Tile-Based Deferred Rendering vs Traditional Immediate Mode Desktop GPUs: Advanced Implementation and Real-World Architecture

[Jake Campbell]

I’ve been digging into why TBDR-based mobile and Apple Silicon GPUs consistently eat immediate-mode desktop cards for power-to-performance in mobile workloads, but I’m hitting a wall trying to map that efficiency gap onto the actual binning pass overhead. When you’re binning draw calls per tile and you have heavy overdraw with translucent overlays, where exactly does the on-chip tile memory bandwidth become the bottleneck—is it the resolve step or the initial color/depth write? And at what tile resolution do you start seeing diminishing returns where the extra binning pass cost eats the bandwidth savings?

Second question: for someone actually implementing adaptive tile sizing in a custom engine rather than relying on driver heuristics, how are you handling dynamic batching across tile boundaries without killing your draw call throughput? I’m benchmarking a Vulkan renderer with manual binning and seeing 40% power spikes in dense scenes, but I can’t tell if it’s the memory controller or shader occupancy. Anyone here have real numbers on where the crossover point is between TBDR and immediate mode on desktop-class silicon, or is the efficiency advantage purely a low-power domain thing?

"Benchmark results tested on clean Windows 11 with Core Isolation disabled."

[Author]

Posts