v146 · Web APIs · Memory Benefit

Memory Benefit

Why TRANSIENT_ATTACHMENT saves memory bandwidth on tile-based GPUs, which GPU architectures benefit, and when to apply the flag in your WebGPU render pipeline.

GPU architecture comparison

Immediate mode GPU (desktop, discrete)

Renders each primitive as it's submitted
Render targets stored in VRAM
All writes/reads go through VRAM bus
TRANSIENT_ATTACHMENT accepted but no bandwidth saving
Examples: NVIDIA, AMD discrete GPUs

Tile-based GPU (mobile, Apple Silicon)

Divides frame into tiles, renders each tile in fast on-chip RAM
Flush tile to main memory only when needed
TRANSIENT_ATTACHMENT → tile buffer never flushed to RAM
Saves memory bandwidth proportional to texture size
Examples: Apple A/M chips, Mali, Adreno, PowerVR

bandwidth saving estimate

Texture	Size (1920×1080)	Tile-based benefit	Immediate GPU benefit
Depth buffer (depth24plus)	~8 MB per frame	Eliminated (tile-local)	None
MSAA resolve target	~32 MB per frame (4× MSAA)	Eliminated if not read back	None
Intermediate render target	~8 MB per frame	Eliminated if used only within pass	None
Colour attachment (final output)	~8 MB per frame	Not transient — must be presented	Not transient

when to use TRANSIENT_ATTACHMENT

// Use TRANSIENT_ATTACHMENT when ALL of these are true:
// 1. The texture is used only as a render target (not sampled)
// 2. You don't need to read the texture after the render pass
// 3. The store op is 'discard'

// ✓ Depth buffer for a single-pass render
const depth = device.createTexture({
  format: 'depth24plus',
  usage: GPUTextureUsage.RENDER_ATTACHMENT | GPUTextureUsage.TRANSIENT_ATTACHMENT,
  size: [w, h],
});

// ✓ MSAA intermediate buffer (discarded after resolve)
const msaa = device.createTexture({
  format,
  sampleCount: 4,
  usage: GPUTextureUsage.RENDER_ATTACHMENT | GPUTextureUsage.TRANSIENT_ATTACHMENT,
  size: [w, h],
});

// ✗ Colour output read as a texture in the next pass
// — cannot use TRANSIENT_ATTACHMENT
const pingPong = device.createTexture({
  usage: GPUTextureUsage.RENDER_ATTACHMENT | GPUTextureUsage.TEXTURE_BINDING,
  // TRANSIENT_ATTACHMENT incompatible with TEXTURE_BINDING
  size: [w, h],
});

scenario focus

Select a scenario to focus its rendered example and summary.