Add B200 v-kmeans vsmooth ablation results to nunchux_low_bit_attention_blackwell.html

blackwell_attention_slides/nunchux_low_bit_attention_blackwell.html

@@ -312,12 +312,23 @@ prmt.b32 P,  p0, p1, 0x5410;</pre>
 
 <section class="slide">
   <div class="eyebrow">B200 / v-smooth ablation</div>
-  <h2>Workload-level Error Only</h2>
-  <div class="placeholder">
-    Reserved for B200 workload-level attention metrics.<br><br>
-    Fill with per-layer latency, speedup, RMSE, and cosine after the rerun.
+  <h2>V-kmeans + V-smooth: Stable Accuracy Gain</h2>
+  <table>
+    <thead>
+      <tr><th>Config</th><th>Latency</th><th>Speedup vs bf16</th><th>TFLOPs</th><th>RMSE</th><th>Cosine</th></tr>
+    </thead>
+    <tbody>
+      <tr><td>bf16 FA4</td><td>68.793 ms</td><td>1.000x</td><td>1422.4</td><td>0</td><td>1.000000</td></tr>
+      <tr><td>no v-kmeans / no v-smooth</td><td><span class="red">51.800 ms</span></td><td><span class="red">1.329x</span></td><td><span class="red">1890.3</span></td><td>8.704e-3</td><td>0.999918</td></tr>
+      <tr><td>v-kmeans + v-smooth</td><td>57.506 ms</td><td>1.197x</td><td>1702.7</td><td><span class="red">7.997e-3</span></td><td><span class="red">0.999928</span></td></tr>
+    </tbody>
+  </table>
+  <div class="metric">
+    <div class="box"><div class="num">1.197x</div><div class="label">still faster than bf16 FA4 with v-kmeans + v-smooth</div></div>
+    <div class="box"><div class="num">8.1%</div><div class="label">RMSE reduction: 8.704e-3 -> 7.997e-3</div></div>
+    <div class="box"><div class="num">1.110x</div><div class="label">latency cost vs no v-kmeans / no v-smooth</div></div>
   </div>
-  <p class="math-note">Target setting: Wan2.1 single-layer workloads, QK-INT8 with Hadamard + Q per-token + K block-128, PV-FP8 per-channel, exp approximation ON (linmant), compare v-smooth OFF vs ON against bf16 FA4 attention output.</p>
+  <p class="math-note">B200, full Wan2.1 p1 set: 18 attention-only workloads. QK-INT8 with Hadamard + Q per-token + K block-128, PV-FP8, exp approximation ON (linmant). Metrics are vs bf16 FA4 attention output.</p>
 </section>
 
 <section class="slide">