25-post thread, 0 media, 2023-12-15. HIGHEST engagement 2023 thread:
112 likes, 25 reposts, 31576 views. 'Holiday GPU thoughts as a
recovering pc/mobile/etc cross platform dev-holic' - the 16-bit
permutation philosophy + post-black-box philosophy in one thread.
[0] STP prototypes: 32-bit multi-dispatch vs packed-16 single-
dispatch-ubershader, 40% faster, 'Industry has huge untapped opt
potential.' [1-7] 16-bit fundamentals: AMD Vega+ for PC, draw
compatibility line at 16-bit, dev on AMD VK RDNA2 + avoid DXC
deoptimizer, Gather4 SoA, range management, denormals, Touch It Once
is top optimization, no pass-graph fragmentation. [8-12] TAA
scaling: 720p/1080p rendering for 4k, large L3 holds full render
targets, RDNA2's 128 MiB L3 was spot on, RDNA3 drop was wrong,
console no-L3 was wrong. [13-17] Triangles obsolete: 8x area scaling
doesn't know connectivity, scaling TAA = mostly tri culling, fix
geometrical aliasing to break black boxes, stratified sampling 2x
geo density, frame viewport jitter EOL, TAA disocclusion deeply
integrated into shading. [18-23] ML/GI/RT critique: per-pixel ML
people wrong, multi-pass vs fewer-passes, GI needs surface shade cache
-> object-space shading -> NUMA, HW RT people wrong (non-ray-traversal
ordered access, occlusion = neighbor-coherent, GI = high-freq
occlusion of low-freq probe domain, HW RT skin-all re-tree won't
scale, stratified visibility = bounded costs). [24] Vote with
engineering: general-purpose CS, say no to black boxes. [25] Scaling
TAA wants sparse striped data per frame, low-freq pixel control
cage displacing high-freq reprojected feedback.
