NVIDIA는 커스텀 XPU와 검증된 NVIDIA 인프라를 결합하여 AI 팩토리 가속화를 위한 플랫폼인 NVLink Fusion을 소개합니다.

AI factory economics are driven by delivered output: tokens per second, tokens per watt, cost per token, utilization and uptime. This demands AI infrastructure designed as a complete factory system rather than isolated accelerators. Hyperscalers and AI-native companies building custom XPUs must architect not just the processor itself but the entire platform, including scale-up and scale-out networking, rack architecture, factory software and supplier relationships. At production scale, this path is complex, expensive and slows time to market.

NVLink Fusion breaks this constraint by pairing custom XPUs with NVIDIA's established infrastructure, letting builders focus innovation on the processor while leveraging mature technology elsewhere.

For trillion-parameter models and mixture-of-experts workloads, scale-up fabric performance is critical. A modern scale-up solution must deliver on three dimensions: end-to-end network performance with in-network compute and mature software integration; factory resiliency through uptime, health monitoring and serviceability during operation; and platform maturity that reduces operational risk.

Sixth-generation NVLink brings XPUs into NVIDIA's scale-up domain across 72-XPU clusters with 3x lower end-to-end latency than off-the-shelf Ethernet solutions and 10x higher packet rates. The GB300 NVL72 system delivers significantly higher throughput and interactivity. NVLink roadmap includes domains up to 1,152 accelerators. NVLink-C2C connects XPUs to NVIDIA Vera CPUs or ecosystem CPUs with 6x better energy efficiency than PCIe.

Developing custom XPUs often involves underestimated complexity: integrating high-speed CPU and scale-up interfaces, sourcing networking solutions, designing compute and switch trays, architecting racks including cooling and power, integrating security and storage, and managing complex supplier ecosystems. NVLink Fusion provides all these components through an ecosystem spanning ASIC design, CPU, and optical interconnect partners. As Tim Wilson, vice president and general manager of data center silicon engineering at Intel, noted, NVLink Fusion lets customers choose the CPU architecture, performance level and software capabilities for their specific workloads.

NVLink Fusion adopters can use NVIDIA's MGX rack-scale architecture and supply chain established for systems like NVIDIA Vera Rubin NVL72. Jack Luoh, head of product and solution at QCT and Quanta Computer, stated that Vera Rubin NVL72 enables nearly 100% automation of system builds in manufacturing, with those investments applicable to XPU deployments using NVLink Fusion.

The NVIDIA AI infrastructure platform is vertically integrated yet horizontally open. Adopters can incorporate NVIDIA Rubin GPUs, Vera CPUs, co-packaged optical switches, ConnectX SuperNICs, BlueField DPUs, Mission Control software and complete rack solutions including Vera Rubin NVL72, Vera CPU Rack, LPX, STX and SPX systems.

AI factory planning begins long before final accelerators arrive, with power procurement, facility design, cooling, rack layout and networking requiring early commitment. Facilities locked to one chip face schedule risks. Different workloads favor different accelerators: XPUs, GPUs, CPUs and LPUs, potentially combined for training, post-training, reasoning, retrieval and serving. Vince Hu, corporate senior vice president and general manager of the data center and computing business group at MediaTek, emphasized that NVLink Fusion lets customers deploy rack solutions with NVIDIA GPUs while developing XPUs at their own pace.

NVLink Fusion uses a unified architecture where XPU and GPU systems share rack footprints, networking, cooling, power and management systems. Operators proceed with buildout while deferring the exact silicon mix, then reprovision capacity as workload demands, silicon supply and business priorities shift. Lie-Szu Juang, chair and chief strategy officer at GUC, noted that NVLink Fusion lets hyperscalers and custom ASIC designers integrate their own custom CPU or XPU while bridging NVIDIA technology with third-party processes into a unified rack-scale architecture.

Factory buildout is expensive and mistakes require costly rework. Infrastructure must be validated before construction. NVLink Fusion aligns with the NVIDIA DSX reference architecture for AI factories, codesigning buildings, power, cooling, compute and networking. The NVIDIA Omniverse DSX AI Factory Blueprint provides a digital twin and open reference design for gigawatt-scale AI factories, enabling partners to model facilities and technology together before deployment.

At rack level, serviceability drives performance. Reference compute trays feature 100% liquid cooling without fans, cables or hoses, and can be removed while the rest of the rack operates. NVLink Switch trays are also liquid cooled and support continued operation during service. CC Lee, senior hardware development manager at Annapurna Labs (Amazon), stated that NVLink Fusion enables access to proven NVL72 rack designs for faster time to market and multiple supplier options for customer delivery.

Software completes the factory. NVIDIA NCCL handles distributed workloads, Dynamo and NIXL manage disaggregation, and Mission Control provides cluster management, telemetry and debugging for mixed AI infrastructure. With NVLink Fusion, XPUs can integrate with a world-class AI platform, enabling hyperscalers and AI-native companies to build unified, semi-custom AI factories.