[{"data":1,"prerenderedAt":157},["ShallowReactive",2],{"run:s3-smoke-1000-depth2":3,"run-findings:s3-smoke-1000-depth2":122},{"run":4,"config":17,"metrics":80,"curve":86,"assets":121},{"id":5,"title":5,"subtitle":6,"eyebrow":7,"date":8,"workspace":9,"commit":10,"tags":11,"verdict":16},"s3-smoke-1000-depth2","S3 smoke arm B (pipelining = depth2): Lego 1k、profile=true で busy% \u002F sync-tax を取る","Run summary · 2026-08 1× 工程表 v2 — S3 (iteration pipelining depth 2) gate","2026-08-16","splat","7e7f875",[12,13,14,15],"p2-s3-iteration-pipelining","s3","gpu-gate","lego","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","1,000",{"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},19.597320556640625,"4.62s",false,6826,"6.568347e-2",{"loss":87,"splats":112},{"iters":88,"values":100},[89,90,91,92,93,94,95,96,97,98,99],1,100,200,300,400,500,600,700,800,900,1000,[101,102,103,104,105,106,107,108,109,110,111],0.593845009803772,0.22888724505901337,0.1168643981218338,0.10670866072177887,0.10142696648836136,0.09738743305206299,0.0876012071967125,0.08004747331142426,0.0743214339017868,0.06973379105329514,0.06568346917629242,{"iters":113,"values":114},[89,90,91,92,93,94,95,96,97,98,99],[115,115,115,115,115,116,117,118,119,120,84],5207,599,960,1598,2704,4343,{"has_renders":83,"has_splat":83},[123],{"id":12,"title":124,"date":125,"status":126,"polarity":127,"category":128,"axes":129,"tags":131,"task_code":140,"related_runs":141,"delta_psnr":151,"delta_wallclock":152,"rank":153,"verdict":154,"impact_summary":155,"detail_path":156},"S3 iteration pipelining depth 2 — 12.5k -18.1% \u002F busy 78 → 91% (default flip)","2026-08-17","stable","positive","experiment",[89,130],3,[132,13,133,134,135,136,137,138,15,139],"1x-roadmap-v2","iteration-pipelining","command-buffer","gpu-busy","sync-tax","double-buffering","scratch-generation","unified-memory","2026-08 1× 工程表 v2 — AC 復帰後 計測 queue #2' (S3 gate)",[142,5,143,144,145,146,147,148,149,150],"s3-smoke-1000-off","s3-12500-off-prof","s3-12500-depth2-prof","s3-12500-off-r1","s3-12500-off-r2","s3-12500-off-r3","s3-12500-depth2-r1","s3-12500-depth2-r2","s3-12500-depth2-r3","12.5k **同一 session 交互 A\u002FB** n=3 mean: depth2 **36.202 dB** (36.195 \u002F 36.179 \u002F 36.231) vs off **36.185 dB** (36.178 \u002F 36.192 \u002F 36.186) = **+0.016 dB** (guard 36.167 ± 0.1 の内側、n=3 mean 判定 — n=1 判定なし)。final splats は depth2 406,439 \u002F off 408,874 で **depth2 のほうが 0.6% 少ない** (前段の profile pair では逆に +0.6% で、方向が一致しない = 既存 atomic 非決定性の chaos 増幅の範囲)","12.5k full run **同一 session 交互 A\u002FB n=3 mean: 185.21 → 151.70s = -18.1%** (train elapsed)。process wallclock では **188.33 → 154.00s = -18.2%**。σ は off 3.33% \u002F depth2 1.18% なので有意閾値 max(2σ,1%) = 6.7%、実測 Δ は**その 2.7 倍**。同一 session の off arm は S2 gate 当日の S2 arm (train 184.61s \u002F process 187.65s) を **+0.3% で再現**しており、session の健全性が確認できている。参考: profile=true の 1 対 1 比較では 180.49 → 162.72s = -9.85% (計装自体が host 側 critical path に乗るため S3 の効果を過小評価する)","high","s3-adopted-default-flip-pipelining-depth2","1× 工程表 v2 が「S2 後の単独本命」に指定していた S3 を gate し、採用した。S2 までの施策が \u003Cstrong>CB の数\u003C\u002Fstrong>を減らすものだったのに対し、S3 は \u003Cstrong>CB の並び方\u003C\u002Fstrong>を変える — iteration N の loss\u002Fbackward CB (B\u003Csub>N\u003C\u002Fsub>) を commit したら wait せずに iteration N+1 の forward CB (F\u003Csub>N+1\u003C\u002Fsub>) を組んで commit し、\u003Ccode>wait(B\u003Csub>N\u003C\u002Fsub>)\u003C\u002Fcode> の後の host readback \u002F \u003Ccode>accumulate\u003C\u002Fcode> を F\u003Csub>N+1\u003C\u002Fsub> の GPU 実行の影に置く。\u003Cstrong>12.5k 同一 session 交互 A\u002FB n=3 で 185.21 → 151.70s = -18.1%\u003C\u002Fstrong> (閾値 max(2σ,1%) = 6.7% の 2.7 倍)。\u003Cstrong>GPU busy 77.5 → 91.3% \u002F sync-tax 3.244 → 1.136 ms\u002Fiter (-65%)\u003C\u002Fstrong> で counter session が置いた「busy 100% = 理論上限」への距離は 22.5 pt → 8.7 pt に縮んだ。\u003Cstrong>1k smoke では busy 65.8 → 72.0% にしか上がらず判定不能だった\u003C\u002Fstrong> — smoke の splat 数 6.8k では forward の GPU 影 (1.06 ms\u002Fiter) が image size 固定の host 仕事 (~1.7 ms\u002Fiter) を隠しきれないためで、これは \u003Cstrong>「pipelining の gate は smoke では代替できない」という harness 上の教訓\u003C\u002Fstrong>として残す。残 bubble の内訳も確定した: \u003Cstrong>refine tick の drain 6.04s (43%) > encode(B) 4.66s (33%) > 残差 3.28s (23%)\u003C\u002Fstrong> — 次の一手が roadmap の「即効ペア (refine capacity スコープバグ修正)」であることが実測で決まった。msplat 同日等価 gap \u003Cstrong>1.105× → 0.907×\u003C\u002Fstrong>、Phase 0 からの累計 766.07 → 154.0s = \u003Cstrong>-79.9%\u003C\u002Fstrong> (全て数学的同値の施策のみ)。","\u002Ffindings\u002Fp2-s3-iteration-pipelining\u002F",1790227325628]