diff --git a/data_dashboard/app.js b/data_dashboard/app.js
new file mode 100644
index 0000000..9fb46fb
--- /dev/null
+++ b/data_dashboard/app.js
@@ -0,0 +1,218 @@
+const kernelSelect = document.getElementById('kernelSelect');
+const metricSelect = document.getElementById('metricSelect');
+const tableHead = document.getElementById('tableHead');
+const tableBody = document.getElementById('tableBody');
+const avgSpeedup = document.getElementById('avgSpeedup');
+const maxSpeedup = document.getElementById('maxSpeedup');
+const darkModeToggle = document.getElementById('darkModeToggle');
+const modeIcon = document.getElementById('modeIcon');
+const modeText = document.getElementById('modeText');
+const downloadBtn = document.getElementById('downloadBtn');
+
+// Dark mode functionality
+function initDarkMode() {
+    const savedMode = localStorage.getItem('darkMode');
+    if (savedMode === 'true') {
+        document.body.classList.add('dark-mode');
+        modeIcon.textContent = '☀️';
+        modeText.textContent = 'Light';
+    }
+}
+
+darkModeToggle.addEventListener('click', () => {
+    document.body.classList.toggle('dark-mode');
+    const isDark = document.body.classList.contains('dark-mode');
+    localStorage.setItem('darkMode', isDark);
+    modeIcon.textContent = isDark ? '☀️' : '🌙';
+    modeText.textContent = isDark ? 'Light' : 'Dark';
+});
+
+// CSV Download functionality
+downloadBtn.addEventListener('click', () => {
+    let csvContent = "Kernel,Metric,Configuration,CUDA,Triton,Helion\n";
+
+    for (const [kernelKey, kernelData] of Object.entries(DATA)) {
+        const kernelName = kernelKey.replace(/_/g, ' ').replace(/\b\w/g, l => l.toUpperCase());
+
+        for (const [metricKey, metricData] of Object.entries(kernelData.metrics)) {
+            const metricName = metricData.name;
+
+            kernelData.configs.forEach((config, idx) => {
+                let configName = config.name;
+                if (config.detail) {
+                    configName += ` (${config.detail})`;
+                }
+
+                const cudaVal = metricData.cuda[idx];
+                const tritonVal = metricData.triton[idx];
+                const helionVal = metricData.helion[idx];
+
+                csvContent += `"${kernelName}","${metricName}","${configName}",${cudaVal},${tritonVal},${helionVal}\n`;
+            });
+        }
+    }
+
+    const blob = new Blob([csvContent], { type: 'text/csv;charset=utf-8;' });
+    const link = document.createElement('a');
+    link.href = URL.createObjectURL(blob);
+    link.download = 'gpu_benchmark_complete_data.csv';
+    link.click();
+});
+
+function populateMetrics() {
+    const kernel = kernelSelect.value;
+    const metrics = DATA[kernel].metrics;
+
+    metricSelect.innerHTML = '';
+    Object.keys(metrics).forEach(metricKey => {
+        const option = document.createElement('option');
+        option.value = metricKey;
+        option.textContent = metrics[metricKey].name;
+        metricSelect.appendChild(option);
+    });
+}
+
+function calculateCellClass(value, values, lowerIsBetter) {
+    const sorted = [...values].sort((a, b) => lowerIsBetter ? a - b : b - a);
+    const best = sorted[0];
+    const second = sorted[1];
+    if (value === best) return 'best';
+    if (value === second) return 'second';
+    return 'worst';
+}
+
+function formatValue(value) {
+    if (value >= 1e6) return value.toExponential(2);
+    if (value >= 1000) return value.toFixed(0);
+    if (value >= 10) return value.toFixed(1);
+    if (value >= 1) return value.toFixed(2);
+    return value.toFixed(4);
+}
+
+function calculateSpeedup(values, lowerIsBetter) {
+    const sorted = [...values].sort((a, b) => lowerIsBetter ? a - b : b - a);
+    return lowerIsBetter ? sorted[2] / sorted[0] : sorted[0] / sorted[2];
+}
+
+function getWinner(values, lowerIsBetter) {
+    const sorted = [...values].sort((a, b) => lowerIsBetter ? a - b : b - a);
+    const best = sorted[0];
+    if (values[0] === best) return 'CUDA';
+    if (values[1] === best) return 'Triton';
+    return 'Helion';
+}
+
+function renderTable() {
+    const kernel = kernelSelect.value;
+    const metric = metricSelect.value;
+    const kernelData = DATA[kernel];
+    const metricData = kernelData.metrics[metric];
+
+    // Extract unit from metric name
+    const metricName = metricData.name;
+    let unit = '';
+    const unitMatch = metricName.match(/\(([^)]+)\)/);
+    if (unitMatch) {
+        unit = unitMatch[1];
+    } else {
+        // For metrics without explicit units in parentheses
+        if (metric === 'registers_per_thread') {
+            unit = 'registers';
+        } else if (metric === 'register_limited_blocks') {
+            unit = 'blocks';
+        } else if (metric === 'gflops') {
+            unit = 'GFLOPs';
+        }
+    }
+
+    // Only show speedup for duration metric
+    const isDuration = metric === 'duration';
+
+    let speedups = [];
+    let winners = { CUDA: 0, Triton: 0, Helion: 0 };
+
+    // Build header with conditional speedup column and units
+    tableHead.innerHTML = `
+        <tr>
+            <th>Configuration</th>
+            <th>
+                <span class="header-impl">CUDA</span>
+                ${unit ? `<span class="header-unit">${unit}</span>` : ''}
+            </th>
+            <th>
+                <span class="header-impl">Triton</span>
+                ${unit ? `<span class="header-unit">${unit}</span>` : ''}
+            </th>
+            <th>
+                <span class="header-impl">Helion</span>
+                ${unit ? `<span class="header-unit">${unit}</span>` : ''}
+            </th>
+            ${isDuration ? '<th><span class="header-impl">Speedup</span><span class="header-unit">×</span></th>' : ''}
+        </tr>
+    `;
+
+    tableBody.innerHTML = '';
+    kernelData.configs.forEach((config, idx) => {
+        const cuda = metricData.cuda[idx];
+        const triton = metricData.triton[idx];
+        const helion = metricData.helion[idx];
+        const values = [cuda, triton, helion];
+
+        const cudaClass = calculateCellClass(cuda, values, metricData.lower_is_better);
+        const tritonClass = calculateCellClass(triton, values, metricData.lower_is_better);
+        const helionClass = calculateCellClass(helion, values, metricData.lower_is_better);
+
+        let speedupCell = '';
+        if (isDuration) {
+            const speedup = calculateSpeedup(values, metricData.lower_is_better);
+            speedups.push(speedup);
+
+            let speedupClass = 'speedup-medium';
+            if (speedup > 50) speedupClass = 'speedup-extreme';
+            else if (speedup > 10) speedupClass = 'speedup-high';
+            else if (speedup < 1) speedupClass = 'speedup-regression';
+
+            speedupCell = `<td><span class="speedup ${speedupClass}">${speedup.toFixed(1)}x</span></td>`;
+        }
+
+        const winner = getWinner(values, metricData.lower_is_better);
+        winners[winner]++;
+
+        const row = document.createElement('tr');
+        row.innerHTML = `
+            <td class="config-cell">
+                ${config.name}
+                ${config.detail ? `<span class="config-detail">${config.detail}</span>` : ''}
+            </td>
+            <td class="${cudaClass}">${formatValue(cuda)}</td>
+            <td class="${tritonClass}">${formatValue(triton)}</td>
+            <td class="${helionClass}">${formatValue(helion)}</td>
+            ${speedupCell}
+        `;
+        tableBody.appendChild(row);
+    });
+
+    // Update stats - show speedup stats only for duration
+    if (isDuration && speedups.length > 0) {
+        const avgSpeed = speedups.reduce((a, b) => a + b, 0) / speedups.length;
+        const maxSpeed = Math.max(...speedups);
+        avgSpeedup.textContent = avgSpeed.toFixed(1) + 'x';
+        maxSpeedup.textContent = maxSpeed.toFixed(1) + 'x';
+        avgSpeedup.parentElement.style.display = 'block';
+        maxSpeedup.parentElement.style.display = 'block';
+    } else {
+        avgSpeedup.parentElement.style.display = 'none';
+        maxSpeedup.parentElement.style.display = 'none';
+    }
+}
+
+kernelSelect.addEventListener('change', () => {
+    populateMetrics();
+    renderTable();
+});
+
+metricSelect.addEventListener('change', renderTable);
+
+initDarkMode();
+populateMetrics();
+renderTable();
diff --git a/data_dashboard/data.js b/data_dashboard/data.js
new file mode 100644
index 0000000..dc1702d
--- /dev/null
+++ b/data_dashboard/data.js
@@ -0,0 +1,2123 @@
+const DATA = {
+  "vector_addition": {
+    "configs": [
+      {
+        "name": "Config A",
+        "detail": "B=32, S=4K, H=256"
+      },
+      {
+        "name": "Config B",
+        "detail": "B=64, S=8K, H=512"
+      },
+      {
+        "name": "BF16",
+        "detail": "B=64, S=8K, H=1024"
+      },
+      {
+        "name": "FP32",
+        "detail": "B=64, S=8K, H=1024"
+      },
+      {
+        "name": "FP16",
+        "detail": "B=64, S=8K, H=1024"
+      },
+      {
+        "name": "Config G",
+        "detail": "B=48, S=12K, H=1536"
+      },
+      {
+        "name": "Config C",
+        "detail": "B=64, S=16K, H=1024"
+      },
+      {
+        "name": "FP16 Large",
+        "detail": "B=64, S=16K, H=2048"
+      },
+      {
+        "name": "Config H",
+        "detail": "B=24, S=48K, H=2048"
+      },
+      {
+        "name": "Config D",
+        "detail": "B=32, S=32K, H=2048"
+      },
+      {
+        "name": "Config E",
+        "detail": "B=16, S=65K, H=4096"
+      },
+      {
+        "name": "Config F",
+        "detail": "B=8, S=131K, H=8192"
+      }
+    ],
+    "metrics": {
+      "duration": {
+        "name": "Duration (ms)",
+        "lower_is_better": true,
+        "cuda": [
+          0.10573,
+          0.83632,
+          1.67,
+          1.95,
+          2.33,
+          2.64,
+          4.58,
+          5.1,
+          7.21,
+          8.78,
+          15.6,
+          31.2
+        ],
+        "triton": [
+          0.08842,
+          0.71251,
+          0.734,
+          1.42,
+          1.97,
+          2.41,
+          3.98,
+          2.85,
+          6.42,
+          6.71,
+          11.71,
+          23.47
+        ],
+        "helion": [
+          0.088,
+          0.70326,
+          0.71,
+          1.4,
+          1.75,
+          2.41,
+          4.09,
+          2.85,
+          6.42,
+          6.98,
+          11.94,
+          23.9
+        ]
+      },
+      "dram_throughput": {
+        "name": "DRAM Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          36.67,
+          38.92,
+          40.06,
+          66.99,
+          28.6,
+          81.78,
+          58.42,
+          50.15,
+          79.83,
+          60.96,
+          68.66,
+          68.66
+        ],
+        "triton": [
+          86.64,
+          89.38,
+          90.8,
+          89.72,
+          33.9,
+          89.67,
+          67.23,
+          89.86,
+          89.72,
+          79.73,
+          91.44,
+          91.26
+        ],
+        "helion": [
+          88.35,
+          92.36,
+          92.48,
+          92.85,
+          38.12,
+          89.58,
+          65.44,
+          89.8,
+          89.7,
+          76.63,
+          89.65,
+          89.61
+        ]
+      },
+      "compute_throughput": {
+        "name": "Compute Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          38.09,
+          38.01,
+          38.21,
+          29.39,
+          23.7,
+          18.52,
+          23.63,
+          35.75,
+          16.64,
+          26.17,
+          29.5,
+          29.34
+        ],
+        "triton": [
+          14.94,
+          14.66,
+          22.51,
+          14.55,
+          11.02,
+          7.43,
+          11.1,
+          12.73,
+          8.88,
+          11.85,
+          12.41,
+          12.33
+        ],
+        "helion": [
+          15.2,
+          14.86,
+          14.88,
+          14.78,
+          11.43,
+          7.27,
+          10.68,
+          14.55,
+          6.16,
+          8.26,
+          5.72,
+          5.71
+        ]
+      },
+      "no_eligible_warps": {
+        "name": "No Eligible Warps (%)",
+        "lower_is_better": false,
+        "cuda": [
+          61.27,
+          61.82,
+          61.72,
+          70.68,
+          63.18,
+          81.46,
+          71.18,
+          67.2,
+          83.34,
+          71.19,
+          70.5,
+          70.65
+        ],
+        "triton": [
+          89.18,
+          89.18,
+          77.34,
+          89.76,
+          82.78,
+          92.55,
+          89.78,
+          87.25,
+          91.12,
+          89.07,
+          87.67,
+          87.68
+        ],
+        "helion": [
+          89.68,
+          90.2,
+          90.21,
+          90.27,
+          86.15,
+          94.91,
+          92.32,
+          92.35,
+          93.83,
+          90.17,
+          95.92,
+          95.93
+        ]
+      }
+    }
+  },
+  "batched_matmul": {
+    "configs": [
+      {
+        "name": "Config A",
+        "detail": "B=32, S=4K, H=256"
+      },
+      {
+        "name": "Config B",
+        "detail": "B=64, S=8K, H=512"
+      },
+      {
+        "name": "FP32",
+        "detail": "B=64, S=8K, H=1024"
+      },
+      {
+        "name": "FP16",
+        "detail": "B=64, S=8K, H=1024"
+      },
+      {
+        "name": "Config G",
+        "detail": "B=48, S=12K, H=1536"
+      },
+      {
+        "name": "Config C",
+        "detail": "B=64, S=16K, H=1024"
+      },
+      {
+        "name": "FP16 Large",
+        "detail": "B=64, S=16K, H=2048"
+      },
+      {
+        "name": "Config D",
+        "detail": "B=32, S=32K, H=2048"
+      },
+      {
+        "name": "Config H",
+        "detail": "B=24, S=48K, H=2048"
+      },
+      {
+        "name": "Config E",
+        "detail": "B=16, S=65K, H=4096"
+      },
+      {
+        "name": "Config F",
+        "detail": "B=8, S=131K, H=8192"
+      }
+    ],
+    "metrics": {
+      "duration": {
+        "name": "Duration (ms)",
+        "lower_is_better": true,
+        "cuda": [
+          2.78,
+          44.19,
+          176.13,
+          174.52,
+          520.74,
+          352.45,
+          1830,
+          1410,
+          2160,
+          5590,
+          22370
+        ],
+        "triton": [
+          0.238,
+          7.95,
+          32.35,
+          2.8,
+          72.66,
+          65.47,
+          26.95,
+          236.11,
+          249.6,
+          916.01,
+          3640
+        ],
+        "helion": [
+          0.198,
+          2.44,
+          101.14,
+          2.27,
+          47.65,
+          17.83,
+          16.37,
+          69.17,
+          168.9,
+          605.19,
+          2420
+        ]
+      },
+      "aggregate_memory": {
+        "name": "Aggregate Memory Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          94.42,
+          94.57,
+          94.6,
+          95.31,
+          95.63,
+          94.6,
+          95.34,
+          94.66,
+          95.56,
+          95.28,
+          95.23
+        ],
+        "triton": [
+          68.77,
+          88.1,
+          92.42,
+          70.87,
+          88.91,
+          92.43,
+          93.9,
+          87.76,
+          88.53,
+          87.33,
+          87.53
+        ],
+        "helion": [
+          74.1,
+          58.96,
+          78.72,
+          68.59,
+          79.98,
+          60.35,
+          79.39,
+          59.81,
+          79.38,
+          79.34,
+          80.39
+        ]
+      },
+      "l1_tex_throughput": {
+        "name": "L1/TEX Cache Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          94.54,
+          94.58,
+          94.6,
+          95.32,
+          95.65,
+          94.6,
+          95.34,
+          94.66,
+          95.58,
+          95.28,
+          95.23
+        ],
+        "triton": [
+          71.85,
+          88.21,
+          92.43,
+          45.1,
+          88.57,
+          92.43,
+          68.56,
+          87.76,
+          88.57,
+          87.32,
+          87.53
+        ],
+        "helion": [
+          76.21,
+          60.06,
+          78.74,
+          66.82,
+          80.4,
+          60.7,
+          64.83,
+          59.94,
+          79.44,
+          79.22,
+          80.42
+        ]
+      },
+      "l2_throughput": {
+        "name": "L2 Cache Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          14.22,
+          14.81,
+          14.65,
+          15.12,
+          13.84,
+          14.63,
+          12.79,
+          14.26,
+          11.9,
+          27.56,
+          27.56
+        ],
+        "triton": [
+          72.5,
+          38.87,
+          50.06,
+          88.88,
+          20.12,
+          49.75,
+          93.89,
+          47.4,
+          20.12,
+          41.65,
+          37.98
+        ],
+        "helion": [
+          55.41,
+          50.75,
+          46.94,
+          68.87,
+          19.98,
+          45.22,
+          78.66,
+          43.01,
+          21.1,
+          22.15,
+          26.1
+        ]
+      },
+      "dram_throughput": {
+        "name": "DRAM Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          1.96,
+          1.04,
+          0.54,
+          0.27,
+          0.35,
+          0.52,
+          0.11,
+          0.26,
+          0.2,
+          16.47,
+          16.41
+        ],
+        "triton": [
+          59.42,
+          27.08,
+          32.89,
+          70.87,
+          2.09,
+          32.65,
+          6.73,
+          30.85,
+          2.09,
+          28.39,
+          26.93
+        ],
+        "helion": [
+          27.12,
+          26.64,
+          29.88,
+          20.96,
+          13.07,
+          25.37,
+          17.03,
+          23.69,
+          13.49,
+          14.72,
+          17.13
+        ]
+      },
+      "compute_throughput": {
+        "name": "Compute Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          73.61,
+          73.92,
+          74.05,
+          74.75,
+          74.85,
+          74.06,
+          74.79,
+          74.16,
+          74.83,
+          74.63,
+          74.61
+        ],
+        "triton": [
+          19.8,
+          15.91,
+          14.26,
+          56.34,
+          11.36,
+          14.26,
+          62.92,
+          15.05,
+          11.5,
+          14.8,
+          14.71
+        ],
+        "helion": [
+          36.37,
+          30.38,
+          53.38,
+          67.16,
+          15.98,
+          33.82,
+          72.98,
+          34.87,
+          15.94,
+          16.03,
+          16.3
+        ]
+      },
+      "no_eligible_warps": {
+        "name": "No Eligible Warps (%)",
+        "lower_is_better": false,
+        "cuda": [
+          47.7,
+          48.75,
+          49.26,
+          47.21,
+          48.03,
+          49.26,
+          46.65,
+          49.49,
+          48.3,
+          45.34,
+          45.48
+        ],
+        "triton": [
+          80.97,
+          84.07,
+          85.74,
+          76.02,
+          86.63,
+          85.74,
+          37.19,
+          84.95,
+          88.5,
+          85.18,
+          85.29
+        ],
+        "helion": [
+          76.1,
+          87.45,
+          80.06,
+          77.41,
+          99.02,
+          89.44,
+          79.01,
+          89.3,
+          96.24,
+          96.41,
+          96.43
+        ]
+      }
+    }
+  },
+  "softmax_forward": {
+    "configs": [
+      {
+        "name": "Batch (B=1)",
+        "detail": ""
+      },
+      {
+        "name": "Batch (B=8)",
+        "detail": ""
+      },
+      {
+        "name": "Batch (B=32)",
+        "detail": ""
+      },
+      {
+        "name": "Batch (B=128)",
+        "detail": ""
+      },
+      {
+        "name": "Data (32, 16K, 1024)",
+        "detail": ""
+      },
+      {
+        "name": "Data (32, 4K, 256)",
+        "detail": ""
+      },
+      {
+        "name": "Data (64, 8K, 1K)",
+        "detail": ""
+      },
+      {
+        "name": "Data (64, 8K, 1K, BF16)",
+        "detail": ""
+      },
+      {
+        "name": "Data (64, 8K, 1K, FP16)",
+        "detail": ""
+      },
+      {
+        "name": "Data (64, 8K, 512)",
+        "detail": ""
+      },
+      {
+        "name": "Hidden (32, 2K, 1K)",
+        "detail": ""
+      },
+      {
+        "name": "Hidden (32, 2K, 4K)",
+        "detail": ""
+      },
+      {
+        "name": "Hidden (32, 2K, 8K)",
+        "detail": ""
+      },
+      {
+        "name": "Seq (32, 4K, 4K)",
+        "detail": ""
+      },
+      {
+        "name": "Seq (32, 8K, 4K)",
+        "detail": ""
+      },
+      {
+        "name": "Seq (32, 16K, 4K)",
+        "detail": ""
+      }
+    ],
+    "metrics": {
+      "duration": {
+        "name": "Duration (ms)",
+        "lower_is_better": true,
+        "cuda": [
+          0.0494,
+          0.3755,
+          1.49,
+          5.93,
+          14.23,
+          0.3867,
+          7.12,
+          7.17,
+          7.16,
+          3.14,
+          0.892,
+          1.49,
+          2.29,
+          2.97,
+          5.93,
+          11.88
+        ],
+        "triton": [
+          0.0177,
+          0.1322,
+          0.5309,
+          2.15,
+          2.02,
+          0.1249,
+          1.02,
+          0.8622,
+          0.7968,
+          0.6206,
+          0.131,
+          0.5249,
+          1.07,
+          1.05,
+          2.14,
+          4.31
+        ],
+        "helion": [
+          0.0174,
+          0.1318,
+          0.5076,
+          1.98,
+          1.99,
+          0.0622,
+          0.9964,
+          0.5413,
+          0.5054,
+          0.4941,
+          0.1919,
+          0.5159,
+          0.9963,
+          0.9951,
+          1.98,
+          4.2
+        ]
+      },
+      "compute_throughput": {
+        "name": "Compute Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          57.7,
+          60.52,
+          60.94,
+          61.04,
+          69.95,
+          84.42,
+          69.94,
+          70.0,
+          69.83,
+          80.32,
+          69.75,
+          61.07,
+          56.19,
+          61.12,
+          61.05,
+          61.16
+        ],
+        "triton": [
+          54.46,
+          62.29,
+          58.58,
+          57.46,
+          76.41,
+          83.98,
+          76.34,
+          89.51,
+          93.53,
+          82.72,
+          77.92,
+          59.01,
+          58.83,
+          57.59,
+          57.35,
+          57.24
+        ],
+        "helion": [
+          34.42,
+          58.0,
+          63.68,
+          64.16,
+          78.01,
+          62.4,
+          78.23,
+          75.23,
+          78.99,
+          67.72,
+          59.61,
+          57.55,
+          60.78,
+          64.49,
+          64.7,
+          69.77
+        ]
+      },
+      "memory_throughput": {
+        "name": "Memory Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          19.67,
+          28.05,
+          29.77,
+          30.01,
+          22.38,
+          28.05,
+          22.38,
+          22.11,
+          22.21,
+          26.16,
+          22.32,
+          29.09,
+          38.76,
+          29.26,
+          30.01,
+          29.38
+        ],
+        "triton": [
+          52.09,
+          79.5,
+          83.71,
+          83.53,
+          86.1,
+          70.65,
+          85.11,
+          50.22,
+          55.5,
+          71.24,
+          80.07,
+          82.71,
+          83.03,
+          82.94,
+          83.7,
+          81.76
+        ],
+        "helion": [
+          47.6,
+          79.2,
+          87.76,
+          89.81,
+          87.45,
+          79.86,
+          87.18,
+          79.86,
+          87.31,
+          89.22,
+          77.96,
+          83.76,
+          89.15,
+          87.4,
+          89.78,
+          83.19
+        ]
+      },
+      "l1_tex_throughput": {
+        "name": "L1/TEX Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          21.2,
+          20.6,
+          20.69,
+          20.68,
+          22.39,
+          28.23,
+          22.39,
+          22.12,
+          22.22,
+          26.18,
+          22.39,
+          20.66,
+          19.77,
+          20.65,
+          20.68,
+          20.65
+        ],
+        "triton": [
+          46.5,
+          45.56,
+          42.64,
+          42.22,
+          58.08,
+          71.91,
+          57.84,
+          48.69,
+          53.0,
+          69.61,
+          60.57,
+          42.17,
+          41.81,
+          41.95,
+          41.8,
+          41.18
+        ],
+        "helion": [
+          40.8,
+          49.76,
+          41.25,
+          40.43,
+          57.55,
+          67.05,
+          57.72,
+          35.87,
+          38.78,
+          64.92,
+          47.98,
+          48.54,
+          39.26,
+          40.52,
+          40.45,
+          50.66
+        ]
+      },
+      "l2_throughput": {
+        "name": "L2 Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          33.35,
+          36.51,
+          37.21,
+          37.17,
+          15.9,
+          18.07,
+          15.95,
+          8.65,
+          8.66,
+          18.03,
+          15.73,
+          37.12,
+          47.62,
+          37.12,
+          37.17,
+          37.09
+        ],
+        "triton": [
+          80.32,
+          88.71,
+          90.25,
+          90.1,
+          92.91,
+          53.62,
+          92.15,
+          61.93,
+          66.63,
+          82.63,
+          90.88,
+          89.85,
+          88.17,
+          90.4,
+          90.39,
+          90.18
+        ],
+        "helion": [
+          81.11,
+          90.52,
+          93.69,
+          94.08,
+          93.69,
+          92.15,
+          93.7,
+          89.55,
+          92.94,
+          93.65,
+          89.42,
+          92.21,
+          93.63,
+          93.78,
+          94.07,
+          89.76
+        ]
+      },
+      "sm_busy": {
+        "name": "SM Busy (%)",
+        "lower_is_better": false,
+        "cuda": [
+          62.19,
+          61.08,
+          60.94,
+          61.04,
+          69.96,
+          84.96,
+          69.96,
+          70.03,
+          69.83,
+          80.32,
+          69.97,
+          61.21,
+          56.19,
+          61.19,
+          61.05,
+          61.18
+        ],
+        "triton": [
+          64.78,
+          62.93,
+          58.58,
+          57.46,
+          76.67,
+          85.49,
+          76.31,
+          89.72,
+          93.53,
+          82.72,
+          79.86,
+          58.49,
+          58.83,
+          58.3,
+          57.35,
+          57.36
+        ],
+        "helion": [
+          41.93,
+          59.0,
+          63.68,
+          64.16,
+          78.17,
+          67.19,
+          78.38,
+          75.73,
+          78.99,
+          67.72,
+          59.91,
+          57.66,
+          60.78,
+          64.7,
+          64.7,
+          69.84
+        ]
+      },
+      "no_eligible_warps": {
+        "name": "No Eligible Warps (%)",
+        "lower_is_better": false,
+        "cuda": [
+          37.84,
+          38.86,
+          38.92,
+          38.91,
+          30.04,
+          15.08,
+          30.04,
+          29.98,
+          30.15,
+          19.63,
+          30.04,
+          38.79,
+          43.73,
+          38.82,
+          38.92,
+          38.82
+        ],
+        "triton": [
+          35.38,
+          35.91,
+          41.01,
+          42.7,
+          23.35,
+          14.68,
+          23.77,
+          10.14,
+          6.18,
+          17.07,
+          20.65,
+          41.24,
+          40.97,
+          40.69,
+          42.91,
+          42.73
+        ],
+        "helion": [
+          57.97,
+          40.41,
+          35.66,
+          35.26,
+          21.84,
+          32.87,
+          21.77,
+          24.23,
+          20.86,
+          31.58,
+          39.9,
+          42.15,
+          39.53,
+          35.38,
+          35.5,
+          30.25
+        ]
+      },
+      "gflops": {
+        "name": "GFLOPs",
+        "lower_is_better": false,
+        "cuda": [
+          0.00472,
+          0.0378,
+          0.151,
+          0.605,
+          1.51,
+          0.0507,
+          0.757,
+          0.757,
+          0.757,
+          0.387,
+          0.0946,
+          0.151,
+          0.227,
+          0.302,
+          0.605,
+          1.21
+        ],
+        "triton": [
+          0.00339,
+          0.0271,
+          0.109,
+          0.434,
+          0.461,
+          0.0199,
+          0.231,
+          0.231,
+          0.231,
+          0.13,
+          0.0288,
+          0.109,
+          0.218,
+          0.217,
+          0.434,
+          0.868
+        ],
+        "helion": [
+          0.00463,
+          0.0417,
+          0.246,
+          0.985,
+          1.03,
+          0.0229,
+          0.516,
+          0.295,
+          0.295,
+          0.185,
+          0.0685,
+          0.167,
+          0.482,
+          0.492,
+          0.985,
+          2.06
+        ]
+      },
+      "registers_per_thread": {
+        "name": "Registers per Thread",
+        "lower_is_better": true,
+        "cuda": [
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18,
+          18
+        ],
+        "triton": [
+          80,
+          32,
+          64,
+          64,
+          32,
+          32,
+          32,
+          64,
+          64,
+          32,
+          47,
+          32,
+          57,
+          64,
+          64,
+          40
+        ],
+        "helion": [
+          36,
+          36,
+          34,
+          34,
+          25,
+          21,
+          25,
+          31,
+          31,
+          22,
+          25,
+          36,
+          39,
+          36,
+          34,
+          36
+        ]
+      },
+      "register_limited_blocks": {
+        "name": "Register-Limited Blocks",
+        "lower_is_better": true,
+        "cuda": [
+          2,
+          2,
+          2,
+          2,
+          2,
+          10,
+          2,
+          2,
+          2,
+          5,
+          2,
+          2,
+          2,
+          2,
+          2,
+          2
+        ],
+        "triton": [
+          6,
+          4,
+          4,
+          4,
+          16,
+          32,
+          16,
+          32,
+          32,
+          16,
+          40,
+          4,
+          4,
+          4,
+          4,
+          3
+        ],
+        "helion": [
+          6,
+          6,
+          6,
+          6,
+          16,
+          21,
+          16,
+          16,
+          16,
+          21,
+          16,
+          6,
+          3,
+          6,
+          6,
+          6
+        ]
+      },
+      "memory_bandwidth": {
+        "name": "Memory Bandwidth (GB/s)",
+        "lower_is_better": false,
+        "cuda": [
+          954.6,
+          1380,
+          1430,
+          1440,
+          602.39,
+          646.03,
+          600.86,
+          297.02,
+          297.64,
+          678.67,
+          580.95,
+          1430,
+          1870,
+          1440,
+          1440,
+          1440
+        ],
+        "triton": [
+          2560,
+          3910,
+          4030,
+          4010,
+          4230,
+          1990,
+          4190,
+          2470,
+          2670,
+          3430,
+          3930,
+          4060,
+          4000,
+          4080,
+          4030,
+          4020
+        ],
+        "helion": [
+          2330,
+          3890,
+          4220,
+          4320,
+          4300,
+          3920,
+          4280,
+          3920,
+          4200,
+          4290,
+          3830,
+          4120,
+          4290,
+          4300,
+          4320,
+          4090
+        ]
+      }
+    }
+  },
+  "softmax_backward": {
+    "configs": [
+      {
+        "name": "Batch (B=1)",
+        "detail": ""
+      },
+      {
+        "name": "Batch (B=8)",
+        "detail": ""
+      },
+      {
+        "name": "Batch (B=32)",
+        "detail": ""
+      },
+      {
+        "name": "Batch (B=128)",
+        "detail": ""
+      },
+      {
+        "name": "Data (32, 16K, 1024)",
+        "detail": ""
+      },
+      {
+        "name": "Data (32, 4K, 256)",
+        "detail": ""
+      },
+      {
+        "name": "Data (64, 8K, 1K)",
+        "detail": ""
+      },
+      {
+        "name": "Data (64, 8K, 1K, BF16)",
+        "detail": ""
+      },
+      {
+        "name": "Data (64, 8K, 1K, FP16)",
+        "detail": ""
+      },
+      {
+        "name": "Data (64, 8K, 512)",
+        "detail": ""
+      },
+      {
+        "name": "Hidden (32, 2K, 1K)",
+        "detail": ""
+      },
+      {
+        "name": "Hidden (32, 2K, 4K)",
+        "detail": ""
+      },
+      {
+        "name": "Hidden (32, 2K, 8K)",
+        "detail": ""
+      },
+      {
+        "name": "Seq (32, 4K, 4K)",
+        "detail": ""
+      },
+      {
+        "name": "Seq (32, 8K, 4K)",
+        "detail": ""
+      },
+      {
+        "name": "Seq (32, 16K, 4K)",
+        "detail": ""
+      }
+    ],
+    "metrics": {
+      "duration": {
+        "name": "Duration (ms)",
+        "lower_is_better": true,
+        "cuda": [
+          0.0442,
+          0.33,
+          1.31,
+          5.2,
+          9.22,
+          0.2556,
+          4.6,
+          4.51,
+          4.45,
+          2.12,
+          0.5773,
+          1.3,
+          2.3,
+          2.58,
+          5.21,
+          10.43
+        ],
+        "triton": [
+          0.0289,
+          0.2394,
+          0.9745,
+          4.22,
+          3.13,
+          0.1129,
+          1.55,
+          0.7338,
+          0.7295,
+          0.7284,
+          0.1865,
+          0.9757,
+          1.59,
+          1.97,
+          4.2,
+          7.96
+        ],
+        "helion": [
+          0.0233,
+          0.1876,
+          0.7271,
+          2.88,
+          2.87,
+          0.0881,
+          1.43,
+          0.7287,
+          0.7174,
+          0.7222,
+          0.3209,
+          0.7596,
+          1.45,
+          1.5,
+          2.88,
+          13.88
+        ]
+      },
+      "compute_throughput": {
+        "name": "Compute Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          45.74,
+          48.03,
+          47.91,
+          48.4,
+          62.96,
+          74.27,
+          63.07,
+          66.8,
+          67.01,
+          68.76,
+          62.84,
+          48.77,
+          43.21,
+          49.18,
+          48.4,
+          48.57
+        ],
+        "triton": [
+          17.54,
+          17.32,
+          15.63,
+          14.34,
+          27.78,
+          64.47,
+          28.02,
+          64.21,
+          57.92,
+          41.77,
+          30.52,
+          15.83,
+          19.38,
+          15.68,
+          14.4,
+          15.47
+        ],
+        "helion": [
+          7.45,
+          21.36,
+          15.12,
+          15.37,
+          30.08,
+          24.63,
+          28.24,
+          52.2,
+          41.83,
+          31.32,
+          54.2,
+          26.69,
+          18.47,
+          25.45,
+          15.2,
+          9.79
+        ]
+      },
+      "memory_throughput": {
+        "name": "Memory Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          40.78,
+          48.89,
+          50.99,
+          51.4,
+          28.41,
+          31.11,
+          28.42,
+          25.57,
+          25.87,
+          31.44,
+          27.98,
+          50.13,
+          58.23,
+          50.77,
+          51.37,
+          50.25
+        ],
+        "triton": [
+          66.8,
+          79.61,
+          81.79,
+          76.22,
+          84.55,
+          70.45,
+          84.87,
+          88.94,
+          91.36,
+          91.5,
+          86.56,
+          79.97,
+          88.19,
+          79.13,
+          76.42,
+          78.6
+        ],
+        "helion": [
+          72.11,
+          85.8,
+          91.54,
+          92.74,
+          91.35,
+          87.42,
+          91.18,
+          89.51,
+          92.67,
+          92.24,
+          84.31,
+          85.91,
+          92.33,
+          86.95,
+          92.75,
+          62.96
+        ]
+      },
+      "l1_tex_throughput": {
+        "name": "L1/TEX Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          25.9,
+          25.41,
+          25.12,
+          25.44,
+          25.57,
+          31.18,
+          25.61,
+          25.58,
+          25.88,
+          28.45,
+          25.68,
+          25.5,
+          26.01,
+          25.79,
+          25.48,
+          25.51
+        ],
+        "triton": [
+          36.24,
+          30.71,
+          28.54,
+          26.43,
+          42.05,
+          62.81,
+          42.3,
+          54.68,
+          54.85,
+          55.17,
+          46.75,
+          28.84,
+          35.73,
+          28.55,
+          28.1,
+          27.75
+        ],
+        "helion": [
+          31.56,
+          44.21,
+          32.99,
+          32.97,
+          46.3,
+          43.74,
+          46.48,
+          52.76,
+          38.99,
+          53.96,
+          45.64,
+          45.32,
+          42.17,
+          39.69,
+          33.02,
+          29.19
+        ]
+      },
+      "l2_throughput": {
+        "name": "L2 Throughput (%)",
+        "lower_is_better": false,
+        "cuda": [
+          56.3,
+          61.89,
+          62.47,
+          62.71,
+          36.78,
+          41.09,
+          36.93,
+          21.18,
+          21.4,
+          39.97,
+          36.77,
+          62.74,
+          70.49,
+          63.45,
+          62.67,
+          62.56
+        ],
+        "triton": [
+          82.27,
+          89.71,
+          92.07,
+          85.44,
+          92.43,
+          84.67,
+          92.97,
+          94.31,
+          94.55,
+          94.59,
+          93.27,
+          91.87,
+          94.35,
+          91.4,
+          88.1,
+          91.4
+        ],
+        "helion": [
+          88.48,
+          92.37,
+          94.09,
+          94.61,
+          95.07,
+          93.86,
+          95.04,
+          94.43,
+          94.94,
+          94.67,
+          89.94,
+          92.17,
+          94.25,
+          92.22,
+          94.62,
+          80.49
+        ]
+      },
+      "sm_busy": {
+        "name": "SM Busy (%)",
+        "lower_is_better": false,
+        "cuda": [
+          49.4,
+          48.56,
+          47.91,
+          48.4,
+          62.98,
+          74.97,
+          63.1,
+          66.83,
+          67.01,
+          68.76,
+          63.19,
+          48.87,
+          43.21,
+          49.21,
+          48.4,
+          48.6
+        ],
+        "triton": [
+          19.9,
+          16.99,
+          15.63,
+          14.34,
+          27.87,
+          65.96,
+          28.06,
+          64.19,
+          57.92,
+          41.77,
+          30.9,
+          15.99,
+          19.38,
+          15.85,
+          14.4,
+          15.43
+        ],
+        "helion": [
+          8.39,
+          21.41,
+          15.12,
+          15.37,
+          30.04,
+          25.89,
+          28.33,
+          52.44,
+          41.83,
+          31.32,
+          55.76,
+          26.49,
+          17.17,
+          24.45,
+          15.2,
+          5.8
+        ]
+      },
+      "no_eligible_warps": {
+        "name": "No Eligible Warps (%)",
+        "lower_is_better": false,
+        "cuda": [
+          50.33,
+          51.1,
+          51.87,
+          51.52,
+          37.01,
+          24.89,
+          36.9,
+          33.18,
+          32.97,
+          31.15,
+          36.88,
+          51.0,
+          56.69,
+          50.75,
+          51.53,
+          51.4
+        ],
+        "triton": [
+          79.93,
+          83.37,
+          84.3,
+          84.62,
+          71.93,
+          33.77,
+          71.85,
+          35.74,
+          41.96,
+          58.19,
+          69.12,
+          84.08,
+          80.66,
+          84.37,
+          85.49,
+          84.52
+        ],
+        "helion": [
+          91.55,
+          78.51,
+          84.81,
+          84.69,
+          70.02,
+          73.96,
+          71.62,
+          47.58,
+          57.88,
+          68.51,
+          43.86,
+          73.52,
+          82.77,
+          75.5,
+          84.62,
+          94.2
+        ]
+      },
+      "fused_operations": {
+        "name": "Fused Operations (instruction-mix)",
+        "lower_is_better": false,
+        "cuda": [
+          262000.0,
+          2100000.0,
+          8390000.0,
+          33600000.0,
+          33600000.0,
+          1050000.0,
+          16800000.0,
+          16800000.0,
+          16800000.0,
+          8390000.0,
+          2100000.0,
+          8390000.0,
+          16800000.0,
+          16800000.0,
+          33600000.0,
+          67100000.0
+        ],
+        "triton": [
+          246000.0,
+          1970000.0,
+          7860000.0,
+          31500000.0,
+          29400000.0,
+          524000.0,
+          14700000.0,
+          14700000.0,
+          14700000.0,
+          6290000.0,
+          1840000.0,
+          7860000.0,
+          15700000.0,
+          15700000.0,
+          31500000.0,
+          62900000.0
+        ],
+        "helion": [
+          254000.0,
+          1840000.0,
+          7340000.0,
+          29400000.0,
+          25200000.0,
+          786000.0,
+          12600000.0,
+          0,
+          0,
+          6290000.0,
+          1570000.0,
+          6290000.0,
+          6290000.0,
+          12600000.0,
+          29400000.0,
+          33600000.0
+        ]
+      },
+      "non_fused_operations": {
+        "name": "Non-Fused Operations (instruction-mix)",
+        "lower_is_better": false,
+        "cuda": [
+          862000.0,
+          6900000.0,
+          27600000.0,
+          110000000.0,
+          240000000.0,
+          8000000.0,
+          120000000.0,
+          120000000.0,
+          120000000.0,
+          61300000.0,
+          15000000.0,
+          27600000.0,
+          44400000.0,
+          55200000.0,
+          110000000.0,
+          221000000.0
+        ],
+        "triton": [
+          688000.0,
+          5510000.0,
+          22000000.0,
+          88100000.0,
+          105000000.0,
+          6820000.0,
+          52400000.0,
+          52400000.0,
+          52400000.0,
+          35700000.0,
+          6550000.0,
+          22000000.0,
+          45100000.0,
+          44000000.0,
+          88100000.0,
+          176000000.0
+        ],
+        "helion": [
+          598000.0,
+          7080000.0,
+          26200000.0,
+          105000000.0,
+          151000000.0,
+          4190000.0,
+          75500000.0,
+          35700000.0,
+          35700000.0,
+          35700000.0,
+          7860000.0,
+          41900000.0,
+          58700000.0,
+          83900000.0,
+          105000000.0,
+          302000000.0
+        ]
+      },
+      "gflops": {
+        "name": "GFLOPs",
+        "lower_is_better": false,
+        "cuda": [
+          0.00139,
+          0.0111,
+          0.0444,
+          0.177,
+          0.307,
+          0.0101,
+          0.154,
+          0.154,
+          0.154,
+          0.0781,
+          0.0192,
+          0.0444,
+          0.0779,
+          0.0887,
+          0.177,
+          0.355
+        ],
+        "triton": [
+          0.00118,
+          0.00944,
+          0.0377,
+          0.151,
+          0.164,
+          0.00786,
+          0.0818,
+          0.0818,
+          0.0818,
+          0.0482,
+          0.0102,
+          0.0371,
+          0.0765,
+          0.0755,
+          0.151,
+          0.302
+        ],
+        "helion": [
+          0.00111,
+          0.0107,
+          0.0409,
+          0.164,
+          0.201,
+          0.00577,
+          0.101,
+          0.0357,
+          0.0357,
+          0.0482,
+          0.011,
+          0.0545,
+          0.0881,
+          0.109,
+          0.164,
+          0.369
+        ]
+      },
+      "registers_per_thread": {
+        "name": "Registers per Thread",
+        "lower_is_better": true,
+        "cuda": [
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16
+        ],
+        "triton": [
+          32,
+          32,
+          32,
+          32,
+          32,
+          32,
+          32,
+          30,
+          32,
+          32,
+          32,
+          32,
+          40,
+          32,
+          32,
+          32
+        ],
+        "helion": [
+          78,
+          46,
+          54,
+          54,
+          22,
+          22,
+          22,
+          22,
+          30,
+          21,
+          22,
+          21,
+          32,
+          22,
+          54,
+          48
+        ]
+      },
+      "register_limited_blocks": {
+        "name": "Register-Limited Blocks",
+        "lower_is_better": true,
+        "cuda": [
+          4,
+          4,
+          4,
+          4,
+          4,
+          16,
+          4,
+          4,
+          4,
+          8,
+          4,
+          4,
+          4,
+          4,
+          4,
+          4
+        ],
+        "triton": [
+          8,
+          8,
+          8,
+          8,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          16,
+          8,
+          3,
+          8,
+          8,
+          8
+        ],
+        "helion": [
+          6,
+          2,
+          2,
+          2,
+          5,
+          42,
+          5,
+          5,
+          4,
+          10,
+          5,
+          2,
+          2,
+          2,
+          2,
+          10
+        ]
+      },
+      "memory_bandwidth": {
+        "name": "Memory Bandwidth (GB/s)",
+        "lower_is_better": false,
+        "cuda": [
+          2000,
+          2470,
+          2450,
+          2400,
+          1400,
+          1530,
+          1400,
+          712.17,
+          720.66,
+          1510,
+          1370,
+          1970,
+          4220,
+          4440,
+          3090,
+          4270
+        ],
+        "triton": [
+          3540,
+          4460,
+          4400,
+          4220,
+          4490,
+          4300,
+          4480,
+          4400,
+          4460,
+          4440,
+          4140,
+          4130,
+          3930,
+          4240,
+          3860,
+          3890
+        ],
+        "helion": [
+          3280,
+          3670,
+          3940,
+          3910,
+          4160,
+          3460,
+          4170,
+          4370,
+          4400,
+          4400,
+          4250,
+          2460,
+          2800,
+          2470,
+          2500,
+          2470
+        ]
+      }
+    }
+  }
+};
\ No newline at end of file
diff --git a/data_dashboard/index.html b/data_dashboard/index.html
new file mode 100644
index 0000000..9d947fb
--- /dev/null
+++ b/data_dashboard/index.html
@@ -0,0 +1,530 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>GPU Kernel Performance Dashboard</title>
+    <style>
+        * {
+            margin: 0;
+            padding: 0;
+            box-sizing: border-box;
+        }
+
+        :root {
+            --bg-primary: #f5f5f7;
+            --bg-secondary: white;
+            --bg-tertiary: #fafafa;
+            --text-primary: #1d1d1f;
+            --text-secondary: #86868b;
+            --border-color: #e5e5e5;
+        }
+
+        body {
+            font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, 'Helvetica Neue', Arial, sans-serif;
+            background: var(--bg-primary);
+            padding: 20px;
+            min-height: 100vh;
+            transition: all 0.3s;
+        }
+
+        body.dark-mode {
+            --bg-primary: #1d1d1f;
+            --bg-secondary: #2d2d2d;
+            --bg-tertiary: #3d3d3d;
+            --text-primary: #f5f5f7;
+            --text-secondary: #98989d;
+            --border-color: #48484a;
+        }
+
+        .container {
+            max-width: 1400px;
+            margin: 0 auto;
+            background: var(--bg-secondary);
+            border-radius: 12px;
+            box-shadow: 0 2px 10px rgba(0,0,0,0.08);
+            overflow: hidden;
+            border: 1px solid var(--border-color);
+        }
+
+        header {
+            background: var(--bg-secondary);
+            padding: 40px;
+            text-align: center;
+            border-bottom: 2px solid var(--border-color);
+            position: relative;
+        }
+
+        h1 {
+            font-size: 28px;
+            font-weight: 600;
+            color: var(--text-primary);
+            letter-spacing: -0.5px;
+        }
+
+        .subtitle {
+            font-size: 14px;
+            color: var(--text-secondary);
+            margin-top: 8px;
+        }
+
+        .header-actions {
+            position: absolute;
+            top: 40px;
+            right: 40px;
+            display: flex;
+            gap: 12px;
+            align-items: center;
+        }
+
+        .dark-mode-toggle, .download-btn {
+            background: var(--bg-tertiary);
+            border: 1px solid var(--border-color);
+            border-radius: 20px;
+            padding: 8px 16px;
+            cursor: pointer;
+            font-size: 13px;
+            font-weight: 500;
+            color: var(--text-primary);
+            transition: all 0.2s;
+            display: flex;
+            align-items: center;
+            gap: 8px;
+            text-decoration: none;
+        }
+
+        .dark-mode-toggle:hover, .download-btn:hover {
+            background: var(--border-color);
+        }
+
+        .controls {
+            background: var(--bg-tertiary);
+            padding: 24px 40px;
+            border-bottom: 2px solid var(--border-color);
+            display: flex;
+            gap: 20px;
+            align-items: flex-end;
+            flex-wrap: wrap;
+            justify-content: center;
+        }
+
+        .control-group {
+            display: flex;
+            flex-direction: column;
+            gap: 8px;
+        }
+
+        .control-group label {
+            font-size: 12px;
+            font-weight: 500;
+            color: var(--text-secondary);
+            text-transform: uppercase;
+            letter-spacing: 0.5px;
+        }
+
+        select {
+            padding: 10px 14px;
+            border: 1px solid var(--border-color);
+            border-radius: 8px;
+            font-size: 14px;
+            font-weight: 400;
+            background: var(--bg-secondary);
+            color: var(--text-primary);
+            cursor: pointer;
+            transition: all 0.2s;
+            min-width: 250px;
+        }
+
+        select:hover {
+            border-color: var(--text-secondary);
+        }
+
+        select:focus {
+            outline: none;
+            border-color: #0071e3;
+            box-shadow: 0 0 0 3px rgba(0, 113, 227, 0.1);
+        }
+
+        .stats-bar {
+            background: var(--bg-secondary);
+            padding: 20px 40px;
+            display: flex;
+            justify-content: center;
+            gap: 60px;
+            border-bottom: 2px solid var(--border-color);
+        }
+
+        .stat {
+            text-align: center;
+            padding: 12px 24px;
+            border: 1px solid var(--border-color);
+            border-radius: 8px;
+            background: var(--bg-tertiary);
+        }
+
+        .stat-value {
+            font-size: 32px;
+            font-weight: 600;
+            color: var(--text-primary);
+        }
+
+        .stat-label {
+            font-size: 12px;
+            color: var(--text-secondary);
+            margin-top: 4px;
+            text-transform: uppercase;
+            letter-spacing: 0.5px;
+        }
+
+        .legend {
+            background: var(--bg-secondary);
+            padding: 20px 40px;
+            border-bottom: 2px solid var(--border-color);
+            display: flex;
+            justify-content: center;
+            gap: 32px;
+            flex-wrap: wrap;
+        }
+
+        .legend-item {
+            display: flex;
+            align-items: center;
+            gap: 10px;
+            font-size: 13px;
+            font-weight: 500;
+            color: var(--text-primary);
+            padding: 8px 16px;
+            background: var(--bg-tertiary);
+            border-radius: 10px;
+            border: 1px solid var(--border-color);
+        }
+
+        .legend-box {
+            width: 20px;
+            height: 20px;
+            border-radius: 6px;
+            border: 1px solid var(--border-color);
+        }
+
+        .table-wrapper {
+            overflow-x: auto;
+            padding: 32px 40px 48px;
+            max-height: 70vh;
+            overflow-y: auto;
+            background: var(--bg-tertiary);
+        }
+
+        .perf-table {
+            width: 100%;
+            border-collapse: separate;
+            border-spacing: 0;
+            font-size: 13px;
+            background: var(--bg-secondary);
+            border-radius: 12px;
+            overflow: hidden;
+            border: 2px solid var(--border-color);
+        }
+
+        .perf-table th {
+            background: linear-gradient(180deg,
+                rgba(29, 29, 31, 0.98) 0%,
+                rgba(29, 29, 31, 0.95) 100%);
+            backdrop-filter: blur(40px) saturate(200%) brightness(1.1);
+            -webkit-backdrop-filter: blur(40px) saturate(200%) brightness(1.1);
+            color: white;
+            padding: 18px 16px;
+            text-align: center;
+            font-weight: 600;
+            font-size: 12px;
+            letter-spacing: 0.8px;
+            border-bottom: 1px solid rgba(255, 255, 255, 0.15);
+            position: sticky;
+            top: -1px;
+            z-index: 1000;
+            box-shadow:
+                0 8px 32px rgba(0, 0, 0, 0.4),
+                0 2px 8px rgba(0, 0, 0, 0.2),
+                inset 0 1px 0 rgba(255, 255, 255, 0.1),
+                inset 0 -1px 0 rgba(0, 0, 0, 0.2);
+            transition: all 0.3s cubic-bezier(0.4, 0, 0.2, 1);
+        }
+
+        body.dark-mode .perf-table th {
+            background: linear-gradient(180deg,
+                rgba(10, 10, 10, 0.98) 0%,
+                rgba(15, 15, 15, 0.95) 100%);
+            backdrop-filter: blur(40px) saturate(200%) brightness(0.9);
+            -webkit-backdrop-filter: blur(40px) saturate(200%) brightness(0.9);
+            color: #f5f5f7;
+            border-bottom: 1px solid rgba(255, 255, 255, 0.08);
+            box-shadow:
+                0 8px 32px rgba(0, 0, 0, 0.6),
+                0 2px 8px rgba(0, 0, 0, 0.4),
+                inset 0 1px 0 rgba(255, 255, 255, 0.05),
+                inset 0 -1px 0 rgba(0, 0, 0, 0.3);
+        }
+
+        .perf-table thead {
+            position: sticky;
+            top: 0;
+            z-index: 1000;
+        }
+
+        .perf-table thead::after {
+            content: '';
+            position: absolute;
+            bottom: -8px;
+            left: 0;
+            right: 0;
+            height: 8px;
+            background: linear-gradient(180deg,
+                rgba(0, 0, 0, 0.15) 0%,
+                transparent 100%);
+            pointer-events: none;
+        }
+
+        .perf-table th::before {
+            content: '';
+            position: absolute;
+            top: 0;
+            left: 0;
+            right: 0;
+            height: 2px;
+            background: linear-gradient(90deg,
+                transparent 0%,
+                rgba(255, 255, 255, 0.15) 20%,
+                rgba(255, 255, 255, 0.15) 80%,
+                transparent 100%);
+            opacity: 0.6;
+        }
+
+        .perf-table th:first-child {
+            text-align: left;
+        }
+
+        .header-impl {
+            display: block;
+            font-size: 13px;
+            font-weight: 600;
+            text-transform: uppercase;
+            letter-spacing: 0.5px;
+        }
+
+        .header-unit {
+            display: block;
+            font-size: 11px;
+            font-weight: 400;
+            color: var(--text-secondary);
+            margin-top: 4px;
+        }
+
+        .perf-table td {
+            padding: 16px;
+            border-bottom: 1px solid var(--border-color);
+            font-weight: 500;
+            color: var(--text-primary);
+            text-align: center;
+        }
+
+        .perf-table td:first-child {
+            text-align: left;
+        }
+
+        .perf-table tr:last-child td {
+            border-bottom: none;
+        }
+
+        .config-cell {
+            font-weight: 600;
+            font-size: 13px;
+            text-align: left;
+            background: var(--bg-tertiary);
+            color: var(--text-primary);
+        }
+
+        .config-detail {
+            display: block;
+            font-size: 11px;
+            color: var(--text-secondary);
+            font-weight: 400;
+            margin-top: 4px;
+            font-family: 'Monaco', 'Courier New', monospace;
+        }
+
+        .best {
+            background: #52c896;
+            color: white;
+            font-weight: 600;
+            text-align: center;
+        }
+
+        body.dark-mode .best {
+            background: #3d9970;
+            color: white;
+        }
+
+        .second {
+            background: #ffa726;
+            color: white;
+            font-weight: 600;
+            text-align: center;
+        }
+
+        body.dark-mode .second {
+            background: #ff8f00;
+            color: white;
+        }
+
+        .worst {
+            background: #ef5350;
+            color: white;
+            text-align: center;
+            font-weight: 600;
+        }
+
+        body.dark-mode .worst {
+            background: #d32f2f;
+            color: white;
+        }
+
+        .speedup {
+            display: inline-block;
+            padding: 6px 12px;
+            border-radius: 6px;
+            font-size: 12px;
+            font-weight: 600;
+        }
+
+        .speedup-high {
+            background: #52c896;
+            color: white;
+        }
+
+        body.dark-mode .speedup-high {
+            background: #3d9970;
+            color: white;
+        }
+
+        .speedup-medium {
+            background: #ffa726;
+            color: white;
+        }
+
+        body.dark-mode .speedup-medium {
+            background: #ff8f00;
+            color: white;
+        }
+
+        .speedup-extreme {
+            background: #ab47bc;
+            color: white;
+        }
+
+        body.dark-mode .speedup-extreme {
+            background: #8e24aa;
+            color: white;
+        }
+
+        .speedup-regression {
+            background: #ef5350;
+            color: white;
+        }
+
+        body.dark-mode .speedup-regression {
+            background: #d32f2f;
+            color: white;
+        }
+
+        footer {
+            background: var(--bg-secondary);
+            color: var(--text-secondary);
+            padding: 32px;
+            text-align: center;
+            font-size: 14px;
+            border-top: 2px solid var(--border-color);
+            font-weight: 400;
+        }
+
+        tr:hover {
+            background: var(--bg-tertiary);
+        }
+    </style>
+</head>
+<body>
+    <div class="container">
+        <header>
+            <h1>GPU Kernel Profiling Data</h1>
+            <p class="subtitle">NVIDIA H200 • CUDA vs Triton vs Helion</p>
+            <div class="header-actions">
+                <button class="download-btn" id="downloadBtn">
+                    <span>📥</span>
+                    <span>Download CSV</span>
+                </button>
+                <button class="dark-mode-toggle" id="darkModeToggle">
+                    <span id="modeIcon">🌙</span>
+                    <span id="modeText">Dark</span>
+                </button>
+            </div>
+        </header>
+
+        <div class="controls">
+            <div class="control-group">
+                <label>Kernel</label>
+                <select id="kernelSelect">
+                    <option value="vector_addition">Vector Addition (4 metrics)</option>
+                    <option value="batched_matmul">Batched Matrix Multiplication (7 metrics)</option>
+                    <option value="softmax_forward">Softmax Forward (11 metrics)</option>
+                    <option value="softmax_backward">Softmax Backward (13 metrics)</option>
+                </select>
+            </div>
+
+            <div class="control-group">
+                <label>Metric</label>
+                <select id="metricSelect">
+                    <!-- Populated dynamically -->
+                </select>
+            </div>
+        </div>
+
+        <div class="stats-bar" id="statsBar">
+            <div class="stat">
+                <div class="stat-value" id="avgSpeedup">-</div>
+                <div class="stat-label">Avg Speedup</div>
+            </div>
+            <div class="stat">
+                <div class="stat-value" id="maxSpeedup">-</div>
+                <div class="stat-label">Max Speedup</div>
+            </div>
+        </div>
+
+        <div class="legend">
+            <div class="legend-item">
+                <div class="legend-box best"></div>
+                <span>Best</span>
+            </div>
+            <div class="legend-item">
+                <div class="legend-box second"></div>
+                <span>Second</span>
+            </div>
+            <div class="legend-item">
+                <div class="legend-box worst"></div>
+                <span>Worst</span>
+            </div>
+        </div>
+
+        <div class="table-wrapper">
+            <table class="perf-table" id="dataTable">
+                <thead id="tableHead"></thead>
+                <tbody id="tableBody"></tbody>
+            </table>
+        </div>
+
+        <footer>
+            <p>Made with ❤️ by Red Hat PyTorch Engineering Interns</p>
+        </footer>
+    </div>
+
+    <script src="data.js"></script>
+    <script src="app.js"></script>
+</body>
+</html>