Hello,
I’m working on an iOS app built around a single skinned character with fur, rendered in Metal and shown in close-up, so it takes up most of the screen. It needs to run at a steady 60fps, and the oldest hardware I want to support is the A15 (iPhone 13). The fur will be alpha cards layered over the base mesh. All of the character’s motion is computed on the GPU each frame, and ARKit and Vision hand tracking are running at the same time.
This is what I have so far. A 180k-triangle test character with no hair cards runs smoothly in the Simulator on my Mac (yes, I know it's not entirely representative of the GPU numbers on a phone). On an actual iPhone 16, a synthetic bench (60k vertices, 4 influences each) put GPU skinning at 1.1 to 1.3 ms whether I used 250 joints or 1000, so I’m not worried about skinning.
The final model is still being built, so I can’t test the real thing yet, but I have to give the modeler a budget now. I’m thinking of 100k to 120k triangles in total, base mesh plus alpha cards. Would that look fairly reasonable for an A15, or not?
Thanks!
Hello @Inkline, thank you for your question!
We don't publish specific triangle count targets per device. A single number wouldn't be accurate for every app, because many things besides triangle count can push a frame past the 16.7 ms you have to hit 60 fps consistently. Draw calls, shader complexity, how many times each pixel is shaded, and whatever else the system is doing that frame can all make a scene costly to render.
Fur cards are a good example. They will most likely cost you more in pixel shading than in triangles. Your character fills most of the screen, so many card layers overlap the same pixels, and the transparent parts of those cards need to be discarded or blended. When that happens, the GPU can't skip shading the layers underneath the way it can for opaque geometry, so each layer may run your fragment shader again. See Harness Apple GPUs with Metal from WWDC20 for more information.
As you mentioned, the Simulator won't tell you much about performance. It runs your Metal calls on your Mac's GPU rather than emulating an iPhone GPU. See Developing Metal apps that run in Simulator for more information about using the Simulator with your Metal app.
Your 180k-triangle test is the right kind of test, but try to run it on an actual iPhone 13, under conditions that match your final app, to converge on the right budget. If you can prototype the same model with fur cards covering the same portion of the screen, with ARKit and hand tracking running, that will give you a much better idea of what you need to hit. Profile it with the Metal debugger and Instruments in Xcode, and look at GPU time per frame and which stage is the bottleneck (Optimize Metal apps and games with GPU counters covers how to read those counters). Your instinct to budget lower than 180k with fur cards included makes sense, but the most important thing is that your benchmark matches the conditions of your final app.
I hope that helps! Let me know if you have any questions.