mcp-server-deep-researchでMCPサーバーのPromptsを試してみる

mcp-server-deep-researchでMCPサーバーのPromptsを試してみる

Clock Icon2025.05.01

こんばんは、製造ビジネステクノロジー部の夏目です。

今回少し面白いMCPサーバーを見つけたので試してみます。

mcp-server-deep-research

https://github.com/reading-plus-ai/mcp-server-deep-research/

MCP Server for Deep Researchは、複雑なトピックに関する包括的なリサーチを実施するために設計されたツールです。質問を深く掘り下げ、関連する情報源を見つけ、構造化された調査レポートを作成するのに役立ちます。

あなた専用のリサーチアシスタントとして、リサーチに関する質問を包括的で引用の多いレポートに仕上げます。

https://github.com/reading-plus-ai/mcp-server-deep-research/?tab=readme-ov-file#mcp-server-for-deep-research

deep-researchと名前がついていますが、ChatGPTのDeep Researchなどとは異なっていました。

何をしているかと言うとMCPのPromptsという機能を使ってプロンプトの生成を行っています。

Prompts

https://modelcontextprotocol.io/docs/concepts/prompts

再利用可能なプロンプトテンプレートとワークフローの作成

プロンプトにより、サーバーは再利用可能なプロンプトテンプレートとワークフローを定義することができます。プロンプトは、共通のLLMインタラクションを標準化し、共有するための強力な方法を提供します。

概要
MCPのプロンプトは定義済みのテンプレートで、以下のことが可能です:

  • 動的な引数を受け入れる
  • リソースからのコンテキストを含む
  • 複数のインタラクションを連鎖させる
  • 特定のワークフローをガイドする
  • UI要素としての表面(スラッシュコマンドなど)

https://modelcontextprotocol.io/docs/concepts/prompts#overview

MPCサーバーの機能の一つがPromptsです。

プロンプトのテンプレートを事前定義する機能で、引数などで受け取った値を使ってプロンプトを生成します。

MCPサーバーの機能としては、クライアントから引数を受け取りサーバー側で処理を行うToolsが有名で主に使われています。

今回のMCPサーバーではPromptsを使っています。

mcp-server-deep-researchが生成するPrompt

https://github.com/reading-plus-ai/mcp-server-deep-research/blob/main/src/mcp_server_deep_research/server.py#L34-L94

こちらが生成するプロンプトのテンプレートです。
{research_question}を引数として渡す形になります。

Find out if there is any further information on NVIDIA's Project DIGITS. という質問を入れて生成されたプロンプトは次のようになります。
(DeepLに NVIDIAのProject DIGITSについて続報がないか調べてくださいという日本語を英語に翻訳してもらった内容)

deep-research.txt
You are a professional researcher tasked with conducting thorough research on a topic and producing a structured, comprehensive report. Your goal is to provide a detailed analysis that addresses the research question systematically.

The research question is:

<research_question>
Find out if there is any further information on NVIDIA's Project DIGITS.
</research_question>

Follow these steps carefully:

1. <question_elaboration>
   Elaborate on the research question. Define key terms, clarify the scope, and identify the core issues that need to be addressed. Consider different angles and perspectives that are relevant to the question.
</question_elaboration>

2. <subquestions>
   Based on your elaboration, generate 3-5 specific subquestions that will help structure your research. Each subquestion should:
   - Address a specific aspect of the main research question
   - Be focused and answerable through web research
   - Collectively provide comprehensive coverage of the main question
</subquestions>

3. For each subquestion:
   a. <web_search_results>
      Search for relevant information using web search. For each subquestion, perform searches with carefully formulated queries.
      Extract meaningful content from the search results, focusing on:
      - Authoritative sources
      - Recent information when relevant
      - Diverse perspectives
      - Factual data and evidence

      Be sure to properly cite all sources and avoid extensive quotations. Limit quotes to less than 25 words each and use no more than one quote per source.
   </web_search_results>

   b. Analyze the collected information, evaluating:
      - Relevance to the subquestion
      - Credibility of sources
      - Consistency across sources
      - Comprehensiveness of coverage

4. Create a beautifully formatted research report as an artifact. Your report should:
   - Begin with an introduction framing the research question
   - Include separate sections for each subquestion with findings
   - Synthesize information across sections
   - Provide a conclusion answering the main research question
   - Include proper citations of all sources
   - Use tables, lists, and other formatting for clarity where appropriate

The final report should be well-organized, carefully written, and properly cited. It should present a balanced view of the topic, acknowledge limitations and areas of uncertainty, and make clear, evidence-based conclusions.

Remember these important guidelines:
- Never provide extensive quotes from copyrighted content
- Limit quotes to less than 25 words each
- Use only one quote per source
- Properly cite all sources
- Do not reproduce song lyrics, poems, or other copyrighted creative works
- Put everything in your own words except for properly quoted material
- Keep summaries of copyrighted content to 2-3 sentences maximum

Please begin your research process, documenting each step carefully.

これをClaude 3.7 Sonnetで日本語に翻訳してもらったものが次のものになります。

# 専門調査研究指示書

あなたは、特定のトピックについて徹底的な調査を行い、構造化された包括的なレポートを作成する専門研究者です。あなたの目標は、研究課題に体系的に取り組む詳細な分析を提供することです。

研究課題は以下の通りです:

<research_question>
NVIDIAのプロジェクトDIGITSに関する追加情報があるか調査してください。
</research_question>

以下の手順に注意深く従ってください:

1. <question_elaboration>
   研究課題を詳しく説明してください。主要な用語を定義し、範囲を明確にし、対処する必要のある中心的な問題を特定してください。質問に関連するさまざまな角度や視点を検討してください。
</question_elaboration>

2. <subquestions>
   あなたの説明に基づいて、研究の構造化に役立つ3〜5個の具体的な小質問を作成してください。各小質問は以下の条件を満たす必要があります:
   - 主要研究課題の特定の側面に対応している
   - ウェブ調査を通じて答えられるように焦点が絞られている
   - 全体として主要質問を包括的にカバーしている
</subquestions>

3. 各小質問について:
   a. <web_search_results>
      ウェブ検索を使用して関連情報を検索します。各小質問について、注意深く作成されたクエリで検索を実行してください。
      検索結果から意味のあるコンテンツを抽出し、以下に焦点を当ててください:
      - 信頼性の高い情報源
      - 関連性がある場合は最新情報
      - 多様な視点
      - 事実データと証拠

      すべての情報源を適切に引用し、広範囲な引用を避けてください。引用は各25語未満に制限し、情報源ごとに1つの引用のみを使用してください。
   </web_search_results>

   b. 収集した情報を分析し、以下を評価してください:
      - 小質問への関連性
      - 情報源の信頼性
      - 情報源間の一貫性
      - カバレッジの包括性

4. 美しくフォーマットされた研究レポートをアーティファクトとして作成してください。レポートには以下が含まれるべきです:
   - 研究課題を枠組みする序論
   - 各小質問の調査結果を含む別々のセクション
   - セクション間の情報の統合
   - 主要研究課題に答える結論
   - すべての情報源の適切な引用
   - 明確にするための表、リスト、その他のフォーマットを適宜使用

最終レポートは、整理され、注意深く書かれ、適切に引用されている必要があります。トピックについてバランスの取れた見解を提示し、制限事項や不確実性の領域を認識し、明確で証拠に基づいた結論を出すべきです。

以下の重要なガイドラインを忘れないでください:
- 著作権のあるコンテンツから広範囲な引用を提供しないでください
- 引用は各25語未満に制限してください
- 情報源ごとに1つの引用のみを使用してください
- すべての情報源を適切に引用してください
- 歌詞、詩、その他の著作権のあるクリエイティブ作品を複製しないでください
- 適切に引用された素材を除いて、すべてを自分の言葉で表現してください
- 著作権のあるコンテンツの要約は最大2〜3文に制限してください

研究プロセスを開始し、各ステップを注意深く文書化してください。

mcp-server-deep-researchでは調査のためのプロンプトを生成してくれます。

使ってみる

PyPIからMCPサーバーを取得する

README.mdによると uvx mcp-server-deep-researchで実行できるとあります。

しかし、PyPIを確認するとその名前のモジュールはありませんでした。

調べてみると、ほぼ同じ内容のモジュールが公開されていたのでそれを利用します。

https://pypi.org/project/deep-research-mcp-server/

READMEの内容が同じだったのでおそらく同じ物だと考えられます。

一応コードを確認したところ一部変更がされていましたが、元のコードで足りていなかったMCPサーバーを起動させる部分が追加されていました。
(他は全く同じだったので同じものと考えています)

追加されていたコード(少しだけ意訳)
import asyncio
# (中略)
if __name__ == '__main__':
    asyncio.run(main())

MCPサーバーを登録する

{
  "mcpServers": {
    "deep-research": {
      "command": "uvx",
      "args": ["deep-research-mcp-server"]
    }
  }
}

Claude Desktopで使う

今回はPromptsを使うので使用できるMCPクライアントが限られます。

CursorやClineなどでは対応していません。

普段から検証に使っているClaude Desktopでは使えるので、そちらで使います。

生成されるプロンプトが英語なので、質問内容は英語で用意します。

日本語

NVIDIAのProject DIGITSについて最初の発表以降続報がないか調べてください

DeepLで英語に翻訳

Find out if there has been any further news about NVIDIA's Project DIGITS since the initial announcement.

また、Promptsを使うときには少し手順があります。

Claude Desktopを立ち上げます。

4a5deffc-8719-480a-99c7-0067bacb358c

矢印の先をクリックします。

2935af89-ac24-464c-b855-a881e3c32706

連携サービスを選択をクリックし、アイコンがある方の deep-researchをクリックします。

d9722f90-ce03-4a21-8ee6-99cdfb66c4a7

Research_question*に先ほどの英文を入れます。

プロンプトを追加をクリックします。

1ceb872f-6265-48ea-9d50-d4ecc3b8fb3b

deep-research.txtというファイルが追加されています。
これをクリックすると生成されたプロンプトが表示されます。

9af5cc18-49dc-403f-91aa-84146ae88012

今回はネットでの検索を行うので、Claude Desktopのウェブ検索をオンにします。

d0dd492c-2555-4435-b82f-ed8c24defc05

矢印の先のアイコンをクリックし、ウェブ検索をオンにします。
ウェブ検索はProプラン以上じゃないと使えないので、そうじゃないときは検索用のMCPサーバーを使ってください。

オレンジのボタンをクリックして、プロンプトを流します。

97c86512-1756-4b10-b832-750b5d1263ab

結果

I'll help you with this research project on NVIDIA's Project DIGITS. Let me break down the research question, develop subquestions, and conduct thorough research to create a comprehensive report.

First, I'll search for information about NVIDIA's Project DIGITS to understand what it is and find recent news about it.

NVIDIA Project DIGITS latest news 2024

Let me search for more specific information about Project DIGITS and find out if there have been any updates since its initial announcement.

NVIDIA Project DIGITS renamed DGX Spark latest updates May 2025

Now I'll search for information about the availability and pricing of NVIDIA's DGX Spark (formerly Project DIGITS) and any comparisons with competitors.

DGX Spark availability pricing competitors Asus Ascent GX10 alternatives

Let me search for any additional information about other developments related to NVIDIA's DGX Spark and any benchmarks or real-world performance data that might be available.

DGX Spark real world performance benchmarks software compatibility latest updates

Based on my research, I'll now create a comprehensive research report on NVIDIA's Project DIGITS (now DGX Spark) as requested.

Research Report: NVIDIA's Project DIGITS (DGX Spark) - Updates Since Initial Announcement

Research Report: NVIDIA's Project DIGITS (DGX Spark) - Updates Since Initial Announcement

Executive Summary

This research report examines the evolution of NVIDIA's Project DIGITS from its initial announcement to its current form as DGX Spark. Originally unveiled at CES 2025 as Project DIGITS, this compact AI supercomputer was officially renamed DGX Spark at NVIDIA's GTC 2025 event in March 2025. This report investigates the specifications, capabilities, availability, pricing, competition, and market implications of this innovative product designed to bring data center-level AI computing to the desktop.

1. Introduction and Background

1.1 Initial Announcement

NVIDIA unveiled Project DIGITS at CES 2025 in January as a personal AI supercomputer designed to provide AI researchers, data scientists, and students with access to the power of the NVIDIA Grace Blackwell platform. This innovative device represents NVIDIA's strategy to democratize advanced AI capabilities by making them accessible at the desktop level.

1.2 Core Purpose

The primary purpose of Project DIGITS/DGX Spark is to enable users to develop and run inference on advanced AI models locally, then seamlessly deploy them on accelerated cloud or data center infrastructure. NVIDIA CEO Jensen Huang described it as bringing "the Grace Blackwell Superchip to millions of developers," empowering data scientists, AI researchers, and students to "engage and shape the age of AI."

2. Key Updates Since Initial Announcement

2.1 Rebranding to DGX Spark

Project DIGITS was officially renamed to DGX Spark at NVIDIA's GTC 2025 event in March 2025. This rebranding aligns the product with NVIDIA's broader DGX product line, which includes server-class AI systems.

2.2 Availability and Pricing

The system is now available for preorder with shipping expected to begin by summer 2025. The price has increased from the initially announced $3,000 to $3,999 for the 4TB model.

NVIDIA also offers a bundle option that includes two DGX Spark units with a proprietary connection cable designed for clustering.

2.3 Introduction of DGX Station

Alongside DGX Spark, NVIDIA also announced the larger DGX Station, a high-performance desktop supercomputer powered by the NVIDIA Blackwell Ultra platform. This more powerful system serves as a companion product targeting enterprise users and researchers with more demanding computational needs.

The DGX Station features a GB300 Grace Blackwell Ultra Desktop Superchip with 784GB of coherent memory space and support for networking at up to 800Gb/s through the NVIDIA ConnectX-8 SuperNIC.

3. Hardware Specifications and Capabilities

3.1 Processor and Architecture

DGX Spark is built around the NVIDIA GB10 SoC, which features a 20-core Armv9 processor consisting of ten Cortex-X925 cores and ten Cortex-A725 cores. This is paired with a Blackwell GPU featuring NVIDIA's latest fifth-generation Tensor Cores.

3.2 Performance

The GB10 Grace Blackwell Superchip delivers up to one petaflop (1,000 TOPS) of AI computing performance at FP4 precision for prototyping, fine-tuning, and running large AI models.

This level of performance was previously only available in data center environments, but now fits "snugly in the hands of Nvidia boss Jensen Huang" according to Tom's Hardware.

3.3 Memory and Storage

Each DGX Spark includes 128GB of unified, coherent memory and up to 4TB of NVMe storage. The system provides:

  • Memory bandwidth of 273 GB/s with 128GB 256-bit LPDDR5x
  • The ability to run AI models with up to 200 billion parameters locally
  • Support for connecting two DGX Spark systems to handle even larger models of up to 405 billion parameters using NVIDIA ConnectX networking
3.4 Connectivity and I/O

DGX Spark provides comprehensive connectivity with WiFi 7, Bluetooth 5.3, HDMI 2.1a video output, 10GbE networking, and four USB4 ports, all in a compact 150 x 150 x 50.5mm (5.9" x 5.9" x 2") form factor.

4. Software Ecosystem

4.1 Operating System

DGX Spark runs NVIDIA DGX OS, a Linux-based operating system based on Ubuntu 22.04. This specialized OS includes NVIDIA's AI software stack, optimizing the platform for AI development and inference.

4.2 AI Software Stack

Users can access an extensive library of NVIDIA AI software for experimentation and prototyping, including software development kits, orchestration tools, frameworks, and models available in the NVIDIA NGC catalog and on the NVIDIA Developer portal.

Specific capabilities include:

  • Fine-tuning models with the NVIDIA NeMo framework
  • Accelerating data science with NVIDIA RAPIDS libraries
  • Running common frameworks like PyTorch, Python, and Jupyter notebooks
  • Supporting NVIDIA Cosmos Reason world foundation model and NVIDIA GR00T N1 robot foundation model

5. Partner Ecosystem and Alternative Versions

5.1 OEM Partners

NVIDIA has partnered with several global system builders to develop versions of DGX Spark, including ASUS, Dell, HP Inc., and Lenovo. These partnerships extend the reach of DGX Spark technology through established vendor channels.

5.2 ASUS Ascent GX10

ASUS has introduced the Ascent GX10, a rival to NVIDIA's DGX Spark that uses the same GB10 Grace Blackwell Superchip. Priced at $2,999, it's $1,000 cheaper than NVIDIA's offering but comes with only 1TB of storage compared to DGX Spark's 4TB.

According to observers at GTC 2025, the ASUS model has a lighter chassis using more plastic compared to NVIDIA's more metal-dense design, potentially making it more portable.

5.3 Other OEM Variants

Dell's Pro Max with GB10 and HP's ZGX Nano AI Station are other alternatives to the DGX Spark, both utilizing the GB10 Grace Blackwell Superchip. These offer similar capabilities within each company's broader AI workstation portfolios.

6. Market Positioning and Competition

6.1 Target Audience

DGX Spark is primarily targeted at:

  • AI researchers and developers
  • Data scientists
  • Educational institutions and students
  • AI startups and small research teams
6.2 Key Competitors

AMD's Strix Halo Ryzen AI MAX+ 395 CPU/GPU offers specifications similar to DGX Spark, with 128GB of unified memory and a comparable memory bandwidth of 256 GB/s. Its main advantages over DGX Spark include x86 compatibility allowing it to run Windows as well as Linux, and a lower price point.

The Apple Mac Studio with M4 Max offers 128GB of RAM with double the memory bandwidth (546 GB/s) compared to DGX Spark, though it's limited to Apple's ecosystem of macOS.

6.3 Differentiation Factors

DGX Spark's key advantages include:

  • The NVIDIA CUDA ecosystem, which remains the dominant platform for AI development
  • Specialized AI optimization through the Grace Blackwell architecture
  • Seamless scaling from desktop to data center through NVIDIA's broader infrastructure

7. Key Use Cases and Applications

7.1 AI Model Fine-Tuning

DGX Spark enables fine-tuning of AI models up to 70 billion parameters, allowing developers to improve the performance of pre-trained models locally.

7.2 AI Inference

The fifth-generation Tensor Cores with FP4 support deliver 1,000 TOPS of AI computing performance, accelerating inference of state-of-the-art AI models for testing, validation, and deployment.

7.3 Data Science Workflows

DGX Spark accelerates end-to-end data science workflows with NVIDIA RAPIDS, from data preparation to model training, inference, and deployment.

7.4 Edge Application Development

The system provides a platform for developing robotics, smart cities, and computer vision solutions through NVIDIA frameworks including Isaac, Metropolis, and Holoscan.

8. Limitations and Considerations

8.1 Architecture Limitations

One criticism of DGX Spark is that, as an ARM system, it will not run Windows and thus is not suitable for gaming or general-purpose computing, though this aligns with its specialized focus on AI development.

Another limitation is the relatively modest memory bandwidth of 273 GB/s, resulting from the use of LPDDR5x RAM, which was likely chosen to balance performance with cost.

8.2 Price Considerations

At $3,999, the DGX Spark represents a significant investment, though NVIDIA justifies this premium through the system's specialized capabilities and integration with its broader AI ecosystem.

9. Conclusion and Future Outlook

NVIDIA's Project DIGITS, now DGX Spark, represents a significant step in democratizing AI computing by bringing data center-class capabilities to the desktop. The rebranding and recent updates have positioned it within NVIDIA's broader DGX portfolio, with clear pathways for scaling from desktop development to cloud or data center deployment.

The emergence of competitive alternatives from both NVIDIA's OEM partners and competitors like AMD and Apple indicates growing momentum in the personal AI computing space. This competition is likely to drive further innovation and potentially more accessible price points for these technologies.

As AI continues to transform various industries, DGX Spark's ability to run substantial models locally addresses a significant gap between consumer-grade AI capabilities and data center solutions. This positions it as an important tool for the next wave of AI innovation, particularly for researchers, startups, and educational institutions that might otherwise lack access to advanced AI infrastructure.

References

  1. NVIDIA Newsroom. (2025, January). "NVIDIA Puts Grace Blackwell on Every Desk and at Every AI Developer's Fingertips."
  2. NVIDIA. (2025). "NVIDIA Project DIGITS: The World's Smallest AI Supercomputer."
  3. Big Data Wire. (2025, January 10). "Inside Nvidia's New Desktop AI Box, 'Project DIGITS'."
  4. NVIDIA Newsroom. (2025, March). "NVIDIA Announces DGX Spark and DGX Station Personal AI Computers."
  5. HPC Wire. (2025, January 8). "NVIDIA Unveils Project DIGITS Personal AI Supercomputer."
  6. Arm Newsroom. (2025, January 7). "Arm-powered NVIDIA Project DIGITS Puts High Performance AI in the Hands of Millions of Developers."
  7. eMarketer. (2025, January 8). "Project Digits is Nvidia's affordable AI supercomputer that rivals some data centers."
  8. Tom's Hardware. (2025, January 7). "Nvidia's Project Digits desktop AI supercomputer fits in the palm of your hand."
  9. Tom's Guide. (2025, January 7). "Nvidia reveals Project DIGITS AI supercomputer."
  10. HotHardware. (2025, March). "NVIDIA Project DIGITS Renamed DGX Spark, DGX Station Unveiled For AI Domination."
  11. NVIDIA. (2025, March). "A Grace Blackwell AI supercomputer on your desk | NVIDIA DGX Spark."
  12. VideoCardz.com. (2025, March). "NVIDIA renames Project DIGITS to DGX Spark, Mini-PC with GB10 Grace Blackwell Superchip to launch at $3999."
  13. Beebom. (2025, March 19). "Nvidia Rebrands Project Digits as 'DGX Spark' and Unveils a Larger DGX Station."
  14. TechRadar Pro. (2025, March 22). "Project Digits is now DGX Spark: Nvidia raises its price by 33%."
  15. ServeTheHome. (2025, March 20). "The NVIDIA DGX Spark is a Tiny 128GB AI Mini PC Made for Scale-Out Clustering."
  16. TechRadar Pro. (2025, March 23). "Asus debuts its own mini AI supercomputer: Ascent GX10."
  17. TechRadar Pro. (2025, March 27). "The Ascent GX10 gets its own photoshoot at GTC 2025 as Asus's DGX Spark rival attracts attention."
  18. Jon Peddie Research. (2025, March 28). "If you wanted a supercomputer on your desk..."
  19. CNX Software. (2025, March 21). "NVIDIA DGX Spark - A desktop AI supercomputer powered by NVIDIA GB10 20-core Armv9 SoC."
  20. WCCFTECH. (2025, March 18). "NVIDIA Enters The AI PC Realm With DGX Spark & DGX Station Desktops."

I've created a comprehensive research report about NVIDIA's Project DIGITS and its evolution into DGX Spark. The report covers all the key aspects of your research question, including:

  1. The initial announcement and subsequent rebranding
  2. Hardware specifications and capabilities
  3. Software ecosystem and compatibility
  4. Partner ecosystem and alternative versions
  5. Market positioning and competition
  6. Key use cases and applications
  7. Limitations and considerations
  8. Future outlook

My research revealed several significant updates since the initial announcement:

  1. Project DIGITS was officially renamed to DGX Spark at NVIDIA's GTC 2025 event in March 2025.

  2. The price has increased from the initially announced $3,000 to $3,999 for the 4TB model.

  3. NVIDIA also introduced the larger DGX Station alongside DGX Spark, a high-performance desktop supercomputer powered by the NVIDIA Blackwell Ultra platform with 784GB of coherent memory space.

  4. Several OEM partners have created their own versions, with ASUS offering the Ascent GX10 at $2,999 (with 1TB storage instead of 4TB), and both Dell and HP developing similar alternatives.

  5. The system has been confirmed to support major AI models, including NVIDIA's own Cosmos Reason world foundation model and GR00T N1 robot foundation model, as well as models from DeepSeek, Meta, and Google.

Is there any particular aspect of the report you'd like me to explain or expand upon?

調査ができました。

英語なので、Sonnetでレポートを日本語に翻訳します。

8b7dbd18-124f-44a2-88fc-96aa43641893

プロンプト

レポートを日本語に翻訳してください

結果

報告書を日本語に翻訳いたします。

研究報告書:NVIDIAのProject DIGITS (DGX Spark) - 初期発表以降の最新情報

研究報告書:NVIDIAのProject DIGITS(DGX Spark)- 初期発表以降の最新情報

エグゼクティブサマリー

本研究報告書では、NVIDIAのProject DIGITSの初期発表からDGX Sparkとしての現在の形態までの進化を検証します。もともとProject DIGITSとして2025年1月のCESで発表されたこのコンパクトなAIスーパーコンピュータは、2025年3月のNVIDIA GTC 2025イベントで正式にDGX Sparkと改名されました。本報告書では、データセンターレベルのAIコンピューティングをデスクトップにもたらすこの革新的な製品の仕様、機能、入手可能性、価格設定、競合状況、および市場への影響について調査します。

1. 序論と背景

1.1 初期発表

NVIDIAは2025年1月のCESでProject DIGITSを発表しました。これはNVIDIA Grace Blackwellプラットフォームの能力にAI研究者、データサイエンティスト、および学生がアクセスできるようにするためのパーソナルAIスーパーコンピュータとして設計されました。この革新的なデバイスは、高度なAI機能をデスクトップレベルでアクセス可能にすることで、AIの民主化を図るNVIDIAの戦略を表しています。

1.2 主な目的

Project DIGITS/DGX Sparkの主な目的は、ユーザーが高度なAIモデルをローカルで開発および推論実行し、その後スムーズに加速されたクラウドまたはデータセンターインフラストラクチャにデプロイできるようにすることです。NVIDIAのCEOであるJensen Huangは、これを「Grace Blackwellスーパーチップを何百万人もの開発者にもたらし」、データサイエンティスト、AI研究者、学生が「AIの時代に参加し、それを形作る」ことを可能にするものと表現しています。

2. 初期発表以降の主な更新

2.1 DGX Sparkへのリブランディング

Project DIGITSは2025年3月のNVIDIA GTC 2025イベントで正式にDGX Sparkに改名されました。このリブランディングにより、製品はサーバークラスのAIシステムを含むNVIDIAの幅広いDGX製品ラインと一致するようになりました。

2.2 入手可能性と価格設定

システムは現在予約注文が可能で、2025年夏までに出荷が開始される予定です。価格は当初発表された3,000ドルから4TBモデルで3,999ドルに引き上げられました。

NVIDIAはまた、クラスタリング用に設計された専用の接続ケーブルを含む2台のDGX Sparkユニットのバンドルオプションも提供しています。

2.3 DGX Stationの導入

DGX Sparkと並行して、NVIDIAはより大型のDGX Stationも発表しました。これはNVIDIA Blackwell Ultraプラットフォームを搭載した高性能デスクトップスーパーコンピュータです。このより強力なシステムは、より高度な計算ニーズを持つ企業ユーザーや研究者をターゲットとした関連製品として位置づけられています。

DGX StationはGB300 Grace Blackwell Ultra Desktop Superchipを搭載し、784GBのコヒーレントメモリ空間とNVIDIA ConnectX-8 SuperNICを通じて最大800Gb/sのネットワーキングをサポートしています。

3. ハードウェア仕様と機能

3.1 プロセッサとアーキテクチャ

DGX SparkはNVIDIA GB10 SoCを中心に構築されており、10個のCortex-X925コアと10個のCortex-A725コアで構成される20コアのArmv9プロセッサを搭載しています。これはNVIDIAの最新の第5世代Tensorコアを搭載したBlackwell GPUとペアになっています。

3.2 性能

GB10 Grace Blackwellスーパーチップは、大規模AIモデルのプロトタイピング、微調整、実行のために、FP4精度で最大1ペタフロップ(1,000 TOPS)のAIコンピューティング性能を提供します。

この性能レベルは以前はデータセンター環境でのみ利用可能でしたが、現在ではTom's Hardwareによれば「NVIDIAのボスJensen Huangの手のひらにすっぽり収まる」サイズになっています。

3.3 メモリとストレージ

各DGX Sparkには128GBの統合コヒーレントメモリと最大4TBのNVMeストレージが含まれています。システムは以下を提供します:

  • 128GB 256ビットLPDDR5xによる273 GB/sのメモリ帯域幅
  • 最大2,000億パラメータのAIモデルをローカルで実行する能力
  • NVIDIA ConnectXネットワーキングを使用して2台のDGX Sparkシステムを接続し、最大4,050億パラメータのさらに大きなモデルを処理するサポート
3.4 接続性とI/O

DGX SparkはWiFi 7、Bluetooth 5.3、HDMI 2.1aビデオ出力、10GbEネットワーキング、4つのUSB4ポートなど包括的な接続性を提供し、すべてをコンパクトな150 x 150 x 50.5mm(5.9" x 5.9" x 2")フォームファクタに収めています。

4. ソフトウェアエコシステム

4.1 オペレーティングシステム

DGX SparkはUbuntu 22.04をベースとしたLinuxベースのオペレーティングシステムであるNVIDIA DGX OSを実行します。この専用OSにはNVIDIAのAIソフトウェアスタックが含まれており、AI開発と推論のためのプラットフォームを最適化しています。

4.2 AIソフトウェアスタック

ユーザーは実験とプロトタイピングのために、NVIDIA NGCカタログやNVIDIA Developerポータルで利用可能なソフトウェア開発キット、オーケストレーションツール、フレームワーク、モデルなど、広範なNVIDIA AIソフトウェアにアクセスできます。

具体的な機能には以下が含まれます:

  • NVIDIA NeMoフレームワークによるモデルの微調整
  • NVIDIA RAPIDSライブラリによるデータサイエンスの高速化
  • PyTorch、Python、Jupyterノートブックなどの一般的なフレームワークの実行
  • NVIDIA Cosmos Reason世界基盤モデルとNVIDIA GR00T N1ロボット基盤モデルのサポート

5. パートナーエコシステムと代替バージョン

5.1 OEMパートナー

NVIDIAはDGX Sparkの開発のために、ASUS、Dell、HP Inc.、Lenovoを含むいくつかのグローバルシステムビルダーとパートナーシップを結んでいます。これらのパートナーシップにより、確立されたベンダーチャネルを通じてDGX Spark技術の普及が拡大されます。

5.2 ASUS Ascent GX10

ASUSは同じGB10 Grace Blackwellスーパーチップを使用したNVIDIAのDGX Sparkのライバル、Ascent GX10を発表しました。価格は2,999ドルでNVIDIAの製品より1,000ドル安いですが、DGX Sparkの4TBに比べて1TBのストレージしか提供していません。

GTC 2025の観察者によると、ASUSモデルはNVIDIAのよりメタルが多いデザインと比較して、より多くのプラスチックを使用したより軽量なシャーシを持っており、潜在的により携帯性が高くなっています。

5.3 その他のOEMバリアント

DellのPro Max with GB10とHPのZGX Nano AI StationはどちらもGB10 Grace Blackwellスーパーチップを使用したDGX Sparkの他の代替品です。これらは各社のより広範なAIワークステーションポートフォリオ内で類似の機能を提供しています。

6. 市場ポジショニングと競合

6.1 ターゲットオーディエンス

DGX Sparkは主に以下を対象としています:

  • AI研究者と開発者
  • データサイエンティスト
  • 教育機関と学生
  • AIスタートアップと小規模研究チーム
6.2 主要な競合

AMDのStrix Halo Ryzen AI MAX+ 395 CPU/GPUはDGX Sparkと類似した仕様を提供し、128GBの統合メモリとDGX Sparkと同等の256 GB/sのメモリ帯域幅を備えています。DGX Sparkに対する主な利点には、Linuxだけでなく Windowsも実行できるx86互換性と、より低い価格設定が含まれます。

Apple Mac Studio with M4 Maxは、DGX Sparkと比較して2倍のメモリ帯域幅(546 GB/s)を持つ128GBのRAMを提供していますが、AppleのmacOSエコシステムに限定されています。

6.3 差別化要因

DGX Sparkの主な利点は以下の通りです:

  • AI開発の主要プラットフォームであり続けるNVIDIA CUDAエコシステム
  • Grace Blackwellアーキテクチャによる特殊なAI最適化
  • NVIDIAの広範なインフラストラクチャを通じたデスクトップからデータセンターへのシームレスなスケーリング

7. 主なユースケースとアプリケーション

7.1 AIモデルの微調整

DGX Sparkは最大700億パラメータのAIモデルの微調整を可能にし、開発者が事前訓練されたモデルのパフォーマンスをローカルで向上させることができます。

7.2 AI推論

FP4サポートを備えた第5世代Tensorコアは1,000 TOPSのAIコンピューティング性能を提供し、テスト、検証、およびデプロイのための最先端AIモデルの推論を加速します。

7.3 データサイエンスワークフロー

DGX SparkはNVIDIA RAPIDSによりデータ準備からモデルトレーニング、推論、デプロイメントまでのエンドツーエンドのデータサイエンスワークフローを加速します。

7.4 エッジアプリケーション開発

このシステムはNVIDIAのIsaac、Metropolis、Holoscanなどのフレームワークを通じて、ロボティクス、スマートシティ、コンピュータビジョンソリューションの開発のためのプラットフォームを提供します。

8. 制限事項と考慮事項

8.1 アーキテクチャの制限

DGX Sparkに対する批判の一つは、ARMシステムとしてWindowsを実行できず、ゲームや汎用コンピューティングには適していないことですが、これはAI開発に特化した専門的な焦点と一致しています。

もう一つの制限は、LPDDR5x RAMの使用により比較的控えめな273 GB/sのメモリ帯域幅です。これはおそらくパフォーマンスとコストのバランスを取るために選択されたものです。

8.2 価格に関する考慮事項

3,999ドルで、DGX Sparkは相当な投資を表しますが、NVIDIAはシステムの特殊な機能とより広範なAIエコシステムとの統合を通じてこのプレミアムを正当化しています。

9. 結論と将来の展望

NVIDIAのProject DIGITS、現在のDGX Sparkは、データセンタークラスの機能をデスクトップにもたらすことで、AIコンピューティングの民主化における重要なステップを表しています。リブランディングと最近の更新により、デスクトップ開発からクラウドまたはデータセンターへのデプロイメントへの明確な道筋を持つ、NVIDIAのより広範なDGXポートフォリオ内に位置づけられています。

NVIDIAのOEMパートナーやAMD、Appleなどの競合他社からの競争的な代替品の出現は、パーソナルAIコンピューティング分野での勢いの高まりを示しています。この競争はさらなるイノベーションと、これらの技術のより手頃な価格設定を促進する可能性があります。

AIがさまざまな産業を変革し続ける中、DGX Sparkが実質的なモデルをローカルで実行する能力は、消費者向けAI機能とデータセンターソリューションの間の重要なギャップに対応しています。これにより、特に研究者、スタートアップ、高度なAIインフラへのアクセスが欠如している可能性のある教育機関にとって、次のAIイノベーションの波のための重要なツールとして位置づけられています。

参考文献

  1. NVIDIA Newsroom. (2025, January). "NVIDIA Puts Grace Blackwell on Every Desk and at Every AI Developer's Fingertips."
  2. NVIDIA. (2025). "NVIDIA Project DIGITS: The World's Smallest AI Supercomputer."
  3. Big Data Wire. (2025, January 10). "Inside Nvidia's New Desktop AI Box, 'Project DIGITS'."
  4. NVIDIA Newsroom. (2025, March). "NVIDIA Announces DGX Spark and DGX Station Personal AI Computers."
  5. HPC Wire. (2025, January 8). "NVIDIA Unveils Project DIGITS Personal AI Supercomputer."
  6. Arm Newsroom. (2025, January 7). "Arm-powered NVIDIA Project DIGITS Puts High Performance AI in the Hands of Millions of Developers."
  7. eMarketer. (2025, January 8). "Project Digits is Nvidia's affordable AI supercomputer that rivals some data centers."
  8. Tom's Hardware. (2025, January 7). "Nvidia's Project Digits desktop AI supercomputer fits in the palm of your hand."
  9. Tom's Guide. (2025, January 7). "Nvidia reveals Project DIGITS AI supercomputer."
  10. HotHardware. (2025, March). "NVIDIA Project DIGITS Renamed DGX Spark, DGX Station Unveiled For AI Domination."
  11. NVIDIA. (2025, March). "A Grace Blackwell AI supercomputer on your desk | NVIDIA DGX Spark."
  12. VideoCardz.com. (2025, March). "NVIDIA renames Project DIGITS to DGX Spark, Mini-PC with GB10 Grace Blackwell Superchip to launch at $3999."
  13. Beebom. (2025, March 19). "Nvidia Rebrands Project Digits as 'DGX Spark' and Unveils a Larger DGX Station."
  14. TechRadar Pro. (2025, March 22). "Project Digits is now DGX Spark: Nvidia raises its price by 33%."
  15. ServeTheHome. (2025, March 20). "The NVIDIA DGX Spark is a Tiny 128GB AI Mini PC Made for Scale-Out Clustering."
  16. TechRadar Pro. (2025, March 23). "Asus debuts its own mini AI supercomputer: Ascent GX10."
  17. TechRadar Pro. (2025, March 27). "The Ascent GX10 gets its own photoshoot at GTC 2025 as Asus's DGX Spark rival attracts attention."
  18. Jon Peddie Research. (2025, March 28). "If you wanted a supercomputer on your desk..."
  19. CNX Software. (2025, March 21). "NVIDIA DGX Spark - A desktop AI supercomputer powered by NVIDIA GB10 20-core Armv9 SoC."
  20. WCCFTECH. (2025, March 18). "NVIDIA Enters The AI PC Realm With DGX Spark & DGX Station Desktops."

お求めいただいたNVIDIAのProject DIGITSに関する研究報告書を日本語に翻訳いたしました。この報告書には、初期発表から現在のDGX Sparkとしての位置づけまでの進化について詳細な情報が含まれています。

主な発見としては:

  1. Project DIGITSは2025年3月のNVIDIA GTC 2025イベントで正式にDGX Sparkと改名されました。

  2. 価格は当初発表された3,000ドルから4TBモデルで3,999ドルに引き上げられました。

  3. NVIDIAはDGX Sparkと並行して、より大型のDGX Stationも発表しました。これはNVIDIA Blackwell Ultraプラットフォームを搭載した高性能デスクトップスーパーコンピュータで、784GBのコヒーレントメモリ空間を備えています。

  4. ASUSはAscent GX10として2,999ドル(4TBではなく1TBストレージ)の独自バージョンを提供し、DellとHPも同様の代替製品を開発するなど、複数のOEMパートナーが独自のバージョンを作成しています。

  5. このシステムはNVIDIA独自のCosmos Reason世界基盤モデルやGR00T N1ロボット基盤モデル、またDeepSeek、Meta、Googleからのモデルなど、主要なAIモデルをサポートすることが確認されています。

報告書の特定の側面について、さらに詳しい説明が必要でしたらお知らせください。

日本語に翻訳できました。

まとめ

mcp-server-deep-researchでMCPのPropmptsを試しました。

MCPの使い道としてPromptsも興味深いですね。

Share this article

facebook logohatena logotwitter logo

© Classmethod, Inc. All rights reserved.