gpu blueprint · cost models

What shader math costs, per GPU architecture.

The cost models SweetOpt uses to find faster shader code, measured on real cards with GPU Blueprint's OpBench. The goal: every architecture that runs DirectX 11.

1.0 = one fma (a fused multiply-add at full rate), measured as throughput: many independent operations in parallel. Lower is better. Numbers are rounded to what the cost models use. Texture costs are on the TexBench page.

Architectures

Greyed-out architectures have no cost model yet. If you have one of those cards, running GPU Blueprint and sending the report puts it on the map.

One operation on every architecture

One group on every architecture

Each row is an architecture, each column an operation, in fma units (lower is better), so the bars of one operation line up and compare directly. The columns after an operation are other ways to write it (named, or numbered like sign→1), from the rewrite library (listed below the table; hover for the formula). Bold: the fastest way to write the operation on that architecture, the operation itself included. Some only hold for certain values: a fixed argument such as pow(x, 2.0), or a range (marked *).

Every OpBench test

Everything OpBench measures, per architecture, in fma units on top of the test's base (lower is better, 0 = free): the middle value of its cards' newest reports. Hover a test for the code it times, an architecture for its cards. A dash: not measured yet (a newer test, or an architecture without a recent report).

Coverage

Cards measured

Cards of the same architecture measure alike, so they share one cost model. Run times: how long each program took on the card, the middle value of its reports with the current set of tests (StageBench: the pixel and the vertex shader run together).

Driver improvements

Cards of one architecture that measure differently, where the driver version tells them apart: what a newer driver made faster (or slower). Keeping your driver up to date pays. Each new report narrows down which driver brought the change. A difference can also come from two different chips of one architecture; more reports will tell. Hover a driver to see the cards.

fp16 (min16float)

How much faster half precision (min16float) runs than the same work in full precision, per architecture (the middle value of its cards; OpBench, Direct3D 11). 2.0× means twice as fast. Some cards accept min16float but run it at full precision, and some do not offer it at all. In ReShade, min16float only takes effect on Direct3D 10 to 12.

Measure your card

Download GPU Blueprint from the releases page, unzip it and double-click GPU-Blueprint.bat. When it is done it offers to send the reports. Only the card, its driver version and the measurements are published.