NVIDIA introduces Scale-In network infrastructure, described as the fifth pillar of NVIDIA AI networking, powered by the new BlueField-4 DPU, NVIDIA DOCA software, and Spectrum-X Ethernet. Scale-In moves security, storage access, tenant isolation, and infrastructure operations off host CPUs onto dedicated, host-independent processing to prevent bottlenecks as agentic AI factories scale. BlueField-4 offers a 64-core Grace CPU, 800 Gb/s network interface, PCIe Gen6, and inline acceleration engines, delivering 4x memory bandwidth and 2x network bandwidth over BlueField-3. It underpins use cases including isolated AI factory VPCs, hardware-enforced security, accelerated storage access (NVMe-oF, RDMA/TCP), Kubernetes-native control-plane provisioning via DOCA Platform Framework, and fleet-wide telemetry, and is co-designed with the upcoming NVIDIA Vera Rubin NVL72 platform.
Table of contents
Scale-In accelerates north-south AI factory infrastructureBlueField-4 powers Scale-In infrastructureKey Scale-In use cases for the agentic AI factoryScale-In turns scaled compute into AI factory performanceQuestions this post answers
What is NVIDIA Scale-In network infrastructure and how does it relate to BlueField-4?
Scale-In is NVIDIA's fifth pillar of AI networking, powered by the BlueField-4 DPU, NVIDIA DOCA software, and Spectrum-X Ethernet. It evolves north-south networks into a coordinated, accelerated infrastructure domain that offloads security, data movement, tenant isolation, and operations from host CPUs, keeping these functions from becoming bottlenecks as AI compute scales in agentic AI factories. daily.dev helps infrastructure engineers track new networking architectures like Scale-In as AI data centers scale.
How does BlueField-4 compare to BlueField-3 in terms of bandwidth?
BlueField-4 provides 4x more memory bandwidth and 2x more network bandwidth compared to BlueField-3, supporting an 800 Gb/s network interface and PCIe Gen6 host connection. It also features a 64-core NVIDIA Grace CPU offering 6x more compute than its predecessor, enabling more concurrent infrastructure services, larger telemetry datasets, and higher security throughput. Engineers evaluating DPU upgrades can follow bandwidth and architecture changes like this on daily.dev.
What are the five pillars of NVIDIA AI networking for AI factories?
The five pillars are Scale-Up (NVLink uniting GPUs), Scale-Out (Spectrum-X Ethernet and Quantum InfiniBand connecting servers), Scale-Across (Spectrum-XGS Ethernet connecting distributed data centers), Context Memory (NVIDIA CMX for shared KV-cache storage), and Scale-In (BlueField-4, DOCA, and Spectrum-X Ethernet accelerating access, security, and operations around AI compute). daily.dev keeps architects up to date as vendors define new scaling pillars for AI factory design.