NVIDIA PerfKit 5: New Tools for 3D Developers

Performance Dashboard

That's the general mode that we have used since the first versions of this utility. It provides detailed monitoring and analysis of general application performance. It shows statistics on the 3D application and GPU and video card usage in real time.

In this mode, PerfHUD shows the main performance graphs. The scrolling graph in the top right corner shows the load of unified processors of a GPU with various shader types: geometry, vertex, and pixel. In case of non-unified architectures, it shows the load of each shader type. This graph can be used to balance the load between various shader types. Another GPU load graph is located to the left of it, the vertex graph is marked red (by default), shader usage graph is green, TMU load is blue. You can add other unit loads to the graph, for example, ROPs. These graphs show sections of the pipeline that limit performance.

The main graph of this PerfHUD mode (it's in the bottom of the screen, in the middle) contains four linear graphs that show the total frame render time, the time spent in the driver, GPU idle time, and the time CPU waits for GPU. The graph is situated in the bottom part of the screen. The Draw Primitives graph above shows a number of DrawPrimitive, DrawPrimitiveUP, and DrawIndexedPrimitive calls in the process of rendering a frame. An indicator of created resources is above the latter. It blinks when various Direct3D resources are created dynamically: various textures, shaders, vertex and index buffers, etc. This indicator is really helpful, as dynamic creation of resources in Direct3D applications has a negative effect on performance. Other interface elements in this mode: info bar on top of the window and several graphs underneath.

These graphs show the number of DrawPrimitive calls and the average batch size, as well as usage of video memory allocated for frame buffers, render targets, and textures. You can replace the Batch Size/Draw Primitives graph with a histogram that shows the number of batches versus geometry size. The first column of the graph contains a bar that indicates the number of batches with the number of triangles below 100, the second - with 100-200 triangles, etc.

There is another interface element in this mode - info bar in the upper part of the screen, which shows average FPS for the last 20 frames, the number of triangles in a frame, the current speed, as well as indicators of render modes: wireframe, 2x2 textures, etc.

These modes are another interesting feature in Performance Dashboard, experiments to locate pipeline bottlenecks by disabling them. This work requires separate analysis of each pipeline stage. Since GeForce 6, new GPUs have corresponding hardware counters that show the load of each unit. But you can disable parts of a pipeline to produce a similar effect. For example, you may force a program to use a special 2C...2 pixel texture instead of all textures in order to reduce the texturing effect on the overall performance. This option can be used to determine whether texturing is the main limiting factor. The same concerns the other parts of the pipeline. You can reduce the effect of ROP and pixel shader units, as well as disable all pipeline stages (ignore DrawPrimitive and DrawIndexedPrimitive calls) in order to determine whether 3D pipe performance is limited by the code of a user application.

Some of the functions can be started with hot keys. For example, you can disable selected pixel shaders of various versions. You can enable render visualization of each version separately, up to 4.0. Rendering results are replaced with a certain color for each shader version. Similarly, you can enable wireframe and overdraw modes.

Let's talk about new features of the latest PerfHUD version. Information panel of the utility has been significantly overhauled. Developers created a convenient fully customizable user interface, so that you can add your own set of graphs of necessary size. You can choose from a full set of dozens Direct3D and GPU counters from PerfSDK. Up to four counters can be selected for each graph:

What's important, you can add custom labels (this feature does not work in our localized Windows XP), and choose location and size of the graphs.

You can resize a graph by pulling the green square in the bottom right corner just like you do with windows in operating systems. The red square closes the application, the blue square calls a configuration menu of the graph (see above). You can create as many graphs as you need, and remove them from the screen. You can save and load your custom interface in the context menu:

Graphs can be locked, so that you couldn't accidentally resize or move them, you can also restore the default interface. What concerns functional changes, we can mention new experiments: minimize geometry processing, highlight results of Pixel Shaders 4.0 and others.