Overview

Cloud-Class AI, Now on Your Desk

From hyperscale cloud to the desk beneath your monitor, a new class of AI workstations is emerging, powered by 20-core Arm configuration and the NVIDIA GB10 Grace Blackwell Superchip. Delivering petaflop-scale compute, unified memory, and support for models up to 200 billion parameters, these systems are redefining what’s possible on an AI workstation. With leadership from top OEMs and native software support from ecosystem partners, Arm-powered AI workstations bring performance per dollar, architectural efficiency, and developer readiness to the forefront of large-screen compute.

Explore Partner Solutions
Performance

From Cloud to Desk: Unmatched Performance with Arm-Powered AI Workstations

Desktop

Purpose-Built for Desktop-Class Performance

Arm architecture delivers workstation-grade performance for today’s most demanding workloads.

Cloud Desktop

From Cloud to Desktop Without Compromise

The same architecture powering the cloud now drives AI workstation innovation.

Performance

Performance Per Dollar that Scales

Arm-based desktops break the cost-performance ceiling.

Partner Solutions

Trusted by Leading OEMs

The world’s leading OEMs are launching Arm-based AI workstations powered by the NVIDIA GB10, offering up to 200-billion-parameter support, petaflop-scale performance, and industry-ready designs.

Acer Veriton GN100 AI Mini Workstation

View Product

ASUS Ascent GX10

View Product

Dell Pro Max with GB10

View Product

GIGABYTE AI TOM ATOM

View Product

HP ZGX Nano AI Station

View Product

ThinkStation PGX

View Product

MSI EdgeXpert MS‑C931

View Product
Technologies

Arm Technology Powers High-Performance AI Workstations

Workloads

Built for Modern Workloads

AI Researchers

Arm-powered AI workstations enable researchers to train and fine-tune large models locally, with support for up to 200 billion parameters. By bringing cloud-class performance to an AI workstation, they allow faster experimentation cycles, reduced infrastructure costs, and greater control over sensitive datasets.

AI Researchers

Healthcare

Arm-powered AI workstations enable local model fine-tuning for diagnostics and medical imaging, ensuring patient data stays secure while accelerating time-to-insight. By eliminating reliance on the cloud, healthcare organizations can achieve both compliance and faster iteration in life sciences and clinical research.

Healthcare

Architecture and 3D

From real-time rendering to large-scale simulations, Arm-based AI workstations provide uncompromising performance for architects, designers, and engineers. Unified memory and high-core efficiency accelerate workflows in 3D modeling, CAD, and visualization without cloud queue bottlenecks.

Architecture and 3D

Enterprise AI

Enterprises can prototype, train, and deploy AI models securely on -device, reducing cloud costs while maintaining control of sensitive data. With NVIDIA GB10 performance and native software compatibility, Arm-powered workstations bring agility and scalability directly into enterprise environments.

Enterprise AI
AI Researchers

Key Takeaways

  • Arm-based AI workstations deliver petaflop-scale AI performance with support for models up to 200 billion parameters, offering high efficiency and performance per dollar.
  • Top OEMs including Dell, HP, Lenovo, and Asus have launched Arm AI workstations, powered by NVIDIA GB10 and optimized for large-screen compute.
  • Arm Cortex-X925 and Arm Cortex-A725 CPUs with unified memory drive demanding workloads, from AI training to 3D rendering and simulation.
  • The NVIDIA AI stack runs natively on Arm, enabling seamless development with PyTorch, TensorFlow, Docker, and Hugging Face.
  • Ideal for AI research, healthcare, and enterprise use, Arm-based AI workstations enable fast iteration, data privacy, and local development of AI models.

FAQs

Arm 架構 NVIDIA GB10 Grace Blackwell 超級晶片的主要功能為何？
  • 20 核心 Arm CPU 配置結合 Arm Cortex-X925 及 Arm Cortex-A725 核心，支援各種人工智慧作業，包括協調管理、標記及強化學習。
  • 統一記憶體架構無需分離 CPU 與 GPU 記憶體，除了可減少負擔，也能在邊緣實現高達 2,000 億個參數的人工智慧模型。
  • 原生工具鏈可攜性能讓開發人員在雲端及邊緣環境中，使用 NVIDIA 人工智慧軟體堆疊與 CNCF 工具 (如 Docker 及 Kubernetes)。

為何開發人員或企業想在人工智慧工作站本機執行人工智慧模型？

在本機執行人工智慧模型有助於提升回應速度、加強資料隱私及作業獨立性。本機推論可協助開發人員及企業減少延遲、加強控制敏感資料，並在連網受限或無連網環境中確保可靠效能。

是什麼讓人工智慧工作站能夠執行大規模人工智慧模型？

GBX 等人工智慧工作站是以全新系統設計為基礎建構而成，將 CPU、GPU 及記憶體整合為單一互連架構。GBX 平台由 Arm 架構 NVIDIA Grace Blackwell 超級晶片驅動運作，結合 20 個 Arm 核心與次世代 Blackwell GPU，每秒可執行 1,000 兆次作業 (TOPS)，並配備 128 GB 的統一 LPDDR5x 記憶體。

此統一記憶體可讓龐大模型 (在單一系統高達 2,000 億個參數) 完全置於本機記憶體中，消除限制傳統 x86 工作站的資料傳輸瓶頸。GBX 也於 Arm 架構上執行 NVIDIA 完整的人工智慧軟體堆疊 (CUDA、TensorRT、PyTorch、TensorFlow 及 DGX 作業系統)，與 DGX 資料中心及雲端基礎設施使用的堆疊相同。

GBX 基本上會將原本需要的伺服器機櫃，縮小為以 200 W USB-C 供電的 6 吋立方體，讓本機人工智慧開發真正成為在雲端租用 GPU 的替代方案。

在工作站執行人工智慧對研究人員、科學家或企業有何實際價值？

對非常大型的模型而言，執行及微調作業過去只能在雲端進行，不過 GBX 等人工智慧工作站能夠實際執行前述作業，還能提供足夠的裝置內效能用於實際實驗及迭代，同時讓資料留在本機及維持成本可預測性。重點並不是達到同等的雲端規模，而是直接將有意義的人工智慧功能導入桌上型設備。

我可以在採用 Arm 技術的人工智慧工作站以原生方式執行 PyTorch 或 TensorFlow 嗎？

可以。NVIDIA 人工智慧軟體堆疊包括 PyTorch、TensorFlow、Docker 及其他廣泛使用的框架，可在 Arm 上原生執行無需模擬。開發人員可在採用 Arm 技術的人工智慧工作站上，使用目前所用的工具鏈建構、訓練及部署模型，確保無縫工作流程及加速迭代。