que paso
NVIDIA introduced Scale-In, a networking architecture centered on BlueField-4 DPUs for the access, security, storage and operational services surrounding AI compute. The company says BlueField-4 can process these functions independently of host CPUs at up to 800 Gb/s.
BlueField-4 is described as the infrastructure processor for Scale-In across GPU servers, agentic CPU systems, storage systems and cloud services. NVIDIA says the chip combines a 64-core Grace CPU with inline acceleration engines, LPDDR5X memory, a PCIe Gen6 host connection and an 800 Gb/s network interface. The company says the inline engines can process packets, RDMA, storage protocols, encryption, firewall rules and policy enforcement at up to 800 Gb/s, reducing work assigned to both the DPU’s CPU and the host CPU.
NVIDIA says BlueField-4 provides four times the memory bandwidth and twice the network bandwidth of BlueField-3. Its DOCA software is intended to turn those hardware functions into deployable services. The post names DOCA Flow for packet-processing pipelines, DOCA PCC for programmable congestion behavior, DOCA Telemetry for device and service health, and the DOCA Platform Framework for provisioning, deployment and updates. The architecture also supports service function chaining so traffic can pass through the required infrastructure services.
The networking layer is NVIDIA Spectrum-X Ethernet. According to the post, BlueField-4 processes infrastructure services at each system while Spectrum-X carries traffic between the AI factory and surrounding users, applications, data sources, services and external storage. NVIDIA says the combination can support virtual private clouds with tenant isolation, enforce security policies outside the host operating system, accelerate storage protocols and provisioning, and provide fleet-wide telemetry.
For a Vera Rubin NVL72 configuration, NVIDIA describes 800 Gb/s of north-south BlueField-4 bandwidth and four 1.6 Tb/s east-west paths per compute tray, totaling 7.2 Tb/s of aggregate interface bandwidth. The company says this allows a common policy model across access and Scale-Out traffic without routing all east-west traffic through the 800 Gb/s interface. These figures are architecture specifications and vendor claims in the source, not independently verified measurements.
Lea la fuente principal: developer.nvidia.com ↗
Por qué es importante
The proposal treats data access, tenant isolation, security enforcement and telemetry as part of the AI infrastructure rather than as separate software layers running on general-purpose servers. If the claimed design works in production, it could reduce host overhead and make shared AI systems easier to operate at scale.
The proposal reflects a shift in how infrastructure vendors are designing for agentic workloads. Traditional cloud systems were built around general-purpose servers, software-defined networking and elastic resource allocation. NVIDIA argues that agentic AI factories create more continuous interaction among users, agents, applications, enterprise data and storage, making software-only control layers harder to scale without consuming compute resources needed by AI workloads.
A dedicated processing domain could have practical value for shared systems. NVIDIA says operators could provision virtual private clouds centrally while keeping tenant traffic on an accelerated fabric. It also says security policies could be enforced in hardware outside the host operating system, making them harder for tenant software to disable or bypass. That separation could simplify isolation between customers or workloads, although the source does not provide an independent security assessment.
Data movement is another important constraint. Training, retrieval and inference systems can have abundant accelerator capacity while waiting for data from storage. NVIDIA claims that its BlueField-4 and Spectrum-X storage path can deliver up to 1.45 times the throughput of off-the-shelf Ethernet. If reproduced across real deployments, such gains could improve accelerator utilization and reduce the amount of host CPU capacity consumed by storage virtualization, protocol processing and data movement.
The architecture also connects infrastructure operations to AI-factory economics. NVIDIA says BlueField-4 can onboard nodes, provision networking and storage, start servers and load host operating systems before tenant software begins running. Centralized provisioning and telemetry could reduce configuration differences and help operators locate access, storage or policy bottlenecks. However, faster deployment and better observability are expected benefits described by NVIDIA, not demonstrated outcomes in the source.
Qué ver a continuación
NVIDIA has not provided customer deployments, pricing, independent performance testing or a release timetable in the source. The key questions are whether the claimed throughput and isolation hold across mixed workloads, how broadly the hardware and software are available, and whether operators can use the architecture without adopting NVIDIA’s wider networking stack.
The immediate question is availability. The blog directs readers to BlueField platform specifications, DOCA resources and programming guides, but it does not state a shipping date, product pricing, regional availability or which parts of the Scale-In stack are ready for production. It also does not clarify whether customers can deploy the services selectively or must adopt BlueField-4, DOCA and Spectrum-X together.
Independent testing will be needed to assess the performance claims. Useful evidence would include results across mixed tenant traffic, storage protocols, encryption, threat detection and telemetry workloads rather than isolated peak throughput. Comparisons should also measure host-CPU use, latency, failure recovery, power consumption and the effect of policy enforcement on real training and inference jobs.
Security claims require particular scrutiny because the design places privileged controls outside tenant hosts. Future documentation should explain the trust model, update process, key management, isolation boundaries and recovery procedures if a DPU, control service or policy-management system is compromised. The source names BlueField Astra as a unified control point across north-south and east-west domains but does not describe its failure modes or provide an external audit.
The broader market test will be interoperability. AI operators may already use different Ethernet fabrics, storage systems, orchestration tools and security platforms. NVIDIA’s post describes a tightly co-designed stack around BlueField-4, DOCA, Spectrum-X and Vera Rubin. It remains unknown how much of the claimed benefit depends on that specific combination, how open the interfaces are, and whether competing infrastructure processors can provide comparable isolation and data-movement performance.
