In traditional CPU-driven rendering, the CPU must iterate over thousands of objects every frame, calculate camera frustum visibility, bind uniform buffers, and record individual draw commands (e.g. pass.drawIndexed()). For scenes with tens of thousands of meshes, CPU driver overhead, validation checks, and command buffer recording saturate the CPU core, dropping framerates even if the GPU has idle render capacity. In GPU-driven rendering, visibility determination (frustum culling, occlusion culling, and LOD selection) is offloaded entirely to a WebGPU Compute Shader. The compute shader dynamically writes the draw arguments directly into a GPUBuffer with GPUBufferUsage.INDIRECT. The CPU issues only a single pass.drawIndexedIndirect(indirectBuffer, 0) call, dropping CPU command recording time from O(N) to O(1).

The WebGPU specification mandates a strict 20-byte struct layout (five 32-bit words) for DrawIndexedIndirectArgs stored in a GPUBuffer with GPUBufferUsage.INDIRECT: 1) indexCount: u32 (byte offset 0, number of indices to read from index buffer); 2) instanceCount: u32 (byte offset 4, number of visible instances to draw, dynamically populated by compute shader atomic compaction); 3) firstIndex: u32 (byte offset 8, starting offset in the index buffer); 4) baseVertex: i32 (byte offset 12, value added to each index before reading vertex attributes); 5) firstInstance: u32 (byte offset 16, starting instance ID).

Each compute invocation evaluates one object instance identified by @builtin(global_invocation_id).x. The shader extracts the object bounding sphere center and radius from a read-only storage buffer, transforms the center into world space, and tests it against the 6 frustum planes (Left, Right, Bottom, Top, Near, Far). For each plane with normal n and distance d, if dot(n, center) + d < -radius, the object lies completely outside the frustum and is culled. If the object passes all 6 planes, the compute shader executes atomicAdd(&indirectArgs.instanceCount, 1u) to allocate a unique slot in a visibleInstancesBuffer, storing the object index. The vertex shader then retrieves instance transform matrices using instance_index into this compacted list.

Vulkan and DirectX 12 support native multi-draw indirect with dynamic draw count (vkCmdDrawIndexedIndirectCount), allowing a compute shader to emit an arbitrary number of distinct draw calls without the CPU knowing the count. In WebGPU 1.0, renderPass.drawIndexedIndirect() executes a single indirect draw call per invocation (though that single call can render arbitrary instances via instanceCount). For multi-mesh architectures with disparate geometries, WebGPU applications either pack meshes into a giant unified index/vertex buffer (megabuffer) and draw via compacted instance lists, or issue a fixed loop of drawIndexedIndirect() calls reading from consecutive 20-byte offsets within the indirect argument buffer.