[자율주행 AI 챌린지] 3D 후보모델 선정 FLOPs + 추론+ Score 추가

<p><span style="color: #ee2323;" data-ke-size="size26">추가! 작년 대회 자료를 확보함 FLOPs(G)와 추론시간을 확보</span></p><div class="figure-file" data-ke-type="file" data-file-src="https://t1.daumcdn.net/cafeattach/1RgNt/c7aa69878b4a25ee4ebe4694d140636e4b18e817?download" data-file-name="3D_객체_검출(차량)_실험결과.xlsx" data-file-size="26138" data-mimetype="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet" data-ke-align="alignCenter"><a href="javascript:checkVirus('grpid%3D1RgNt%26fldid%3DRoVa%26dataid%3D2335%26fileid%3D8%26regdt%3D20250826005946&url=https%3A%2F%2Ft1.daumcdn.net%2Fcafeattach%2F1RgNt%2Fc7aa69878b4a25ee4ebe4694d140636e4b18e817')"><div class="image"></div><div class="desc"><div class="filename"><span class="name">3D_객체_검출(차량)_실험결과</span><span class="type">.xlsx</span></div><div class="size">25.53KB</div></div></a></div><div class="figure-file" data-ke-type="file" data-file-src="https://t1.daumcdn.net/cafeattach/1RgNt/60b78e7f372ed8e8ab80e12a623dcf38745c46ea?download" data-file-name="3D_객체_검출(인프라)_실험결과.xlsx" data-file-size="15331" data-mimetype="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet" data-ke-align="alignCenter"><a href="javascript:checkVirus('grpid%3D1RgNt%26fldid%3DRoVa%26dataid%3D2335%26fileid%3D9%26regdt%3D20250826005946&url=https%3A%2F%2Ft1.daumcdn.net%2Fcafeattach%2F1RgNt%2F60b78e7f372ed8e8ab80e12a623dcf38745c46ea')"><div class="image"></div><div class="desc"><div class="filename"><span class="name">3D_객체_검출(인프라)_실험결과</span><span class="type">.xlsx</span></div><div class="size">14.97KB</div></div></a></div><p> </p><p>원 논문의 GPU---> 4090 정확한 환산 방법을 찾는중 ----> 밴치마크로 근사화 하는방법이 TFLOPS를 활용하는 방법보다 </p><p><br><b><span data-ke-size="size23">추론관련 근거자료</span></b><br><br><a href="https://bizon-tech.com/gpu-benchmarks/NVIDIA-A100-80-GB-(SXM4)-vs-NVIDIA-RTX-4090/625vs637#benchmarks" target="_blank" class="ke-link">https://bizon-tech.com/gpu-benchmarks/NVIDIA-A100-80-GB-(SXM4)-vs-NVIDIA-RTX-4090/625vs637#benchmarks</a><br><br></p><div class="figure-open" contenteditable="false" data-ke-type="opengraph" data-ke-align="alignCenter" data-og-type="website" data-og-title="NVIDIA A100 80 GB (SXM4) vs. NVIDIA RTX 4090| GPU Benchmarks for AI/ML, LLM, deep learning 2025 | BIZON" data-og-description="In this article, we are comparing the best graphics cards for deep learning in 2025: NVIDIA RTX 5090 vs 4090 vs RTX 6000, A100, H100 vs RTX 4090" data-og-host="bizon-tech.com" data-og-source-url="https://bizon-tech.com/gpu-benchmarks/NVIDIA-A100-80-GB-(SXM4)-vs-NVIDIA-RTX-4090/625vs637#benchmarks" data-og-url="https://bizon-tech.com/gpu-benchmarks/NVIDIA-A100-80-GB-(SXM4)-vs-NVIDIA-RTX-4090/625vs637" data-og-image="https://scrap.kakaocdn.net/dn/biQ1v2/hyZCZN785D/nfz81RJI56LA011Rmnmq6K/img.jpg?width=1920&height=341&face=0_0_1920_341,https://scrap.kakaocdn.net/dn/jfI6n/hyZDaa26jT/Mq97NCNviyqKfXa0GB0rPk/img.jpg?width=1920&height=341&face=0_0_1920_341"><a href="https://bizon-tech.com/gpu-benchmarks/NVIDIA-A100-80-GB-(SXM4)-vs-NVIDIA-RTX-4090/625vs637#benchmarks" target="_blank" data-source-url="https://bizon-tech.com/gpu-benchmarks/NVIDIA-A100-80-GB-(SXM4)-vs-NVIDIA-RTX-4090/625vs637#benchmarks"><div class="og-image"><img class="thumb_img" src="https://scrap.kakaocdn.net/dn/biQ1v2/hyZCZN785D/nfz81RJI56LA011Rmnmq6K/img.jpg?width=1920&height=341&face=0_0_1920_341,https://scrap.kakaocdn.net/dn/jfI6n/hyZDaa26jT/Mq97NCNviyqKfXa0GB0rPk/img.jpg?width=1920&height=341&face=0_0_1920_341" alt="" xxxxonerror="this.src="//img1.kakaocdn.net/thumb/C200x200/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fcafe_image%2Fcafe_meta_image_190529.png""></div><div class="og-text"><p class="og-title">NVIDIA A100 80 GB (SXM4) vs. NVIDIA RTX 4090| GPU Benchmarks for AI/ML, LLM, deep learning 2025 | BIZON</p><p class="og-desc">In this article, we are comparing the best graphics cards for deep learning in 2025: NVIDIA RTX 5090 vs 4090 vs RTX 6000, A100, H100 vs RTX 4090</p><p class="og-host">bizon-tech.com</p></div></a></div><p> </p><p><a href="https://compute.hivenet.com/post/why-more-developers-are-choosing-rtx-4090-over-a100" target="_blank" class="ke-link">https://compute.hivenet.com/post/why-more-developers-are-choosing-rtx-4090-over-a100</a></p><div class="figure-open" contenteditable="false" data-ke-type="opengraph" data-ke-align="alignCenter" data-og-type="website" data-og-title="Why more developers are choosing RTX 4090 over A100 for AI workloads | Hivenet" data-og-description="GPU supply is tight—developers need smarter options. Here's how the RTX 4090 stacks up to the A100 for AI, RAG, and LLMs, especially when cost matters." data-og-host="compute.hivenet.com" data-og-source-url="https://compute.hivenet.com/post/why-more-developers-are-choosing-rtx-4090-over-a100" data-og-url="https://compute.hivenet.com/post/why-more-developers-are-choosing-rtx-4090-over-a100" data-og-image="//img1.kakaocdn.net/thumb/C200x200/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fcafe_image%2Fcafe_meta_image_190529.png"><a href="https://compute.hivenet.com/post/why-more-developers-are-choosing-rtx-4090-over-a100" target="_blank" data-source-url="https://compute.hivenet.com/post/why-more-developers-are-choosing-rtx-4090-over-a100"><div class="og-image"><img class="thumb_img" src="//img1.kakaocdn.net/thumb/C200x200/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fcafe_image%2Fcafe_meta_image_190529.png" alt="" xxxxonerror="this.src="//img1.kakaocdn.net/thumb/C200x200/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fcafe_image%2Fcafe_meta_image_190529.png""></div><div class="og-text"><p class="og-title">Why more developers are choosing RTX 4090 over A100 for AI workloads | Hivenet</p><p class="og-desc">GPU supply is tight—developers need smarter options. Here's how the RTX 4090 stacks up to the A100 for AI, RAG, and LLMs, especially when cost matters.</p><p class="og-host">compute.hivenet.com</p></div></a></div><p> </p><p> </p><p>아래의 표는 Waymo L2 mAP와 논문에서 사용한 GPU +추론속도를 표로 정리함 4090 정확한 환산수식을 확보하면 계산 예정</p><p>1차 평가 기준: 베이스라인 모델(CenterPoint-Pillar) FLOPs(G)의 2배를 넘기면 탈락 </p><p>정확한 FLOPs는 베이스라인 코드로 실시 해야함 현재는 논문에 기재된 다른 모델들의 FLOPs 값을 보고 유추한 결과임</p><p> </p><p>추론 지표는 논문에 기재된 FPS or ms 값을 벤치마크 기준으로 환산함 </p><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/a822416dc238eb2f1ab83e8c75915690bed17021" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/a822416dc238eb2f1ab83e8c75915690bed17021" data-origin-width="587" data-origin-height="110"><div class="figcaption">2차 평가 기준</div></div><p> </p><div class="table-wrap"><table data-ke-type="table" data-ke-align="alignLeft" style="width: 101.13%; height: 288px;" border="1"><tbody><tr style="height: 54px;"><td style="width: 17.2762%; height: 54px;">모델</td><td style="width: 24.7115%; height: 54px;"><b> Waymo L2 mAP (Veh+Ped+Cyc 평균) </b></td><td style="width: 25.4675%; height: 54px;">원 논문 GPU 및 추론 속도</td><td style="width: 12.9198%; height: 54px;"><b> RTX 4090 추론 시간 (추정) </b></td><td style="width: 12.7941%; height: 54px;"><b>FLOPs(G)</b></td><td style="width: 7.83367%; height: 54px;"><b> 점수 (Score) </b></td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;">CenterPoint-Pillar</td><td style="width: 24.7115%; height: 18px;"><b>예외 작년 기준 약 42%</b></td><td style="width: 25.4675%; height: 18px;">NVIDIA A100 @ ~28 FPS <br>35.7 ms</td><td style="width: 12.9198%; height: 18px;">작년 기준 32</td><td style="width: 12.7941%; height: 18px;">작년기준 493</td><td style="width: 7.83367%; height: 18px;">39</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;">PointPillars</td><td style="width: 24.7115%; height: 18px;"><b>62.8%</b></td><td style="width: 25.4675%; height: 18px;">GTX 1080 Ti @ ~42 FPS<br>23.8 ms</td><td style="width: 12.9198%; height: 18px;">4.8</td><td style="width: 12.7941%; height: 18px;"></td><td style="width: 7.83367%; height: 18px;">63</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;">SECOND</td><td style="width: 24.7115%; height: 18px;"><b>61.0%</b></td><td style="width: 25.4675%; height: 18px;">GTX 1080 Ti @ ~20 FPS<br>50 ms </td><td style="width: 12.9198%; height: 18px;">10.08</td><td style="width: 12.7941%; height: 18px;">91.2</td><td style="width: 7.83367%; height: 18px;">61</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;">Part-A2-anchor</td><td style="width: 24.7115%; height: 18px;"><b>66.9%</b></td><td style="width: 25.4675%; height: 18px;">Tesla V100 @ ~15 FPS<br>66.6 ms</td><td style="width: 12.9198%; height: 18px;">23.1</td><td style="width: 12.7941%; height: 18px;"></td><td style="width: 7.83367%; height: 18px;">65</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;">PV-RCNN</td><td style="width: 24.7115%; height: 18px;"><b>66.8%</b></td><td style="width: 25.4675%; height: 18px;">TITAN RTX @ ~3.3 FPS<br>303ms</td><td style="width: 12.9198%; height: 18px;">108</td><td style="width: 12.7941%; height: 18px;"></td><td style="width: 7.83367%; height: 18px;">34</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;"><b>PV-RCNN++</b></td><td style="width: 24.7115%; height: 18px;"><b>68.4%</b></td><td style="width: 25.4675%; height: 18px;">TITAN RTX @ <b>~10 FPS<br>100 ms</b></td><td style="width: 12.9198%; height: 18px;">35.7</td><td style="width: 12.7941%; height: 18px;">166.6</td><td style="width: 7.83367%; height: 18px;">62</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;"><b>CenterPoint</b></td><td style="width: 24.7115%; height: 18px;"><b>69.8%</b></td><td style="width: 25.4675%; height: 18px;">TITAN RTX @ <b>~11 FPS<br>90.9 ms</b></td><td style="width: 12.9198%; height: 18px;">32.4</td><td style="width: 12.7941%; height: 18px;">141.2</td><td style="width: 7.83367%; height: 18px;">65</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;">Voxel R-CNN</td><td style="width: 24.7115%; height: 18px;">~65% (추정)</td><td style="width: 25.4675%; height: 18px;">RTX 2080 Ti @ <b>25 FPS<br>40 ms</b></td><td style="width: 12.9198%; height: 18px;">13.0</td><td style="width: 12.7941%; height: 18px;"></td><td style="width: 7.83367%; height: 18px;">65</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;"><b>DSVT (Voxel)</b></td><td style="width: 24.7115%; height: 18px;"><b>73.0%</b></td><td style="width: 25.4675%; height: 18px;">NVIDIA A100 @ <b>~10 FPS<br>100 ms</b></td><td style="width: 12.9198%; height: 18px;">123.4</td><td style="width: 12.7941%; height: 18px;"></td><td style="width: 7.83367%; height: 18px;">31</td></tr><tr style="height: 18px;"><td style="width: 17.2762%; height: 18px;">DSVT (Pillar)</td><td style="width: 24.7115%; height: 18px;">71.1% </td><td style="width: 25.4675%; height: 18px;">NVIDIA A100 @ <b>15 FPS<br>66.6 ms</b></td><td style="width: 12.9198%; height: 18px;">82.19</td><td style="width: 12.7941%; height: 18px;"></td><td style="width: 7.83367%; height: 18px;">46</td></tr><tr style="height: 54px;"><td style="width: 17.2762%; height: 54px;"><b>MPPNet (16frm)</b></td><td style="width: 24.7115%; height: 54px;"><b>~76.4%</b></td><td style="width: 25.4675%; height: 54px;">NVIDIA RTX 4090 @ <b>5FPS<br>200 ms</b></td><td style="width: 12.9198%; height: 54px;"><b>≫200 ms</b> (실시간 불가)</td><td style="width: 12.7941%; height: 54px;"><b> 220</b></td><td style="width: 7.83367%; height: 54px;"><b>0.0</b></td></tr></tbody></table></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/e87bf8b493b8361467177ce4d02ccabccce330d4_re_1756137565783" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/e87bf8b493b8361467177ce4d02ccabccce330d4_re_1756137565783" data-origin-width="926" data-origin-height="551"><div class="figcaption">MPPNet</div></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/52407e21624ec464484e53d87a96cc0a9f7d0fe9_re_1756137565783" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/52407e21624ec464484e53d87a96cc0a9f7d0fe9_re_1756137565783" data-origin-width="623" data-origin-height="308"><div class="figcaption">MPPNet(4/16 frame)</div></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/880735e22a18f12e2cf6f949d38fb09c3246af59_re_1756137565782" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/880735e22a18f12e2cf6f949d38fb09c3246af59_re_1756137565782" data-origin-width="1107" data-origin-height="368"><div class="figcaption">pv-rcnn++</div></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/31071fad6241431dbd92d0149eed0d0f110accc2_re_1756137565783" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/31071fad6241431dbd92d0149eed0d0f110accc2_re_1756137565783" data-origin-width="539" data-origin-height="234"><div class="figcaption">DSVT single A100</div></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/3602b5a6cbe02c7e7959d0c2997b1846dc101028" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/3602b5a6cbe02c7e7959d0c2997b1846dc101028" data-origin-width="771" data-origin-height="297"><div class="figcaption">SEED</div></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/98468fa7b2756d88c24e4520c9b07ed855fd9a27" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/98468fa7b2756d88c24e4520c9b07ed855fd9a27" data-origin-width="1275" data-origin-height="745"><div class="figcaption">DSVT논문중</div></div><p> </p>