NVIDIA is on stage at Hot Chips 2026, walking through its BlueField-4 Processor, the fourth generation of its data center processing unit. I think we are getting more of the Grace with ConnectX-9 version, not the bigger Vera version, so that is neat.
This is being done live, so please excuse any typos.
NVIDIA BlueField-4 DPU at Hot Chips 2026
BlueField-4 sits inside NVIDIA’s larger argument that agentic AI is the most complex computing workload yet. NVIDIA positions the DPU as the component that ties the CPU, GPU, and network together, enabling an agent to run through the observe, reason, and act steps. I think this is the third time we have seen this image at Hot Chips?
NVIDIA frames BlueField-4 within the Vera Rubin platform, which it describes as a full-stack AI factory spread across seven chips and five racks. BlueField-4 is one of those seven chips, working alongside the Vera CPU and Spectrum-6. Again, it seems like one we have seen before a few times this conference.
Here, NVIDIA has a DPU handling networking, security, and storage in every data center node. General-purpose compute and business logic run on the CPU, while parallel AI and HPC workloads run on the GPU.
NVIDIA casts the DPU as the trusted entity of the platform, a server in front of the server that isolates the host from the data center fabric and owns the tenant data path. That framing is central to everything that follows in the talk.
Scale-up is the domain NVIDIA covers through NVLink, letting multiple GPUs operate as one system inside a tenant. That keeps GPUs fed within a single node before any fabric comes into play.
Scale-out is where the fabric takes over, spanning ConnectX and Spectrum-X switches to turn the data center into a single unit of compute. NVIDIA keeps scale-up and scale-out as separate layers here before introducing scale-in later.
Every Vera Rubin system gets a BlueField co-designed in, NVIDIA says, combining security, data, and infrastructure services at the platform level. That is a shift from the optional plugging in of a DPU card to an existing server. Really, this is a central idea since NVIDIA is now focused on data center scale. Hyper-scalers run machines at scale and use DPUs. So it makes sense that the DPUs are part of the NVIDIA AI Factory design.
Now NVIDIA announces the BlueField-4 Spectrum-X scale-in network, an end-to-end data center networking solution aimed at mixed agentic workloads. Scale-in handles the traffic between the host and the wider data center, often across multiple tenants and applications.
Scale-in gets its first numbers with storage access. NVIDIA shows Spectrum-X Ethernet delivering 1.5x faster access for 5 GB files compared to off-the-shelf Ethernet, with 1.4x for 10 GB files and 1.3x for 50 GB files.
A useful DPU needs both serious network acceleration and serious compute. NVIDIA pairs ConnectX-9-class networking with Grace CPU-class cores to run control plane tasks, encryption, and storage I/O at high rates, overlapping with our earlier look at NVIDIA’s Vera CPU.
Earlier cloud DPUs were built for general-purpose cloud computing, recalling the one-size-fits-all approach shown at Hot Chips 2021. That older model treated the DPU as a replaceable add-in, and we have tracked NVIDIA’s DPU roadmap as it folded Arm cores, GPU, and networking into one package.
Agentic AI needs a different class of infrastructure, and NVIDIA lays out the contrast in terms of bandwidth alone, moving from 200 Gb/s-class cloud DPUs to 7 Tb/s aggregate bandwidth on the AI DPU. That older card plugs into a PCIe slot while BlueField-4 is co-designed at the system level.
BlueField-4 combines a Grace CPU with 64 Arm Neoverse V2 cores at 1.7 GHz, LPDDR5 bandwidth of 275 GB/s, and ConnectX-9 networking with 800G Ethernet across 200G PAM4 SerDes, all with inline encryption and a PCIe Gen6 x16 host link. That 1.7GHz seems slower than what we are seeing on the GB300 AI station systems and what we have seen in Grace servers. Likely it is just to operate at a lower power envelope. Also, there are only 64 Arm cores active. This slide probably could have used a bit more on SPECINT 220. SPEC CPU2017/CPU2026, int rate/peak, estimated? That is really academic because more bigger and faster cores are going to get you huge performance gains.
If you are wondering, we expect these to be much faster than the Xsight Labs E1 DPU because these are V2 versus N2 cores.
Vera Rubin presents an extreme bandwidth problem. Every compute tray needs 7 Tb/s of aggregate network bandwidth, built from four 1.6 Tb/s GPU scale-out links plus 800 Gb/s of scale-in to the DPU
Without a DPU owning the data path, NVIDIA argues that the GPU network cannot be secured end-to-end and isolated from the tenant. Scale-in traffic gets isolated under the traditional approach, but scale-out traffic does not.
A non-DPU approach also multiplies power consumption, because each NIC requires a separate CPU, memory subsystem, and management connection. NVIDIA puts that at 4x higher power versus putting BlueField-4 in front of the host.
BlueField-4 changes that with Astra, NVIDIA says, delivering 7 Tb/s of CSP-grade security and management where the DPU owns the data path. Zero trust is enforced and telemetry collected at line rate without touching the host. Astra uses links (sound like PCIe in Vera Rubin and from what the demo looks like) for all of the management and so forth so that way NVIDIA does not need a BlueField-4 DPU on all of the ConnectX-9 links.
This scaling story holds up on paper here, with measured bandwidth tracking ideal linear scaling as NIC count rises. Encrypted and virtualized scale-out traffic remains secure and monitored as the platform scales to multi-terabit aggregates. Broadcom just showed a more detailed performance view, but BlueField-4 is a completely different class of device versus Thor Ultra.
A quick live demonstration shows the 7 Tb/s platform-level DPU running in a Vera Rubin system. Get your ticket to Hot Chips 2026 virtual if you want to see that. NVIDIA is showing how the BlueField-4 DPU can see all of the ConnectX-9 NICs and the traffic going across those NICs.
Storage access for the Rubin GPUs is being highlighted here. NVMe over fabrics running on BlueField-4 Grace delivers 1.6 Tb/s with 8 cores and reaches 20 million IOPS with 16 cores, providing a 2x faster path to data for the GPUs. This is a neat chart.
On the Vera CPU side, BlueField-4 makes agentic security transparent through in-silicon policy enforcement. DOCA Flow, DOCA Vault, and DOCA Argus map file, object, processor, and network context to keep policy enforcement at the speed of the agents.
Small-packet traffic is where control planes usually fall apart, and this figure centers line-rate networking for the AI infrastructure. BlueField-4 keeps forwarding and security checks moving without collapsing under tiny frames.
Storage-Scale extends agent context across a tiered memory map. NVIDIA shows 3.2 Tb/s storage access with 10x IOPS and 5x efficiency handled by DOCA MEMOS, moving key-value context from GPU HBM through system memory to local and network storage. This is “Two BlueField-4 Vera devices”. NVIDIA uses “BlueField-4” for both the Grace and single ConnectX-9 version up to the single and dual Vera with several ConnectX-9 versions. I have given the feedback that it would make it easier to have BlueField-4G (Grace) and BlueField-4V (Vera) or something just to help differentiate the versions.
NVIDIA BlueField-4 is framed as the first AI-native DPU across scale-up, scale-out, scale-in, and storage-scale.
Neat talk and a super presentation.
Final Words
NVIDIA’s pitch is that the DPU stops being an add-in and becomes the trusted front end of the AI factory, and that is a meaningful shift for anyone building rack-scale systems. BlueField-4 relies on a considerably more powerful CPU than prior parts, giving the architecture the headroom to handle security and storage workloads that earlier generations struggled with. BlueField-4 feels like an enormous generational jump.
