three.js's EffectComposer clones its render target, multisampling and all, for its second buffer

Fact

new EffectComposer(renderer, target) makes its second buffer with target.clone(), and RenderTarget.copy carries samples and the depth buffer. Hand it a multisampled target and both ping-pong buffers are multisampled, though only the scene's own render gains anything from it; every full-screen pass after that draws into, and resolves, a multisampled target.

Why it matters

Multisampled storage is samples × (colour + depth) bytes per pixel: with 4× MSAA and half-float colour, about 48 bytes per pixel, roughly 250 MB per buffer on a 2880×1800 screen. Doubled, that is half a gigabyte of GPU memory for one picture, which risks losing the context on phones and tablets.

How to apply

• Draw the scene yourself into one multisampled target with a DepthTexture, then run the passes (pass.render(renderer, write, read), swapping only when pass.needsSwap) between two plain targets, the last with renderToScreen = true.
• Or give the composer a target without samples and smooth edges afterwards (FXAA or SMAA, after the output pass, on sRGB values).
• Cap the post-processing resolution by a pixel count, not a fixed pixel ratio: memory and fill cost grow with the screen, not with the scene.