Every number on this page was measured on one machine with the sgcGrid QA performance gates and two benchmark programs that ship with the QA suite. Nothing is estimated or extrapolated. Run the same programs on your own hardware and database to get your numbers.
| Measurement | Win32 | Win64 |
|---|---|---|
| 1920 x 1080 frame of a 100,000 row grid, 12 columns (QA gate C1) | 1.47 ms | 1.29 ms |
| The same frame with flashing cells and 3 format rules | 1.67 ms | 1.46 ms |
| Live updates from a worker thread into a sorted 1,000 row grid, frames rendered at 60 Hz | 1,000 updates/s at 60.0 fps, 0 missed frames | 1,000 updates/s at 60.0 fps, 0 missed frames |
| Highest update rate of one worker thread while the same grid holds 60 fps | 1,552,858 updates/s | 2,001,921 updates/s |
| Server mode, SQLite table of 10,000,000 rows: open to the first painted page | not run | 117.3 ms (2.1 ms with the estimated count) |
Measured on 2026-11-04. The tables below give the method, the machine and every figure.
A frame is the visible window of a 1920 x 1080 grid, rendered headless by TsgcGridRenderer to the raster canvas, the same code that paints the VCL and FireMonkey grids. There is no window and no GPU in the loop. The grid scrolls a little every frame, as a live grid does. C1 and the other frame gates take the median of 5 runs of 60 frames after a warm-up.
Grouping, filtering, sorting and summaries run on the grid data model with no renderer. Each figure is the median of 5 runs after a warm-up, or the best of 12 interleaved runs where the gate says so. Every gate has a budget, and the QA run fails when a figure goes over it.
A worker thread writes price updates with the thread safe UpdateRow, as the LiveTicker demo does. The main thread renders a frame every 16.67 ms (60 Hz). Every frame re-sorts the changed view, as a painted grid does. The benchmark reports the updates per second written, the frames per second rendered, the frames that missed their slot and the render time per frame.
TsgcGridServerProvider_DataSet pages a generated table through FireDAC. Each step ends when its rows are loaded and the frame is painted. Each figure is the median of 5 runs after a warm-up, with the database file in the operating system cache. Jumps land 2,000 rows apart from run to run, so every run reads from the database and not from the page cache of the grid.
| CPU | Intel Core Ultra 5 245T, 14 cores, 14 logical processors |
| Memory | 31.5 GB |
| Disk | NVMe SSD, Phison 1 TB |
| Operating system | Windows 11 Pro 26H2, build 26300.9711 |
| Compiler | Delphi 13 (compiler 37.0), Win32 and Win64 release builds |
| Database | SQLite 3.42.0 through FireDAC, embedded and statically linked |
| Source | sgcGrid development tree of 2026-11-04 (svn revision 20169 with local changes) |
The machine was a developer workstation, not an isolated benchmark server. Each heavy run held a machine wide lock, so no build or test ran at the same time.
The perf group of the sgcGrid QA suite, run alone on Win32 and Win64. The budget is the limit the gate enforces (Win32 / Win64).
| Gate | What is measured | Win32 | Win64 | Budget |
|---|---|---|---|---|
| C1 | Flat frame 1920 x 1080, 12 columns, 100,000 rows | 1.47 ms | 1.29 ms | 4.09 / 4.32 ms |
| Styled | The same frame, flashing cells and 3 format rules (average of 60 frames) | 1.67 ms | 1.46 ms | 25 / 16 ms |
| GW3 | Frame with find highlights, 100,000 rows | 1.43 ms | 1.75 ms | 25 / 16 ms |
| GW5 | Frame with 20 columns, 3 band levels, 2 frozen columns, 100,000 rows | 1.62 ms | 1.42 ms | 25 / 16 ms |
| GW7 | Frame with image, progress and 30 point sparkline columns, 100,000 rows | 2.20 ms | 3.20 ms | 25 / 16 ms |
| F1 | Frame with 2 frozen columns and a footer, scrolled both ways, 100,000 rows | 0.56 ms | 0.50 ms | 25 / 16 ms |
| G4 | Frame of 1,000,000 rows grouped by 2 columns, all expanded | 0.49 ms | 0.48 ms | 25 / 16 ms |
| D1 | Frame with 3 visible detail bands of 50 rows each | 2.47 ms | 2.65 ms | 25 / 16 ms |
| G1 | Group 100,000 rows by 2 columns, 2 summaries, build and flatten | 15.11 ms | 11.98 ms | 60 / 40 ms |
| G2 | Group 1,000,000 rows by 2 columns, 2 summaries, build and flatten | 155.37 ms | 132.61 ms | 700 / 450 ms |
| G3 | Expand or collapse one group of the 1,000,000 row view | 4.16 ms | 3.28 ms | 25 / 15 ms |
| G5 | Footer summaries, 5 columns of 100,000 rows | 5.64 ms | 2.34 ms | 25 / 15 ms |
| T1 | Tree of 100,000 rows, depth up to 8, build and flatten fully expanded | 20.42 ms | 22.80 ms | 90 / 60 ms |
| GW0 | Sorted view rebuild, 1,000,000 rows | 253.45 ms | 240.29 ms | 268.24 / 259.62 ms |
| GW1 | Distinct values of a 1,000 member text column, with a cascading filter, 1,000,000 rows | 55.61 ms | 48.42 ms | 450 / 300 ms |
| GW2 | Filter with a 500 of 1,000 value list plus a condition, 1,000,000 rows | 50.75 ms | 48.02 ms | 225 / 150 ms |
| M1 | Memory of the line view above the data, 1,000,000 rows grouped by 2 columns | 20.07 MB | 20.09 MB | 32 MB |
| X1 | XLSX export of 100,000 rows x 12 columns, 2 group levels, summaries | 1,693.87 ms | 1,002.71 ms | 4,500 / 3,000 ms |
| L2 | Log viewer following the tail while a thread appends 10,000 lines per second: frames painted per second | 38.87 fps | 38.67 fps | 20 fps minimum |
| L2 | The same run: 99th percentile of the input delay | 4.22 ms | 2.83 ms | 33 ms |
sgcGridBench_Live: a worker thread writes price, change, volume and time of a random row with UpdateRow, while the main thread renders a sorted, flashing 1920 x 1080 grid every 16.67 ms. 5 seconds per case. A missed slot is a frame that finished after the start of the next one.
| Rows | Target rate | Platform | Updates/s | Frames/s | Missed slots | Render median | Render p99 | Render max |
|---|---|---|---|---|---|---|---|---|
| 1,000 | 1,000/s | Win32 | 1,000 | 60.0 | 0 | 1.28 ms | 2.46 ms | 2.57 ms |
| 1,000 | 1,000/s | Win64 | 1,000 | 60.0 | 0 | 1.36 ms | 2.41 ms | 2.45 ms |
| 1,000 | 10,000/s | Win32 | 9,998 | 60.0 | 0 | 1.43 ms | 2.79 ms | 2.98 ms |
| 1,000 | 10,000/s | Win64 | 9,998 | 60.0 | 0 | 1.29 ms | 2.46 ms | 3.17 ms |
| 1,000 | as fast as possible | Win32 | 1,552,858 | 60.0 | 0 | 1.26 ms | 3.09 ms | 3.35 ms |
| 1,000 | as fast as possible | Win64 | 2,001,921 | 60.0 | 0 | 1.26 ms | 2.54 ms | 3.28 ms |
| 100,000 | 1,000/s | Win32 | 1,000 | 60.0 | 0 | 10.52 ms | 13.72 ms | 15.92 ms |
| 100,000 | 1,000/s | Win64 | 1,000 | 60.0 | 0 | 10.50 ms | 13.99 ms | 14.68 ms |
| 100,000 | 10,000/s | Win32 | 9,966 | 47.6 | 62 | 14.70 ms | 40.33 ms | 42.03 ms |
| 100,000 | 10,000/s | Win64 | 9,997 | 54.2 | 29 | 13.59 ms | 35.10 ms | 39.86 ms |
A sorted grid of 100,000 rows re-sorts its view in every frame where the sorted column changed, so its render time grows with the row count. Updates to columns that are not sorted keep the view, and MaxFPS can lower the frame rate of a grid that only needs to look live.
sgcGridBench_Server on SQLite 3.42.0, Win64. The table has 12 columns of mixed types (integer, text, float, date and time, boolean), an indexed integer key and an indexed customer column, with deterministic data. Every step ends with its rows loaded and the 1920 x 1080 frame painted. Median of 5 runs after a warm-up.
| Step | 1,000,000 rows | 10,000,000 rows | What runs on the database |
|---|---|---|---|
| Open to the first painted page | 2.8 ms | 117.3 ms | Columns from the schema, exact COUNT(*), first page |
| Open with the estimated count | 2.1 ms | 2.1 ms | CountMode = gscEstimated: first page only, the count grows while pages come back full |
| Exact row count | 0.2 ms | 115.6 ms | SELECT COUNT(*) |
| Jump to the middle | 40.1 ms | 378.8 ms | One page at the middle row, LIMIT and OFFSET |
| Jump to the end | 75.0 ms | 735.4 ms | The last page, LIMIT and OFFSET |
| Sort by an indexed column | 2.5 ms | 119.3 ms | ORDER BY customer through its index, plus the exact count of the new query |
| Sort by a column with no index | 106.9 ms | 1,169.3 ms | ORDER BY amount DESC, a full scan |
| Filter builder condition | 97.5 ms | 975.7 ms | WHERE region = 'North' AND qty > 50, count and first page (40,000 and 400,000 rows match) |
| Server grouping, group rows painted | 266.5 ms | 3,917.8 ms | GROUP BY region (12 groups) with the row count of each group, a full scan |
| Process memory after the run | 76.5 MB | 75.7 MB | Working set of the benchmark process, SQLite page cache included |
The grid itself reads only the pages on screen, so its own cost stays flat. What grows with the table is the work of the database: the exact count, the OFFSET of a deep jump, and a sort or a grouping on a column with no index. An index on the sorted or grouped column, or the estimated count, keeps those steps short. Generating the 10,000,000 row table took 42.5 s, and a second run on the same file gave figures within 25% of these.
SQL Server, PostgreSQL, Firebird and tables of 100,000,000 rows are measured with the same harness. Their figures will be added here when they have been run. Until then, no number is published for them.
The benchmark programs are in the QA folder of the sgcUI source, QA\Grid\Bench. They are console programs with no IDE project.
build_bench.bat builds sgcGridBench_Server.exe and sgcGridBench_Live.exe for Win32 and Win64 with Delphi 13 into bin\win32 and bin\win64, and copies the ini file next to them.
sgcGridBench_Server.exe --rows=10000000 creates the table when it is missing, then measures every step and writes a Markdown report and a CSV line per figure. Options: --ini, --rows, --runs, --db, --generate, --out.
Edit sgcGridBench_Server.ini: DriverName is the FireDAC driver (MSSQL, PG, FB or SQLite), every key of [Params] goes to the connection, and Dialect selects the SQL dialect. The ini holds a sample for each engine.
sgcGridBench_Live.exe --seconds=5 runs the live cases above. The client gates are the perf group of the QA suite, sgcGridQA.exe --group=perf, which appends every figure to sgcGridQA_perf.txt. Run it alone on an idle machine.
[Connection]
DriverName=PG
[Params]
Server=localhost
Port=5432
Database=bench
User_Name=postgres
Password=secret
[Bench]
Dialect=postgresql
Table=bench_orders
Rows=10000000
Runs=5