Update README.md

README.md

@@ -154,25 +154,28 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <th></th>
       <th></th>
       <th></th>
+      <th></th>
       <th style="text-align: center;" colspan="2" >Document Visual Question Answering<br>1680W x 2240H<br>64/128</th>
       <th style="text-align: center;" colspan="2" >Visual Reasoning <br>640W x 480H<br>128/128</th>
       <th style="text-align: center;" colspan="2" >Image Captioning<br>480W x 360H<br>0/128</th>
     </tr>
     <tr>
       <th>Hardware</th>
+      <th>Number of GPUs</th>
       <th>Model</th>
       <th>Average Cost Reduction</th>
       <th>Latency (s)</th>
-      <th>QPD</th>
+      <th>Queries Per Dollar</th>
       <th>Latency (s)th>
-      <th>QPD</th>
+      <th>Queries Per Dollar</th>
       <th>Latency (s)</th>
-      <th>QPD</th>
+      <th>Queries Per Dollar</th>
     </tr>
   </thead>
   <tbody>
     <tr>
-      <td>A100x4</td>
+      <th rowspan="3" valign="top">A100</td>
+      <td>4</td>
       <td>Qwen/Qwen2.5-VL-72B-Instruct</td>
       <td></td>
       <td>6.4</td>
@@ -183,7 +186,7 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <td>113</td>
     </tr>
     <tr>
-      <td>A100x2</td>
+      <td>2</td>
       <td>neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w8a8</td>
       <td>1.85</td>
       <td>7.0</td>
@@ -194,7 +197,7 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <td>211</td>
     </tr>
     <tr>
-      <td>A100x1</td>
+      <td>1</td>
       <td>neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16</td>
       <td>3.33</td>
       <td>9.4</td>
@@ -205,7 +208,8 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <td>420</td>
     </tr>
     <tr>
-      <td>H100x4</td>
+      <th rowspan="3" valign="top">H100</td>
+      <td>4</td>
       <td>Qwen/Qwen2.5-VL-72B-Instruct</td>
       <td></td>
       <td>4.3</td>
@@ -216,7 +220,7 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <td>100</td>
     </tr>
     <tr>
-      <td>H100x2</td>
+      <td>2</td>
       <td>neuralmagic/Qwen2.5-VL-72B-Instruct-FP8-Dynamic</td>
       <td>1.79</td>
       <td>4.6</td>
@@ -227,7 +231,7 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <td>177</td>
     </tr>
     <tr>
-      <td>H100x1</td>
+      <td>1</td>
       <td>neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16</td>
       <td>5.66</td>
       <td>4.3</td>
@@ -240,6 +244,9 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
   </tbody>
 </table>
 
+**Use case profiles: Image Size (WxH) / prompt tokens / generation tokens
+
+**QPD: Queries per dollar, based on on-demand cost at [Lambda Labs](https://lambdalabs.com/service/gpu-cloud) (observed on 2/18/2025).
 
 ### Multi-stream asynchronous performance (measured with vLLM version 0.7.2)
 
@@ -258,16 +265,16 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <th>Model</th>
       <th>Average Cost Reduction</th>
       <th>Maximum throughput (QPS)</th>
-      <th>QPD</th>
+      <th>Queries Per Dollar</th>
       <th>Maximum throughput (QPS)</th>
-      <th>QPD</th>
+      <th>Queries Per Dollar</th>
       <th>Maximum throughput (QPS)</th>
-      <th>QPD</th>
+      <th>Queries Per Dollar</th>
     </tr>
   </thead>
   <tbody style="text-align: center">
     <tr>
-      <td>A100x4</td>
+      <th rowspan="3" valign="top">A100x4</th>
       <td>Qwen/Qwen2.5-VL-72B-Instruct</td>
       <td></td>
       <td>0.4</td>
@@ -278,29 +285,27 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <td>595</td>
     </tr>
     <tr>
-      <td>A100x2</td>
       <td>neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w8a8</td>
       <td>1.80</td>
-      <td>0.6</td>
+      <td>1.2</td>
       <td>289</td>
-      <td>2.0</td>
+      <td>4.0</td>
       <td>1020</td>
-      <td>2.3</td>
+      <td>4.6</td>
       <td>1133</td>
     </tr>
     <tr>
-      <td>A100x1</td>
       <td>neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16</td>
       <td>2.75</td>
-      <td>0.7</td>
+      <td>2.8</td>
       <td>341</td>
-      <td>3.2</td>
+      <td>12.8</td>
       <td>1588</td>
-      <td>4.1</td>
+      <td>16.4</td>
       <td>2037</td>
     </tr>
     <tr>
-      <td>H100x4</td>
+      <th rowspan="3" valign="top">H100x4</th>
       <td>Qwen/Qwen2.5-VL-72B-Instruct</td>
       <td></td>
       <td>0.5</td>
@@ -311,26 +316,30 @@ The following performance benchmarks were conducted with [vLLM](https://docs.vll
       <td>379</td>
     </tr>
     <tr>
-      <td>H100x2</td>
       <td>neuralmagic/Qwen2.5-VL-72B-Instruct-FP8-Dynamic</td>
       <td>1.73</td>
-      <td>0.9</td>
+      <td>1.8</td>
       <td>247</td>
-      <td>2.2</td>
+      <td>4.4</td>
       <td>621</td>
-      <td>2.4</td>
+      <td>4.8</td>
       <td>669</td>
     </tr>
     <tr>
-      <td>H100x1</td>
       <td>neuralmagic/Qwen2.5-VL-72B-Instruct-quantized.w4a16</td>
       <td>8.27</td>
-      <td>3.3</td>
+      <td>13.2</td>
       <td>913</td>
-      <td>3.3</td>
+      <td>13.2</td>
       <td>913</td>
-      <td>24.8</td>
+      <td>99.2</td>
       <td>6777</td>
     </tr>
   </tbody>
-</table>
+</table>
+
+**Use case profiles: Image Size (WxH) / prompt tokens / generation tokens
+
+**QPS: Queries per second.
+
+**QPD: Queries per dollar, based on on-demand cost at [Lambda Labs](https://lambdalabs.com/service/gpu-cloud) (observed on 2/18/2025).