Date: 14 May 2026 Author: OpenCode Technical Research Unit Citation style: Harvard (alphabetical) Language: British English Document ID: 0321
Open-world titles such as Grand Theft Auto VI exert ferocious pressure on the texturing pipeline. A single Vice City block can host hundreds of bespoke surfaces — neon signage, sun-bleached stucco, sand-pitted asphalt, hand-painted murals — and the rendering engine must page all of them in and out of GPU memory whilst sustaining a sixty-hertz frame budget on consumer hardware. The two dominant architectural answers to this problem trace back to a common ancestor: id Software's MegaTexture, and the more generalised family of techniques now grouped under "virtual texturing" (Wikipedia, 2026). This report contrasts the original clipmap-derived MegaTexture lineage with modern sparse virtual texturing (SVT), tile-based residency in DirectX 12 Ultimate and Vulkan, and the runtime virtual texturing (RVT) systems implemented in commercial engines, with specific reference to how Rockstar's RAGE engine is reputed to handle this workload in GTA VI.
MegaTexture is, formally, a clipmap implementation. A clipmap clips a mipmap pyramid to the subset of texels actually visible to the camera, so that only a moving window of high-resolution detail need reside in graphics memory at any one time (Wikipedia, 2026). John Carmack adapted this approach for Enemy Territory: Quake Wars (2007) on id Tech 4, where the entire terrain was painted as one enormous unique texture and streamed from disk on demand. The technique allowed artists to abandon the visually monotonous "tile-and-decal" workflow inherited from Quake III: every square metre of ground could be hand-authored without repetition.
The id Tech 5 incarnation, shipped in Rage (2011), pushed the idea much further. Textures were authored at resolutions up to 128,000 × 128,000 texels, and — crucially — virtual texturing was extended beyond terrain to cover characters, props and sprites (Wikipedia, 2026). The engine subdivided the virtual address space into fixed-size pages (typically 128² texels including a border), and a feedback pass rendered page identifiers into an offscreen target so the streaming subsystem could prioritise loads. The cost was infamous "texture pop-in" when the camera turned faster than the disk could feed it, a symptom of an entirely CPU-bound transcoding pipeline.
Sean Barrett's GDC 2008 talk on Sparse Virtual Textures (SVT) formalised what Carmack had described in scattered forum posts and emails (Barrett, 2008). The SVT architecture maintains an indirection page table mapping virtual tile coordinates to physical cache slots, and a pixel shader performs the redirection at sample time. Mip-level mismatch is handled by walking up the page table — or, as Barrett later conceded after audience feedback, by a bounded loop in the pixel shader itself. The technique is largely orthogonal to the source content: it works equally well for a single 256k² landscape or for thousands of conventional textures packed into a virtual atlas.
Two structural problems dogged early SVT. First, decompression: DXT/BCn blocks must be transcoded from an intermediate disk format (often a custom JPEG-XR variant) on the CPU, since GPUs of the era could not decompress on-the-fly. Second, anisotropic filtering across page borders required either explicit gutter pixels or shader-side filtering, both of which incurred bandwidth penalties (Barrett, 2008).
The advent of Tier-2 Tiled Resources in Direct3D 11.2 — and their cross-vendor descendants VK_EXT_sparse_residency and D3D12_TILED_RESOURCEs — moved much of the SVT bookkeeping into the driver. The GPU's MMU now resolves virtual texel addresses against a sparse residency map, and an undefined-tile read returns deterministic zero rather than a fault. This eliminates the explicit shader-side page-table walk and permits anisotropic hardware filtering across resident tiles without gutter padding (Microsoft, 2024).
Modern engines layer two complementary systems on top. Streaming virtual textures page authored, baked content from disk, much as id Tech 5 did, but using hardware residency and GPU-resident transcoders (BC7 via compute shaders, or ASTC on mobile). Runtime virtual textures, popularised by Unreal Engine 4.23 and onwards, composite material-graph outputs — terrain layer blends, decal projection, puddle masks — into virtual texture tiles on the GPU itself, caching the result for many frames so that expensive per-pixel material work is amortised (Epic Games, 2024). The conceptual descent from MegaTexture is direct, but the implementation is unrecognisable: no disk streaming, no CPU transcoding, and the cache is invalidated by gameplay events rather than by camera motion alone.
Rockstar's RAGE engine has historically used a bespoke streaming texture system rather than a literal MegaTexture, favouring instanced, repeated materials with high-resolution decal overlays — a pragmatic choice given the PS3/Xbox 360 origins of GTA V's pipeline. Public technical material on GTA VI remains limited, but pre-release footage suggests an evolution towards hardware-tiled streaming virtual textures for the open world, with RVT-style composition for the wet-surface and sand-displacement effects prominent in the Vice City reveal trailer. The hybrid approach side-steps the principal MegaTexture liability — the inability to author dynamic, gameplay-reactive surfaces — whilst retaining the artistic freedom of unique texel authoring across the map.
The trade-off is residency pressure on a console memory budget shared with high-resolution shadow maps, GI probes and ray-tracing acceleration structures. A reasonable estimate, extrapolating from public Doom Eternal numbers, places the physical virtual-texture cache at 1–2 GiB on PS5 and Xbox Series X, with a tile size of 128² or 256² BC7 texels and SSD-fed streaming at multiple gigabytes per second. The PCIe-direct storage path on current consoles makes the historic id Tech 5 pop-in artefacts largely a thing of the past.
MegaTexture was a watershed: it proved that unique, non-repeating texel data could clothe an entire game world. Sparse virtual texturing generalised the idea away from terrain and away from id Software, and hardware tiled resources finally moved the indirection out of the shader and into silicon. GTA VI almost certainly inherits this lineage rather than reinventing it, combining streaming virtual textures for authored content with runtime virtual textures for material composition. The net effect for the player is invisible — exactly as a successful texturing system should be.
Barrett, S. (2008) Sparse Virtual Textures. Available at: https://silverspaceship.com/src/svt/ (Accessed: 14 May 2026).
Epic Games (2024) Virtual Texturing in Unreal Engine. Available at: https://docs.unrealengine.com/ (Accessed: 14 May 2026).
Microsoft (2024) Tiled resources — Direct3D 12. Microsoft Learn. Available at: https://learn.microsoft.com/windows/win32/direct3d12/ (Accessed: 14 May 2026).
van Waveren, J.M.P. (2009) id Tech 5 Challenges: From Texture Virtualization to Massive Parallelization. id Software. Available at: http://mrl.cs.vsb.cz/people/gaura/agu/05-JP_id_Tech_5_Challenges.pdf (Accessed: 14 May 2026).
Wikipedia (2026) Clipmap. Available at: https://en.wikipedia.org/wiki/Clipmap (Accessed: 14 May 2026).