, JULY 2026 Researched 2026-07-31 from primary sources for the Nishi sovereign library. Every claim carries a source and a status: SHIPPING / PREVIEW / ANNOUNCED / RESEARCH / DEPRECATED. Purpose: rank what an independent sovereign team should build next, and what it should refuse to build. === THE HEADLINE: THE NEURAL API YOU WOULD HAVE TARGETED NO LONGER EXISTS === DirectX Cooperative Vectors was DEPRECATED, not shipped. Microsoft DirectX team, 2026-02-26: "Cooperative Vector has been deprecated in favor of a future design unifying matrix-matrix and vector-matrix operations, coming in Shader Model 6.10." https://devblogs.microsoft.com/directx/shader-model-6-9-retail-and-more/ Anyone who spent 2025 building on CooperativeVector is rewriting now. This is the strongest single argument in this document for where NOT to spend the next six months. === 1. API STATUS AS OF 2026-07-31 === Shader Model 6.9 SHIPPING RETAIL Agility SDK 1.619 + DXC 1.9.2602.16, 2026-02-26 Cooperative Vectors DEPRECATED never reached retail SM 6.10 / DX LinAlg PREVIEW AgilitySDK 1.720-preview + DXC 1.10.2605.2, 2026-04-27 LinAlg VectorAccumulate PREVIEW AgilitySDK 1.721-preview, 2026-05-28 DX Compute Graph Compiler ANNOUNCED MLIR-based, private preview summer 2026 DXR 2.0 ANNOUNCED spec v0.30 dated 2026-07-10, preview late summer 2026 Work Graph MESH NODES NOT SHIPPING spec v1.012: "proposed ... aren't supported yet"; Draw/DrawIndexed nodes CUT WHAT SM 6.9 ACTUALLY DELIVERED, AND WHY IT MATTERS: Long Vectors -- native DXIL vectors up to 1024 elements -- plus mandatory 16-bit and 64-bit shader ops. Hardware: AMD RX 9000 (driver 26.2.1), Intel Arc B-series, NVIDIA all RTX (driver 595+). ★LONG VECTORS ALONE ARE ENOUGH TO WRITE A RESPECTABLE MLP INFERENCE KERNEL WITH NO MATRIX HARDWARE. That is the on-ramp for a sovereign implementation: no vendor extension, no preview SDK. DX LinAlg (SM 6.10) shape, for when it goes retail: three scopes (thread = matrix-vector, the neural-shading case; wave = matrix-matrix on MMA/WMMA/XMX; threadgroup = driver-tiled). Component types FP32, FP16, F8_E4M3FN, F8_E5M2, U8, S8, U32. Tier 1.0 mandates U8xU8->S32 and S8xS8->S32. ⚠FP8 matrix inputs must be ACCEPTED but MAY be EMULATED -- do not assume FP8 is fast. Layouts ROW_MAJOR / COLUMN_MAJOR / MUL_OPTIMAL / OUTER_PRODUCT_OPTIMAL; 128-byte buffer base alignment, 16-byte stride multiples. VectorAccumulate is the gradient-accumulation primitive, i.e. what makes IN-SHADER TRAINING viable rather than inference only. Vendor commitment to LinAlg + CGC: AMD, Intel, NVIDIA and Qualcomm (GDC 2026). ⚠VULKAN PARITY IS UNVERIFIED. VK_NV_cooperative_vector exists (NVIDIA-only). Whether a cross-vendor KHR/EXT equivalent exists as of July 2026 could NOT be confirmed. Do not assume parity. === 2. NEURAL TEXTURE COMPRESSION -- THE ONE NEURAL TECHNIQUE WITH REAL NUMBERS === NVIDIA RTX NTC SDK, still v0.9.2 BETA. https://github.com/NVIDIA-RTX/RTXNTC Compresses ALL PBR channels of a material JOINTLY (9-10 channels, about 64 bits/texel baseline) down to about 5 bits/texel -- roughly 12.8:1 versus uncompressed -- at quality comparable to BCn. Quality bar from the SDK's own docs: PSNR 35-40 dB "sufficient", 50 dB "perceptually lossless". Real bundles land 40-50 dB. Concrete: a 2K PBR material = 2.5 MB compressed vs 12 MB uncompressed. Two modes: INFERENCE ON SAMPLE (decode in the pixel shader, 2.5 MB VRAM, Ada+ recommended) and INFERENCE ON LOAD (transcode to BCn at load; 2.5 MB disk / 12 MB VRAM, and RUNS ON ANY SHADER MODEL 6 GPU). Inference on Feedback reaches up to 6x memory saving vs fully-mapped BCn. Cooperative Vector gave 2-4x inference throughput on Ada/Blackwell -- meaning the SCALAR PATH WORKS, it is merely slower. That is the sovereign door in. ⚠NO VERIFIED RETAIL GAME SHIPS NTC OR NEURAL MATERIALS. Treat engine-side neural MATERIAL claims as unproven. === 3. GEOMETRY: VIRTUALIZED MICROPOLYGONS === Nanite reference: Karis, Stubbe, Wihlidal, SIGGRAPH 2021 Advances. The load-bearing pieces in dependency order: 1. Cluster the mesh into ~128-triangle clusters (graph partitioning). 2. Build a cluster DAG: group neighbouring clusters, simplify the GROUP as a unit, re-split into new clusters. ★THE GROUP BOUNDARY IS LOCKED DURING SIMPLIFICATION -- that is what prevents cracks, and it is the whole trick. Because groups differ between levels, locked edges migrate instead of accumulating. 3. Monotonic error metric: a parent's error must be >= its children's, so ANY view-dependent cut through the DAG is a valid crack-free LOD, selectable per-cluster on the GPU with no CPU involvement and no popping between neighbours. 4. Two-pass occlusion culling against a hierarchical Z-buffer built from last frame's visible set. 5. Software rasterisation in a compute shader with 64-bit atomics (depth in the high bits, payload in the low) for sub-pixel triangles -- hardware raster loses badly below about one pixel per triangle because it is quad-based. 6. Visibility buffer: write only {instanceID, triangleID}; resolve attributes and shade deferred, with a material classification pass. ★★★MOST OF THE HARD HALF IS NOW A PERMISSIVELY-LICENSED C LIBRARY: meshoptimizer (github.com/zeux/meshoptimizer) ships meshopt_buildMeshlets / buildMeshletsSpatial, meshopt_partitionClusters (explicitly for Nanite-style hierarchical LOD), meshopt_simplify with meshopt_SimplifyLockBorder (the locked-boundary simplification), meshopt_SimplifySparse, and meshopt_computeMeshletBounds for sphere + normal-cone culling. 2026 STRUCTURAL CHANGE -- DXR 2.0 STANDARDISES CLUSTERED GEOMETRY FOR RAY TRACING. Spec v0.30, 2026-07-10: CLAS (Cluster Level Acceleration Structure, max 256 triangles / 256 vertices per cluster), position-invariant pre-baked cluster templates, Cluster BLAS, Partitioned TLAS, GPU-driven ExecuteIndirectRTASOperations(), and a ClusterID() HLSL intrinsic (needed because primitive indices are no longer unique within a BLAS). Microsoft states it "can work well on any hardware with existing raytracing support given a driver update" -- i.e. NVIDIA's RTX Mega Geometry becoming a cross-vendor standard. ⛔DO NOT PLAN A GEOMETRY PIPELINE AROUND WORK-GRAPH MESH NODES IN 2026. Spec v1.012 (2026-02-04): mesh nodes are proposed but unsupported, and Draw / DrawIndexed nodes were CUT "in favor of starting experimentation with mesh nodes only". Compute-only work graphs are specified. UE 5.7 (Nov 2025) added Nanite Foliage (experimental, built on Nanite Voxels), Nanite Assemblies and Nanite Skinning; Substrate and PCG went Production-Ready. === 4. REAL-TIME GLOBAL ILLUMINATION === LUMEN, the budget to beat: Epic quality is about 8 ms at 1080p internal on next-gen consoles for GI AND reflections combined, relying on temporal super resolution to reach near-4K. High targets 60 FPS. Architecture: screen traces first, falling back to software (mesh SDF + merged global SDF) or hardware ray tracing; a Surface Cache parameterises each mesh from 12 capture cards by default, updated amortised across frames, which is what makes multi-bounce cheap. Epic deprecated software-RT detail tracing in 5.6 and is converging on a single hardware-RT path at 60 Hz. UE 5.7: Lumen High saves about 0.5 ms on console at 1080p; MegaLights saves 0.1-0.2 ms via vectorised shading. ReSTIR, the 2020 paper (Bitterli, Wyman, Pharr, Shirley, Lefohn, Jarosz): 6x-60x faster than prior art unbiased, 35x-65x biased; Zero Day 15 ms and Amusement Park 50 ms at 1920x1080; up to 3.4 MILLION dynamic emissive triangles at 8 rays per pixel or fewer. The algorithm in four moves: (a) RIS -- draw M cheap candidates, resample one proportional to an unnormalised target; (b) weighted reservoir sampling so state stays O(1); (c) temporal reuse against the reprojected previous-frame reservoir; (d) spatial reuse against neighbours. ★THE CORRECTNESS OF (c) AND (d) LIVES ENTIRELY IN THE MIS WEIGHTS. Naive reuse is BIASED because neighbours have different integration domains -- different visibility, different geometric terms. Budget real time for getting the unbiased weights right. ★★★THE 2026 DIRECTION IS RADIANCE CACHING, AND THE EVIDENCE IS UNAMBIGUOUS: RTXGI 2.0 = NRC + SHaRC, and DDGI HAS BEEN MOVED OUT to a legacy v1.x repo. SHaRC runs on any DXR GPU; NRC needs Turing+. AMD shipped FSR Radiance Caching (Preview) in FidelityFX SDK 2.3.0. EA SEED presented ORCA (Online Radiance Cache Acceleration) at SIGGRAPH 2026. The industry moved from probe-based DDGI to hashed / neural radiance caches. Other shipping stacks: RTXDI 3.0.0 (ReSTIR DI, GI v2.0, PT v3.0, D3D12 + Vulkan); RTX Path Tracing SDK 1.8.1 (claims 25-30 percent gain over 1.7.x at comparable or better quality). ★READ THE OPEN-SOURCE REFERENCE FIRST: Bevy Solari (Bevy 0.17, 2025-09-30) implements ReSTIR DI + ReSTIR GI + a world-space irradiance cache + DLSS Ray Reconstruction denoising, in the open, in Rust/WGSL. Explicitly experimental, Vulkan-only. Closest readable reference implementation of the modern stack. Failure modes to instrument from day one: bias (energy gain/loss -- validate against a reference path tracer on a fixed scene); BOILING (reservoir variance under temporal reuse -- cap M); ghosting on disocclusion; light leaking through thin geometry. === 5. RADIANCE FIELDS: 3D GAUSSIAN SPLATTING === COMPRESSION IS SOLVED. Mip-NeRF 360 leaderboard (3DGS Compression Survey, Eurographics 2025 STAR): method PSNR SSIM LPIPS size MB Scaffold-GS 27.50 0.806 0.252 156.0 Compact3DGS 27.08 0.798 0.247 48.8 LightGaussian 27.28 0.805 0.243 42.0 Self-Organizing Gaussians 27.08 0.799 0.230 40.3 HAC (high-rate) 27.77 0.811 0.230 22.9 ContextGS (high-rate) 27.75 0.811 0.231 12.4 SmolGS-large 27.86 0.812 0.230 10.4 The frontier method is simultaneously HIGHER QUALITY and about 15x SMALLER than Scaffold-GS. HAC++ reaches 30.34 dB at 5.5 MB on Deep Blending. Anyone shipping raw .ply splats in 2026 is two years behind. RENDERING: 3DGRUT (NVIDIA). On RTX 5090, NeRF Synthetic: 3DGRT (ray tracing) 33.87 PSNR 347 FPS 479.3 s train 3DGUT (unscented transform, rast) 33.88 PSNR 846 FPS 214.6 s train ★3DGUT MATCHES RAY-TRACED QUALITY AT 2.4x THE FRAMERATE AND 2.2x FASTER TRAINING, and both support fisheye and rolling-shutter cameras, which vanilla 3DGS CANNOT because its EWA projection assumes a pinhole model. === 6. UPSCALING AND FRAME GENERATION === DLSS 4 (shipping since Jan 2025): transformer super-resolution with 2x the parameters of the CNN model, using self-attention across the whole frame and over multiple frames. Available on ALL RTX GPUs. Multi Frame Generation emits up to 3 generated frames per rendered frame; the new FG model is 40 percent faster using 30 percent less VRAM. Blackwell adds hardware Flip Metering. MFG is RTX 50-exclusive. DLSS 5 (ANNOUNCED, ships fall 2026): a COMPACT DIFFUSION TRANSFORMER run AFTER conventional rendering, taking game colour plus motion vectors and infusing photoreal lighting anchored to the source 3D content. Three selectable models, per-object artist controls, 4K (about 8M pixels) inside a 16 ms budget, causal one-frame-in / one-frame-out, single GPU. ⛔NOT REPLICABLE BY AN INDEPENDENT TEAM -- track it, do not chase it. FSR 4 / Redstone (shipping): FSR 4.1.1 (June 2026) added RDNA 3 support; requires SM 6.6, auto-falls back to FSR 3.1.5 below that. ★THE BUDGET TARGETS: RX 9070 XT about 1.3 ms at 4K; RX 7900 XTX about 3.1 ms at 4K; RX 7800 XT about 2.1 ms at 1440p. A TEMPORAL UPSCALER MUST LAND IN THE 1-3 ms RANGE AT 4K OR IT IS NOT WORTH SHIPPING. ⚠AMD's FSR 4 model architecture (CNN vs transformer, FP8 vs INT8) is NOT DISCLOSED. === 7. RANKED ROADMAP FOR A SOVEREIGN TEAM === Assumed starting point: a software rasteriser, PBR/GGX, z-buffer, texturing, AA, SSAO, skinning, scene graph, Cornell-box radiosity. No neural rendering, no virtualized geometry, no real-time GI, no splats. TIER 1 -- do these, in this order. 1. TEMPORAL UPSAMPLING (TAAU). ~3-4 weeks. HIGHEST RATIO ON THE LIST -- and not because upscaling is glamorous. ★IT IS THE SUBSTRATE EVERY OTHER ITEM ASSUMES: ReSTIR at 1 spp, neural denoising and all stochastic effects are designed around a temporal reconstruction pass existing downstream. Building GI first means building GI twice. It also forces motion vectors to be correct across skinning and alpha-test, which is needed regardless. The five load-bearing pieces: (a) sub-pixel jitter with Halton(2,3), phase count scaled to the upscale factor; (b) motion vectors for EVERYTHING including skinned and alpha-tested geometry -- ★MISSING OR WRONG MOTION VECTORS ARE THE NUMBER ONE CAUSE OF A BAD TAAU, AHEAD OF THE RESOLVE FILTER; (c) history reprojection with Catmull-Rom, not bilinear -- bilinear history is the main softness source; (d) neighbourhood clamping in YCoCg, building a colour AABB from the 3x3 neighbourhood and clipping history toward the current colour, which kills most ghosting; (e) disocclusion detection via depth/motion divergence resetting history locally. 2. VISIBILITY BUFFER + DEFERRED MATERIALS. ~3-4 weeks. Do it BEFORE any geometry work. Independently valuable, decouples raster cost from shading cost, and is the prerequisite for virtualized geometry. For a team that already owns the rasteriser this is a SIMPLIFICATION of existing code -- write {instID, triID} instead of shading. 3. BVH + ReSTIR DI + A-trous denoise. ~2-3 months. The biggest single capability jump available. Published bar: 15 ms at 1080p with 3.4M emissive triangles at <=8 rays/pixel -- six-year-old numbers, so a modern implementation should beat them comfortably. TIER 2 -- high value, larger commitment. 4. MESHLET CLUSTER-LOD DAG (Nanite-lite). ~2-3 months, gated on item 2. Ship in stages: meshlets + cone culling, then cluster DAG + GPU cut, then HZB two-pass, then the 64-bit-atomic compute rasteriser LAST (it is an optimisation, not a requirement). 5. FORWARD-ONLY 3D GAUSSIAN SPLAT RENDERER. ~1-3 weeks. ★THE CHEAPEST CREDIBLE CHECKBOX ON THE LIST, because a tile-based software rasteriser is the hard prerequisite and we already own it. Load pretrained .ply; do NOT build training. The only real work is (i) projecting the 3D covariance to 2D via the Jacobian of the perspective projection, Sigma' = J W Sigma W^T J^T, plus a small isotropic screen-space low-pass term -- this is the one piece of real maths and the one people get wrong -- and (ii) a global radix sort by (tileID, depth), which is both the performance and the correctness bottleneck. Then front-to-back alpha compositing per tile with early termination when transmittance drops below about 1e-4. 6. WORLD-SPACE HASH-GRID RADIANCE CACHE. ~1 month, after item 3. Simpler than ReSTIR GI, degrades gracefully, and matches where the industry actually went. TIER 3 -- deliberate bounded bets. 7. NEURAL TEXTURE COMPRESSION, DECODE-ON-LOAD PATH. ~1 month. Transcode-to-BCn at load runs on ANY Shader Model 6 GPU -- no cooperative vectors, no LinAlg, no vendor lock. Target about 5 bits/texel at 40+ dB PSNR. A real measurable neural-rendering result with almost no API risk. 8. TINY-MLP NEURAL MATERIAL IN A COMPUTE SHADER. ~1-2 months. Use SM 6.9 LONG VECTORS, not Cooperative Vectors. ★THE INPUT ENCODING IS THE LOAD-BEARING PART, NOT THE NETWORK: a multiresolution latent/feature grid plus frequency encoding, with the material parameterised in Rusinkiewicz (half-vector/difference-vector) coordinates. Typical shape is 2-4 hidden layers, 32-64 wide. Train offline, ship weights as a constant buffer. ⛔DO NOT BUILD: anything on Cooperative Vectors (deprecated 2026-02-26); work-graph mesh nodes (unsupported, draw nodes cut); FRAME GENERATION -- worst effort-to-differentiation ratio here, it costs about one frame of latency BY CONSTRUCTION and its entire value is artefact quality and pacing tuning. ★ASYNC REPROJECTION / SPACE WARP IS THE TRACTABLE COUSIN because it REDUCES latency rather than adding it. === 8. EXPLICITLY UNVERIFIED -- DO NOT PROPAGATE AS FACT === Cross-vendor Vulkan cooperative-vector / LinAlg equivalent (Khronos registry returned 403). Whether NTC or neural materials ship in any retail game -- no evidence either way was found. Nanite competitor survey (id Tech 8, Snowdrop, Decima, Frostbite) -- not researched. Path-traced shipping games 2025-26 with measured performance -- not researched. 3DGS adoption in production engines and splat format standardisation -- not researched. Intel XeSS 2/3 and Microsoft DirectSR status -- not verified. ★This field carries a heavy layer of AI-generated SEO content that invents products and dates. Every source cited here is a vendor engineering blog, a Khronos or Microsoft spec, an SDK repository, official engine documentation, a peer-reviewed conference page, or a named practitioner blog. license_tier: ORIGINAL -- Nishi-authored survey and ranking; no third-party text reproduced.