Upgrade to Pro — share decks privately, control downloads, hide ads and more …

SSII2026 [PT1] アクセラレーテッド・コンピューティングが切り拓く知能の最前線 ~...

Sponsored · Your Podcast. Everywhere. Effortlessly. Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.

SSII2026 [PT1] アクセラレーテッド・コンピューティングが切り拓く知能の最前線 ~生成AIからエージェンティックAI、そしてフィジカルAIへの進化~

More Decks by 画像センシングシンポジウム

Transcript

  1. Accelerated Computing Platform Company  1993 年創業  創業者及び CEO

    ジェンスン フアン  従業員 45,000人  2026会計年度売上高 2,159億ドル  時価総額 5.1兆ドル
  2. 一年周期リズム|スーパークラスター規模|フルスタック|CUDAエブリウェア NVIDIAで構築するAIインフラ Grace Blackwell MGX ノード NVLink スイッチ Quantum スイッチ

    Spectrum-X スイッチ AIスーパーコンピューティング専用チップ GPU | CPU | DPU | NIC | NVLink スイッチ | IB スイッチ | ENET スイッチ CUDA • DOCA • NCCL Cluster-Scale Software System Software Chip Software CUDA-X ライブラリ NIM CUDAで高速化された エージェント型AIライブラリ Omniverse CUDAで高速化された フィジカルAIライブラリ 加速された ソフトウェアスタック GB200 NVL72 SuperPOD
  3. AI の進化 エージェント AI は、より強力な AI アプリケーションを実現します AGENTIC AI PERCEPTION

    AI 音声認識 DEEP RECSYS 医療画像診断 GENERATIVE AI デジタルマーケティング コンテンツ作成 コーディングアシスタント カスタマーサービス 患者ケア 2012 ALEXNET PHYSICAL AI 家事アシスト 医療・介護サービス 製造・物流支援
  4. NVIDIA NIM NVIDIA NeMo 生成AIに向けたNVIDIA AI Enterprise デジタル ヒューマン 企業向け

    RAG パイプライン AI Agentを洗練する データフリーホイール 動画検索と要約 企業調査向け AI エージェント Blueprints
  5. SMBC-GIC: 資産運用レポーティ ング業務をNVIDIA Nemoと NIMで変革 NVIDIA Nemoと NIMを活用し た金融ド メ

    イン特化型AI開発を飛躍的に高速化 https://www.nvidia.com/en-us/on-demand/session/gtc25-s73894/ https://www.sbbit.jp/article/sp/162473 課題 資産運用レポート 件数の増大による、 作業量・ 修正 業務の負荷増大。 また、 担当者間での表現の不統 一・ 属人的な知識への依存。 ソリ ューショ ン NVIDIA Nemoを活用し 、 日本語LLM ”Swallow” を独自データ でファ インチューニング。 開発し たカスタ ムLLMをNVIDIA NIMでオンプレミ ス環境にデプロイ。 成果 開発効率の向上: Hugging Faceと 比べ学習速度 が最大44%向上。 推論高速化: 従来環境比(vLLM)でスループッ ト が 1.4倍に向上。
  6. リーズニングが100倍のトークンと収益を生み出す コンピュートによって増大するコスト - コスト/トークンを下げるにはフルスタックのイノベーションが必要 知識 事前学習済みモデルはインターネットの知識をシ ミュレートします 100 トークン |

    $0.50 / M トークン リーズニング リーズニングモデルは問題について考え抜きます ~10K トークン  100X コスト NVIDIA は、グラフィックスプロセッシングユニッ ト (GPU)、ハイパフォーマンスコンピューティ ングハードウェア、および人工知能 (AI) テクノ ロジーの設計と製造を専門とする多国籍テクノロ ジー企業です。同社は 1993 年に、Jensen Huang 、Chris Malachowsky、および Curtis Priem に よって設立されました。 NVIDIA はカリフォルニア州サンタクララに本社を 置き、コンピュータグラフィックス、ゲーム、AI の分野におけるリーディングカンパニーのひとつ に成長しました。 NVIDIAとは何ですか? Okay, so I need to figure out seating for a family dinner with 8 people. Let's list out the requirements first to make sure I don't miss anything. First, the family has 8 members. Let me confirm who they are. The user mentions: in-laws (which I assume are the wife's parents), parents (user's own parents), sister, little brother, wife, and the user themselves. That adds up to 8: user, wife, user's mom, user's dad, wife's mom, wife's dad, sister, and little brother. Wait, that's 7. Hmm, maybe I miscounted. Let me check again. The user says "my family of 8" which includes the user, their spouse (wife), their parents (mom and dad), their in-laws (wife's parents: mom and dad), sister, and little brother. That's 8: user, wife, user's mom, user's dad, wife's mom, wife's dad, sister, little brother. Yes, that's 8 people. Now, the constraints: 1. In-laws (wife's parents) don't get along with user's parents. So they should not be seated next to each other. 2. Sister needs to sit next to little brother to help him eat. So sister and little brother must be adjacent. 3. Wife doesn't want to sit next to user's mom. So wife and user's mom should not be adjacent. Additional considerations: Typically, seating arrangements at a rectangular table have two ends and sides. But sometimes it's a round table. The problem doesn't specify, so maybe I should assume a round table because it's common for family dinners and allows for a circular arrangement. If it's a round table, seating is in a circle, so everyone has two neighbors. If it's rectangular, the ends have only one neighbor, but the problem might be easier with a round table. Let me proceed with a round table unless specified otherwise. So, the goal is to seat 8 people around a round table with the given constraints. First, list all the people: 1. User (let's call him U) Wife (W) User's Mom (UM) User's Dad (UD) Wife's Mom (WM) Wife's Dad (WD) 7. Sister (S) Little Brother (LB) Constraints: • WM and WD (in-laws) should not be next to UM and UD (parents). So WM and WD cannot be adjacent to UM or UD. • S must sit next to LB. • W should not sit next to UM. Let me start by placing some people to satisfy the strictest constraints first. The sister and little brother needing to be adjacent is a strict requirement. So let's fix S and LB next to each other. Let's place them in two adjacent seats. Let's say positions 1 and 2: S at 1, LB at 2. Now, the next strict constraint is the in-laws (WM and WD) not being next to parents (UM and UD). So we need to ensure that WM and WD are separated from UM and UD by at least one seat. Also, W doesn't want to sit next to UM. Another thing to consider is that in many families, couples might want to sit together, but the problem doesn't mention that. However, it's possible that UM and UD would sit together, and WM and WD would sit together. But the problem doesn't specify, so maybe we can assume that couples can be separated if needed, but perhaps it's better to keep them together unless conflicting with constraints. Let's see. But let's proceed step by step. First, place S and LB next to each other. Let's fix them at positions 1 and 2. Now, we need to place the rest: U, W, UM, UD, WM, WD. We need to keep WM and WD away from UM and UD. Also, W should not be next to UM. Let me try to place the user and wife in positions that can act as buffers. Maybe placing U and W opposite each other or in positions that separate the conflicting parties. Alternatively, since in-laws (WM and WD) don't get along with parents (UM and UD), maybe we can place them on opposite sides of the table. Let's try to group the parents and in-laws into separate groups with others in between. Let me try to create two blocks: one with UM and UD, and another with WM and WD, separated by other people. But also, W shouldn't be next to UM. So W needs to be away from UM. Let me attempt a possible arrangement. Starting with S at 1, LB at 2. Then moving clockwise, let's try to place U and W in positions that can help separate the conflicting groups. Let's try: 1: S 2: LB 3: ? 4: ? 5: ? 6: ? 7: ? 8: ? We need to fill positions 3-8 with U, W, UM, UD, WM, WD. Let me consider placing the user (U) next to LB (position 2) to have a family member next to the little brother. So position 3: U. Then position 8 (next to S) could be someone else. But then U at 3 would be next to LB at 2 and position 4. Let's see. Alternatively, maybe place U and W opposite S and LB. Let's see, in an 8-seat round table, opposite of S (position 1) would be position 5. So position 5 could be U or W. Let's try placing U at 5 and W at 6, but need to check constraints. Wait, maybe a better approach is to divide the table into sections. Let's say S and LB are at 1 and 2. Then, to separate the in-laws from parents, we can place the parents (UM and UD) on one side and in-laws (WM and WD) on the other side, with U and W in between. 8人家族を円卓に調和よく着席させるには すべての制約を尊重した上で、最適な配置は次の とおりです。 : あなた 妻 お父さん お母さん 姉 弟 義母 義 父 主な制約事項の達成状況: • 義理の家族対実の家族:あなたの姉、兄、そし てあなたによって隔てられています • 姉と兄:助け合うために隣り合っています。 • 妻と母:隣り合っていません(あなたと父に よって隔てられています) 8人家族で円卓の夕食の席順を決めなければなりません。義理の両 親は私の両親と仲が良くありません。妹は弟の食事を手伝うため、 弟の隣に座らなければなりません。妻は私の母親の隣に座りたく ないと言っています。各人はどこに座ればよいでしょうか? • AI Factory Value 収益の増加 顧客体験 トークンの増加の 必要性 コストを削減します AIファクトリーの価値を加速する >$1/M トークン | >1T パラメータ | >300 TPS
  7. リーズニング時代におけるポストトレーニング スケーリング モデルに考えることを教えるには、多くの計算が必要 モデルの更新 学習 質問 100 トピック領域 1 Million

    トピックごとの質問 100トピック×100万質問×100回答×10,000トークン 100兆トークン … スコアの確認 学習モデル 1兆パラメータ 100 回答候補 それぞれ10,000トークン LLama3の事前学習の7倍以上
  8. NeMo Curator – 機能 重複排除と分類  語彙の重複排除– 同一(完全一致) またはほぼ同一(あいまい一致) 

    意味的重複排除– 正確な文章よりも、 その意味に重点  分類器モデル-データに情報を追加し たり、フィルタリングしたりするた めの最先端のオープンモデル. RAPIDS による GPU アクセラレーション  cuDF -重複排除および分類モデル用  cuML -セマンティック重複排除にお けるK-meansクラスタリング用  cuGraph –ファジー重複排除用 合成データの生成  事前構築済みのパイプライン -プロ ンプト生成、対話生成、エンティ ティ分類などのタスク向け  モジュール式 - NeMo Curatorの機能 を既存のパイプラインに簡単に統合 できます  OpenAI API に対応- カスタム 「Instruct」および「Reward」モデ ルを統合する GPUによる高速化機能を駆使して、より高い精度を実現
  9. 富士通: マルチエージェントで購買調達業務の自動化を実現 NVIDIA NIMを統合したFujitsu Kozuchi Physical AI 1.0を開発 課題 部署間や企業間の調整を含む複雑な

    業務ワークフロー、特に機密情報を 含む調達業務の自動化は困難だった ソリューション NVIDIA NIMと自社技術を統合した 「Fujitsu Kozuchi Physical AI 1.0」 を開発。 帳票理解・購買規約解析・適合 チェックにそれぞれ特化させた3種 類のAIエージェント(Takaneベース) が連携するマルチAIエージェントフ レームワークを構築。 成果 自社の購買部門の実証実験において、 発注確認業務工数を約50%削減。 NIMの活用により推論速度50%向上 を見込む https://global.fujitsu/ja-jp/pr/news/2025/12/24-02
  10. 博報堂テクノロジーズ: Agentic AIで自己強化型の広告運用サイクルを構築中 NVIDIA RAG BlueprintとNeMo Agent Toolkitを活用した広告特化型AIエージェントの開発 課題 広告・マーケティング領域において複雑化

    する消費者行動やリアルタイム性の高い施 策運用に対応するためには、複数のAIが協 調し自律的に判断する仕組みが不可欠。 ソリューション RAG BlueprintとNeMo Agent Toolkitを 活用して、広告制作を自律的に遂行する Agentic AIを開発。 実運用データから精度を継続的に改善す る”Data Flywheel ”の考え方も導入。 成果 マーケティング投資効率の向上: キャン ペーン成果を即座に次の施策へ反映する 「自己強化型の広告運用サイクル」の実現 を目指す 生活者体験の向上: パーソナライズされた 広告体験により煩わしさを軽減 https://prtimes.jp/main/html/rd/p/000000034.000113498.html
  11. 動画分析AIエージェント–全く新しい種類のアプリケーション カメラストリームやアーカイブビデオから知識と洞察を引き出す ビジュアルAI エージェント ビジョン言語 モデル コンピュータビ ジョンと追跡 大規模言語モデル グラフDB

    ベクトル DB 要約 Q & A アラート ユーザー ストリーミング 動画 画像/動画 アーカイブス エージェントにタスクを 割り当ててください。 希望する洞察の内容と、その共有 方法について説明してください。 AC1
  12. スライド 15 AC1 Should we really limit to three use

    cases? Alvin Clark, 2025-02-24T16:59:50.863
  13. NVIDIA AI Blueprint: 動画検索と要約(VSS) 動画分析AIエージェントの開発プラットフォーム AI Blueprint VLM NIMs CV

    NIMS LLM NIMs Graph DB Vector DB 要約 Q & A アラート ユーザー ストリーミング 動画 画像/動画 アーカイブス RAG NIMs エージェントにタスクを 割り当ててください。 希望する洞察の内容と、その共有 方法について説明してください。
  14. Metropolis VSS Blueprint ビジネス上のメリット: • ペガトロンは、毎月約1500万の複雑なデバイスを組み立 てています。 • 複雑な手作業の正確性と効率性を向上させることで、コ スト削減を実現

    • VSSはペガトロンの組み立て生産性のコパイロットとし て機能し、手作業による組み立てミスを検知して削減し ます 活用事例
  15. AIエージェントの進化 LLMやマルチターン・プロンプトから、自己修正を行う長期稼働エージェントまで ChatGPT Nov 2022 DeepSeek Jan 2025 OpenClaw Jan

    2026 基盤LLM、 マルチターン・プロンプト より良い出力を得るために振り返りを 行うリーズニング あなたの代わりに働く 自己進化型エージェント 100x 1000x
  16. エージェントAIの能力を飛躍させる新たな第4のスケーリング則 “インテリジェンス” 知覚AI 生成AI エージェントAI フィジカル AI Pre-training scaling Post-training

    scaling Test-time scaling “long thinking” Agentic Scaling “AI talking to AI” Pre-training scaling Post-training scaling Test-time scaling “long thinking” Agentic Scaling “AI talking to AI” Get models trained Scale and scaling efficiency Fine-Tune Pre-Trained Models Quantize to NVFP4 Optimize MoE models NVL72, Dynamo Low-latency and large context required at scale
  17. Alpamayo Autonomous Vehicle BioNeMo Biomedical AI GR00T Robotics Earth-2 AI-Physics

    Nemotron Agentic AI Cosmos Physical AI NVIDIAオープンソースAI分野へ貢献 NVIDIAがオープンモデルエコシステムを牽引 2025年にリリースされた言語モデル数 Source: Hugging Face Count excludes quantized models, but includes variants (e.g., reasoning vs. instruct) and fine-tunes of other models 0 25 50 75 100 Google IBM Granite AI2 Microsoft Mistral Z.ai DeepSeek Kimi Minimax Meta Llama OpenAI Alibaba Qwen NVIDIA VLM3D ViDoRe, MTEB, MMTEB OpenASR Physical Reasoning MVPBench, IntPHys, CasualVQA OCRBench PAIBench Artificial Analysis Output Artificial Analysis Intelligence Index
  18. データ ライブラリ モデル Blueprint 学習スクリプト NVIDIA Nemotron:エージェントAI向けのオープンモデル、データ、ライブラリ 事前学習トークン10兆個 教師ありファインチューニング サンプル1,300万件

    強化学習環境12 種類 強化学習タスク 90万件 セーフティトレース 1万1,000件 Nemotron 3 Nemotron VL Nemotron Speech Nemotron RAG Nemotron Safety NeMo Data Designer NeMo Curator Megatron Core NeMo RL NeMo Gym NeMo Evaluator NeMo Agent Toolkit Nemotron Cookbook Nemotron Recipes AI-Q Video Search and Summarization RAG
  19. NVIDIA Nemotron 3 ファミリー:オープンで効率的、かつインテリジェントなモデル群 Nano Super Ultra AVAILABLE NOW 30B-A3B

    120B-A12B 550B-A55B Nemotron 3 オープンデータセット NeMo Gym オープンリサーチ 学習前のトークン、学習後のサン プル、強化学習タスク 強化学習の環境とスキル 論文、サンプル AVAILABLE NOW Nano Omni New 30B-A3B New
  20. 現在利用可能なNVIDIA Nemotron モデル Nemotron 3 Nano Highest efficiency SLM Nemotron

    3 Super Highest efficiency LLM Llama Nemotron Ultra Highest accuracy powered by world knowledge リーズニング ビジョン 情報検索 コンテンツの安全性 Nemotron Nano 2 VL Video understanding and Image reasoning Nemotron Parse 1.1 Document extraction Nemotron 3 Nano Omni (coming soon) Enterprise-grade multimodal understanding Nemotron Page Elements Structured images detection & classification Nemotron Table Structure Rows, columns detection & preservation in markdown Nemotron Graphic Elements Chart component detection &classification Llama Nemotron Embed Multilingual text question-answering retrieval Llama Nemotron Rerank Multilingual fine-tuned reranking model Llama Nemotron Safety Guard Multilingual content safety Llama Nemotron Content Safety Multilingual content safety Nemotron Content Safety Reasoning Domain-specific safety policies Llama Nemotron Topic Control Prevent off-topic / unsafe deviations Nemotron Jailbreak Detection Detect adversarial or jailbreak prompts to block unsafe behavior スピーチ Nemotron ASR Real-time English speech recognition Magpie TTS Multilingual Research-ready, text-to-speech for natural, fine-tunable voices across 9 languages Nemotron 3 VoiceChat Full duplex speech-to-speech 12B model built for voice agents (Early Access)
  21. エージェント向け OpenShell セキュアランタイム 自律的で、長時間稼働し、自己進化するエージェントを安全に構築 Policy Engine Privacy Router Network Guardrail

    API / CLI / MCP Nemotron 3 Super Dynamo, TRT-LLM Cloud Frontier Models Filesystem Web Search IoT Dev Machine (CLI) Coding Agents OpenClaw Sandbox Guardrail Definition Terminal Skills Memory Omniverse CAD Isaac Sim ClawHub (脆弱なスキル) OpenShell
  22. OpenShell Policy Engine Network Guardrail Cloud Frontier Models Definition Terminal

    Skills Memory OpenClaw Runtime NemoClaw OpenClaw向けNVIDIA NemoClawの発表 Local Frontier Model Nemotron 3 Super Privacy Router
  23. AI の進化 エージェント AI は、より強力な AI アプリケーションを実現します AGENTIC AI PERCEPTION

    AI 音声認識 DEEP RECSYS 医療画像診断 GENERATIVE AI デジタルマーケティング コンテンツ作成 コーディングアシスタント カスタマーサービス 患者ケア 2012 ALEXNET PHYSICAL AI 家事アシスト 医療・介護サービス 製造・物流支援
  24. フィジカルAIはAIの次なる潮流 フィジカルAI 基盤モデル 工場 ロボット センサー トークン テキスト トークン アクション

    トークン “何でも聞いてください、または 私に何か指示してください” シティー 自動車 センサー入力 手順
  25. ロボット基盤モデル シミュレーションフレームワーク 合成データ生成パイプライン 学習 GB300 シミュレーション 推論 GR00T OMNIVERSE COSMOS

    Jetson Thor RTX 6000 Pro CUDA-X ライブラリ フィジカルAIの開発には3つのコンピューターが必要
  26. スライド 39 MH1 [@Pranjali Joshi] - colored ones Madison Huang,

    2026-05-31T02:33:31.739 PJ1 0 Done. Will add GIFs at the end. Pranjali Joshi, 2026-05-31T17:52:33.487
  27. データ ライブラリ モデル Blueprint 学習スクリプト NVIDIA Cosmos:フィジカルAI向けオープンモデル、データ、ライブラリ さまざまな運転シーンの動画 27万 合成ビデオ8万2千

    Cosmos Predict Cosmos Transfer Cosmos Reason Cosmos RL Cosmos Curator Cosmos Cookbook Cosmos Training Recipes Cosmos Dataset Search Video Search and Summarization GR00T-Mimic GR00T-Dreams
  28. 汎用ヒューマノイドロボットのための「Open Vision Language Action」 GR00T 1.7 Model • [New]商用利用可能: 商用ライセンスにより、本番環境への導入

    が可能になります。製品チームによる完全なサポートが受けられ ます。 • 長期タスクに対する論理的推論: タスクおよびサブタスクレベ ルでの論理的推論により信頼性が向上 • 性能向上: 動作の質と異体間の汎化性能が向上。1.6と比べて DROIDにおける成功率が10%以上向上 • 能力拡張: 新たに双腕およびロコマニピュレーションタスクに 対応 • 利用可能状況: 1.7EA版利用可能; 1.7 GA は7月予定 • 実世界のユースケースの解決を目指しています. • 6 LHA (Foxconnを含む) GitHub およびHuggingFaceにて入手可能
  29. データ ライブラリ モデル Blueprint 学習スクリプト NVIDIA Isaac GR00T:ロボット工学向けオープンモデル、データ、ライブラリ ロボットの軌跡 50万件

    SimReadyのアセット800個 ロボットによる把持 5,700万個 GR00T N1.6 Isaac Lab Isaac Lab-Arena Isaac Sim Cosmos-RL GR00T-Mimic GR00T-Dreams GR00T Training Scripts
  30. 操作のための合成軌道データの生成 わずかな人間のデモンストレーションから、 膨大な合成操作データへと拡張 新しい環境における新たな行動に関する 大量の合成データを生成 Pick up the object and

    place it on the pan Cosmos Predict 2 Unlimited “”Dreams” In New Environments Scalable Training Data Single Real-World Image Cosmos Reason Action Labelling Trained policy running on a real robot GR00T-Dreams GR00T-Mimic 100s of synthetic motion generations 1 million X training data 10s of demonstrations Omniverse Cosmos Transfer
  31. Isaac GR00T-Mimic Blueprint 数十の人間のデモンストレーションを100万倍のデータに増幅するデータパイプライン GR00T-Gen GR00T-Mimic Isaac Lab Isaac Lab

    数100 の合成モーション生成 100万倍 の学習データ 3D Cosmos Omniverse 数10 のデモンストレーション GR00T-Teleop Isaac Lab 迅速なデータ収集 データ収集時間を数時間から数分に短縮す る ソリューションまでの時間の短縮 新しいデータをより速く生成し、解決まで の時間を短縮する シミュレーションと現実のギャッ プを縮小する COSMOS WFMで生成データを必要な フォトリアリズムに拡張する
  32. . Unitree H2 Plus ロボット工学の研究開発向け、世界初のオープン・フロンティア型ヒューマノイドロボットのリファレンスデザイン Unitree H2 Body 6ft tall

    · 150lbs 31-DOF · WBC Jetson AGX Thor T5000 Blackwell GPU, 2070 FP4 TFLOPS • フルスタックなH2 Plusリファレンスデザインによる ヒューマノイドロボットの研究開発の拡張 • NVIDIA Jetson Thorのオンボードコンピューターにより、リアルタイム論理的推論と 制御を実現 • 設定済みのGR00Tワークフローにより、セットアップ時間を数日から数時間に短縮 • シミュレーションから実環境での導入に至るまで、人間サイズのヒューマノイドの試験 を加速 Training Scripts (Imitation learning) Training Evaluation Isaac Lab Arena (Sim evaluation) Isaac Teleop (Real & Sim) Data Creation Isaac Lab (Sim environment) Setup Deployment Isaac ROS (Real evaluation) Developer Validated Policy Robot Dual Sharpa Wave Hands 22-DOF · 5-finger dexterity Coming soon