSQL infrastructure, without an analytical database migration and without new application code for supported queries. That said, not every SQL query is accelerated — actual gains depend on query structure and the operations the GPU implementation supports, so representative queries should be evaluated before drawing conclusions.
Query Patterns and Data Types on the GPU Path
Analytical workloads typically scan large tables, apply filters, join datasets, and aggregate across millions or billions of rows — the kind of similar, parallel work GPUs are built for. The accelerated patterns CYBERTEC documents include:
- Scalar aggregates such as
COUNT,SUM,AVG,MIN, andMAX. GROUP BYon supported keys.DISTINCTandCOUNT(DISTINCT).- Supported star-schema joins — inner, left, full, semi, and anti joins.
- Joins combined with grouping and supported aggregate filters.
- Range filters that can be pushed down into the scan.
Selected integer, date, timestamp, numeric, floating-point, text, and UUID types are supported, each with its own implementation rules. Coverage is deliberately selective: certain floating-point aggregates remain on the CPU to preserve PostgreSQL's numerical behaviour, and complex query shapes outside the planner's supported patterns also continue to use CPU execution.
Correctness Under Concurrent Writes
Execution speed matters only if results stay correct, and the design accounts for PostgreSQL's transaction visibility rules. Data cached on the GPU is used only while visibility information allows it to be trusted for the query's snapshot. Once a page is no longer provably visible to that snapshot, the affected portion is recomputed on the CPU within the same query. Concurrent writes can therefore shrink the amount of work done on the GPU, but they do not change the result the query must return.
These statements describe the documented validation work and are not a guarantee that every possible query or deployment has been tested.
Benchmark Results and Their Limits
The published figures show large speedups on selected analytical workloads, but each must be read with its baseline, dataset, and GPU configuration in mind.
| Workload | CPU baseline | GPU configuration | GPU execution time | Reported speedup |
| Orders scan, filter, and aggregation, 1 billion rows | 159 seconds on 1 CPU core | 8 NVIDIA L40S GPUs | 2.8 ms | 57063x |
| Instant KPI dashboard, 500 million rows | 30,993 ms on 1 CPU core | NVIDIA GB10 | 14.7 ms | 2,109x |
| TPC-DS revenue by item, SF100 | 141,454 ms on 1 CPU core | NVIDIA GB10 | 263.9 ms | 536x |
These are specific documented tests rather than guaranteed speedups. The 57,063x result was measured on eight NVIDIA L40S GPUs, while the KPI and TPC-DS results used NVIDIA GB10 hardware, so they are not directly interchangeable. CYBERTEC also notes that speedups above 100,000× can be observed on NVIDIA H200, B200, or B300 GPUs.
In the TPC-DS work, the standard 99-query set runs against a schema at SF100 (a 100 GB database for this test). The report identifies 63 queries with a shape the GPU engine can accelerate; 33 accelerated results passed its correctness gate. A revenue-by-item query completed in 263.9 milliseconds on NVIDIA GB10 against a measured single-core PostgreSQL baseline of 141,454 milliseconds — a reported 536x for that query. These are not audited official TPC-DS results: the report states that it computes no official composite metric and runs no full set of benchmark phases.
Tested Hardware and Multi-GPU Scaling
The benchmark materials cover two primary configurations:
| NVIDIA hardware | Documented use |
| NVIDIA GB10 | Development, continuous integration, and TPC-DS benchmark testing |
| Eight NVIDIA L40S GPUs | Multi-GPU testing and the 1-billion-row orders benchmark |
| Four NVIDIA H200 GPUs | Additional GPU benchmark configuration |
The technical overview also lists supported NVIDIA architectures and compatibility requirements, but appearing in a compatibility matrix does not mean a given model was used for the results above. Performance depends on the workload, the chosen execution path, available GPU resources, and whether the required data already sits in GPU memory.
More GPUs generally help workloads that can use multiple devices and add total VRAM for caching larger datasets, but scaling is not necessarily linear and should be measured on representative workloads. The number of GPUs used for supported aggregate operations is configurable without application changes; whether a specific query improves is best established through measurement and execution-plan inspection.
Operational Fit
Enterprise use also depends on inspection and control. The technical overview documents inspectable execution plans, GPU memory controls, and ways to investigate why a query did or did not take a GPU path. CYBERTEC, which develops PostgreSQL products and offers consulting, performance tuning, infrastructure services, and enterprise support, provides 24x7 enterprise support for teams running this in production.
A practical evaluation starts with identifying representative analytical queries, reviewing their execution plans, and measuring performance on the intended hardware and data.
How PGEE GPU Edition Fits Into an Existing PostgreSQL Deployment
PGEE GPU Edition is not a new database engine. It extends CYBERTEC PostgreSQL Enterprise Edition so that supported analytical queries can execute on GPUs, and PostgreSQL stays the system of record. GPU execution is an optional additional path rather than a substitute for the existing engine. Because queries flow through the standard PostgreSQL SQL interface, neither application code nor SQL statements need to change.
Benchmark Numbers and Their Scope
CYBERTEC's documented benchmark reports an 11,200× speedup on a 2.4-billion-row workload. That workload completed in 12.3 seconds on 32 CPU cores, versus 1.10 milliseconds on four NVIDIA H200 GPUs.
The comparative measurements using the TPC-DS query set and schema are not an audited official TPC-DS result; the extended report states this explicitly. Results in general depend on the workload and configuration.
Execution Behavior: CPU, GPU and Snapshot Consistency
Not every query reaches the GPU. The query planner picks the appropriate execution path, and queries that fall outside the supported acceleration patterns — or whose GPU path is unsuitable — continue to run on the CPU.
Ongoing writes do not disqualify a query from acceleration. The documented design consults PostgreSQL visibility information to decide whether cached data is usable for a query's snapshot. Any portion that cannot be proven visible is recomputed on the CPU within that same query.
Scaling and Support
- Scaling characteristics — for workloads that benefit from parallel execution, scaling is in most cases close to linear, and additional GPUs raise total GPU memory so larger datasets fit. Actual scaling depends on the workload and configuration.
- Vendor services — CYBERTEC offers 24x7 enterprise support alongside PostgreSQL products, consulting, performance tuning, and infrastructure services.



