[{"data":1,"prerenderedAt":173},["ShallowReactive",2],{"run:final-parity-12500-r2":3,"run-findings:final-parity-12500-r2":135},{"run":4,"config":17,"metrics":80,"curve":86,"assets":134},{"id":5,"title":5,"subtitle":6,"eyebrow":7,"date":8,"workspace":9,"commit":10,"tags":11,"verdict":16},"final-parity-12500-r2","msplat parity (最終スタック) — Lego 12500 iter、profile=false","Run summary · 2026-08 msplat 等価比較 最終スタック (matched 断面込み)","2026-08-17","splat","9f16280",[12,13,14,15],"p2-msplat-parity-final","msplat-parity","final-stack","lego-12500","partial",[18,21,24,27,30,33,36,39,42,45,48,51,54,57,60,63,66,69,72,75,77],{"key":19,"value":20},"dataset","\u002FUsers\u002Fotkrickey\u002Fdev\u002F3dgs-workspace\u002Fdatasets\u002Fnerf_synthetic\u002Flego",{"key":22,"value":23},"gt_convention","Premultiplied",{"key":25,"value":26},"iterations","12,500",{"key":28,"value":29},"seed","42",{"key":31,"value":32},"capacity","1,000,000 splats",{"key":34,"value":35},"sh_degree","3",{"key":37,"value":38},"loss","L1Ssim",{"key":40,"value":41},"lambda","0.200",{"key":43,"value":44},"ssim","window=7 sigma=1",{"key":46,"value":47},"backend.backward","Simd",{"key":49,"value":50},"backend.loss_path","Gpu",{"key":52,"value":53},"backend.emit_pairs","Original",{"key":55,"value":56},"backend.forward","F32",{"key":58,"value":59},"backend.dispatch","PerStage",{"key":61,"value":62},"backend.radix_scan","GpuHierarchical",{"key":64,"value":65},"backend.binning","Resident",{"key":67,"value":68},"backend.tile_sort","GlobalRadix",{"key":70,"value":71},"backend.submission","Chained",{"key":73,"value":74},"backend.target_cache","true",{"key":76,"value":74},"backend.shader_fast_math",{"key":78,"value":79},"os_build","26.5.2 (25F84)",{"psnr":81,"wallclock":82,"wallclock_regress":83,"splats":84,"final_loss":85},36.138099670410156,"2m 32s",false,408111,"5.556727e-3",{"loss":87,"splats":118},{"iters":88,"values":103},[89,90,91,92,93,94,95,96,97,98,99,100,101,102],1,1000,2000,3000,4000,5000,6000,7000,8000,9000,10000,11000,12000,12500,[104,105,106,107,108,109,110,111,112,113,114,115,116,117],0.5938453674316406,0.05447784811258316,0.014408562332391739,0.008355962112545967,0.007251642644405365,0.006458763964474201,0.006039143539965153,0.00591091625392437,0.005777014885097742,0.00558779202401638,0.005536865442991257,0.0055298833176493645,0.005409435369074345,0.005556726828217506,{"iters":119,"values":120},[89,90,91,92,93,94,95,96,97,98,99,100,101,102],[121,122,123,124,125,126,127,128,129,130,131,132,133,84],5207,8232,182122,298988,350854,376411,390649,398275,402174,404261,405028,405835,407228,{"has_renders":83,"has_splat":83},[136],{"id":12,"title":137,"date":8,"status":138,"polarity":139,"category":140,"axes":141,"tags":144,"task_code":157,"related_runs":158,"delta_psnr":167,"delta_wallclock":168,"rank":169,"verdict":170,"impact_summary":171,"detail_path":172},"最終スタック msplat 等価比較 — wallclock 0.927× \u002F throughput 1.135× \u002F energy 0.906×、matched 断面は convention で割れる","stable","mixed","experiment",[89,142,143],2,3,[145,146,147,148,149,150,151,152,153,154,155,156],"1x-roadmap-v2","msplat","parity","baseline","cross-eval","eval-convention","thermal-interleave","energy","throughput","lego","m4-max","thesis-ready","2026-08 1× 工程表 v2 — 最終段 (msplat 等価比較 matched 断面込み)",[159,5,160,161,162,163,164,165,166],"final-parity-12500-r1","final-parity-12500-r3","final-parity-12500-r4","final-parity-12500-r5","final-parity-13300","final-parity-6000","final-parity-30000","final-parity-30000-r2","brush convention (同一 evaluator q8、msplat 30K best 34.243 dB 基準): 本実装 12.5k **+1.97 dB** \u002F 13.3k (matched-wallclock) **+1.94 dB** \u002F 6.0k **+0.67 dB** \u002F 30k **+2.26 dB**。white convention (msplat 30K 最小 36.252 dB 基準): 12.5k **-1.10 dB** \u002F 13.3k **-1.17 dB** \u002F 30k **-0.81 dB** — **matched 点は 30k まで延ばしても存在しない**","12.5k 152.68s vs msplat 30K 164.69s = **0.927×** (train elapsed、**block 1 interleave の n=2 mean 同士**)。msplat の characterization 値は n=4 mean **165.42s** だが、gap の分母には interleave pair である block 1 n=2 (164.69s) を使う (n=4 を分母にすると 0.923×)。process wallclock では 155s vs 168.5s = **0.920×**。matched-wallclock は 13,300 iter = 165.85s (msplat block1 mean 164.69s の **1.007×**)。matched-PSNR (brush) は 6,000 iter = 51.60s = **0.313×**。matched-iteration (30K vs 30K、co-located anchor) は 479.91s vs 166.15s = **2.888×**。前回 (2026-08-16) の 2.10× から反転","high","wallclock-and-energy-won-psnr-convention-split","1× 工程表 v2 の最終段を通し、卒論 §5 の direct comparison を確定させた。\u003Cstrong>本実装 Lego 12.5k = 152.68s \u002F msplat 30K = 164.69s (block 1 interleave pair の n=2 mean 同士) で wallclock gap 0.927×\u003C\u002Fstrong> — 2026-08-16 の同一 protocol 実測 2.10× から \u003Cstrong>2.27 倍の改善\u003C\u002Fstrong>で、初めて convention 非依存の軸で msplat を下回った。\u003Cstrong>per-splat·iter throughput も 2.62× → 1.135×\u003C\u002Fstrong> と near-parity まで詰まっており、「本実装は workload 形状 (少 iter × 多 splats) で勝っているだけで kernel は遅い」という 2026-08-16 時点の解釈は\u003Cstrong>もはや成立しない\u003C\u002Fstrong>。エネルギーも \u003Cstrong>GPU 0.906× \u002F package 0.896×\u003C\u002Fstrong> と反転したが、機構は counter session (低 W × 長時間) から \u003Cstrong>同 W × 短時間\u003C\u002Fstrong>へ完全に入れ替わっている (W 比 0.704× → 0.978×) — S3 が GPU busy を 77.5 → 91.3% に上げた以上、低 W は原理的に維持できない。\u003Cstrong>一方 PSNR の matched 断面は convention 依存のまま\u003C\u002Fstrong>: brush では 6.0k (0.313×) で msplat 30K を上回り、white では 30k (2.888×) でも -0.81 dB。roadmap v2 が「msplat native 比 -0.34 dB」と留保していた差は、\u003Cstrong>30k まで延ばしても self-eval が 36.455 dB で天井\u003C\u002Fstrong>になることから \u003Cstrong>iter 予算の不足ではなく評価 convention の差\u003C\u002Fstrong>と確定した (Phase H の 36.384 dB 天井とも整合)。よって当該留保は「解消」ではなく\u003Cstrong>撤回\u003C\u002Fstrong>し、convention 別の 2 断面に置き換える。","\u002Ffindings\u002Fp2-msplat-parity-final\u002F",1790227323415]