skip to content

As rendering lead, how would you set the block-compression format standard per texture class and platform tier, weighing quality, memory and hardware support?

level: principalimportance: should knowfreq 34%

answer

  1. decide per class, not per texture
  2. fixed-size blocks, random access
  3. correlated versus independent channels
  4. one block size, many footprints
  5. measure, default, allow exceptions

basics

~20 s

Decide per texture class, not per texture: BC7 or BC1 sRGB for colour, BC5 for normals, BC4 for masks, BC6H for HDR on desktop-class GPUs; ASTC block sizes or ETC2 on mobile. Measure quality; make exceptions cost budget.

solid answer

~50 s

GPUs sample **block-compressed** formats directly because fixed-size blocks allow random access, so the standard picks a native format per **texture class** and **platform tier**. Desktop-class GPUs use BCn: BC7 sRGB for hero colour, BC1 (half the size) where noisy content hides its banding, BC5 for normals because x and y are independent channels, BC4 for single masks, BC6H for HDR. Mobile GPUs mostly decode ASTC, whose 128-bit block covers 4x4 at 8 bits per texel down to 12x12 at about 0.89, so block size becomes a per-class quality dial; ETC2 and EAC cover older devices. I would set defaults by declared texture type in import, measure candidates on real content with an error metric plus lit review, let artists request a richer format with its memory cost shown, and re-run the table for each new platform tier. There is no single right table; the art style and the memory left decide.

go deeper

for a junior

Know that GPUs read block-compressed textures directly, name the common BCn, ASTC and ETC2 formats, and remember that colour, normal and mask textures use different ones.

for a middle

Explain bits per texel for BC1, BC5, BC7 and ASTC block sizes, why normals want two independent channels, and which formats carry sRGB variants.

for a senior

Diagnose format artefacts such as banding, hue shifts and blocky normals, and argue from measurement on real content which format fixes them at what memory cost.

for a principal

Own the table: defaults per texture class and platform tier, the measurement that justifies them, an exception process tied to the budget, and when a tier earns its own ship set.

## What the standard decides A texture-format standard answers, for every **class** of texture and every **platform tier**, which GPU-native format it ships in. It exists because the choice repeats thousands of times: left to each artist, the same kind of map ends up in several formats, quality is inconsistent and memory goes where nobody planned. The memory budget itself is set with the engine's memory planning, and when and where conversion runs belongs to the asset pipeline; the standard decides what the conversion produces. GPUs sample **block-compressed** formats directly: every block of texels compresses to a fixed number of bits, so any texel's block can be located by arithmetic and decoded in hardware as it is filtered. Image formats with variable-size output cannot be sampled that way, which is why textures do not stay in general-purpose image formats in GPU memory. ## The formats on the table | Format | Channels | Bits per texel | Typical use | sRGB variant | |---|---|---|---|---| | BC1 | RGB + 1-bit alpha | 4 | opaque colour, noisy content | yes | | BC3 | RGBA | 8 | colour with smooth alpha, older choice | yes | | BC4 | 1 | 4 | single masks, roughness, height | no | | BC5 | 2 | 8 | tangent-space normal maps | no | | BC6H | RGB half-float | 8 | HDR environment and emissive maps | no | | BC7 | RGB or RGBA | 8 | high-quality colour and alpha | yes | | ETC2 RGB / RGBA | 3 or 4 | 4 / 8 | colour on mobile GPUs | yes | | EAC R11 / RG11 | 1 or 2 | 4 / 8 | masks and normals on mobile GPUs | no | | ASTC | 1 to 4, LDR or HDR | 8.00 (4x4) to 0.89 (12x12) | every class, where supported | yes | Uncompressed RGBA at 8 bits per channel is 32 bits per texel, so BC1 is 8:1 and BC7 4:1 against it. Every ASTC block is 128 bits whatever its footprint, so its rate is 128 divided by the texels per block: 6x6 is about 3.56 bits per texel and 8x8 exactly 2. The BCn family is what desktop-class GPUs support; ASTC is widespread on recent mobile GPUs and uncommon on desktop-class ones; ETC2 is the older mobile baseline. Not every GPU that decodes LDR ASTC decodes the HDR profile. ## A per-class default | Texture class | Desktop-class tier | Mobile tier with ASTC | Fallback tier | |---|---|---|---| | Hero albedo, skin, UI | BC7 sRGB | ASTC 4x4 or 5x5 sRGB | ETC2 sRGB | | Bulk albedo, rock, terrain detail | BC1 or BC7 sRGB | ASTC 6x6 sRGB | ETC2 sRGB | | Normal maps | BC5 | ASTC 4x4 or 5x5, two channels | EAC RG11 | | Single-channel masks | BC4 | ASTC 6x6 to 8x8 | EAC R11 | | HDR environment, emissive | BC6H | ASTC HDR where supported, else a float format | float format | This is a starting proposal, not a law. The channel layout of packed material masks is decided with the material model; the standard only picks the format for the result. ## The trade-offs a lead weighs - **Quality against size.** BC1 stores two endpoint colours and two interpolated ones per block, so smooth gradients band and neighbouring hues shift; BC7's modes and partitions fix that at twice the memory. Noisy, high-frequency content hides BC1's errors well. - **Correlated against independent channels.** Colour formats assume the channels move together; a normal map's x and y do not, so BC5's two independent channels beat any colour format for them. - **Support against reach.** A format the GPU cannot sample must be decoded or transcoded at load into one it can, often a larger one, costing load time and several times the memory; a tier without ASTC needs ETC2 or its own ship set. - **One download against several.** A transcodable intermediate format converts to each device's native block format at load, trading some quality and load time for a single package; native sets per tier cost build time and storage instead. - **Encode time.** High-quality BC7 and ASTC encoders are slow, which hurts iteration; a fast preset for development and a thorough one for release is common. ## Running the standard 1. **Default by texture type.** Import assigns the format from the texture's declared class, so artists rarely choose. 2. **Measure.** Compare candidate formats on a reference set of real textures with an error metric such as PSNR or SSIM and an in-engine review under lighting, because some errors only show in specular highlights. 3. **Allow exceptions with a visible cost.** An artist can request BC7 over BC1 or 4x4 over 6x6, and the request shows against the memory budget. 4. **Re-run per platform tier.** Each new device class gets the table re-evaluated rather than inheriting the nearest tier's answer. ## Where judgement remains There is no single right table. Whether bulk albedo deserves BC7, how coarse an ASTC block terrain masks tolerate, or whether a low-end tier earns a separate ship set depends on the art style, the content mix and the memory left after everything else. The lead's job is to make the default sensible, the measurement repeatable and the exceptions explicit.

  • When would you accept BC1 for albedo instead of BC7?
    For noisy, high-frequency content such as rock, gravel or terrain detail, where BC1's limited palette per block is masked, and when memory is tight. I would reject it for smooth gradients, skin, skies and UI, where its banding and hue shifts show. The decision comes from measuring both on representative textures under game lighting, not from a blanket rule.
  • How do you handle a platform tier whose GPUs lack ASTC?
    Either ship ETC2 and EAC for that tier, accepting lower quality at similar sizes, or use a transcodable intermediate format that converts to whatever the device decodes at load. Decompressing to uncompressed formats is the last resort because it multiplies memory several times. The choice depends on how many players the tier holds and the cost of a separate ship set.
  • Why should normal maps not share the albedo's format?
    Colour formats such as BC1 and BC7 assume correlated channels and spend their bits on colour lines, while a normal map's x and y vary independently and must stay accurate for lighting. BC5, EAC RG11 or two-channel ASTC store the two components separately and let z be rebuilt from unit length, giving clearly better lighting at a comparable or smaller size.

  • One format for every texture keeps the pipeline simple, so that is the right standard.
  • BC7 is always worth its extra memory over BC1.
  • ASTC suits every platform, so desktop and mobile should share one format set.
  • Format choice belongs to the asset pipeline team, not to rendering.
  • A format the GPU cannot sample is fine because the driver handles it for free.