BlueField-3 & AI Factory Firewall · Lesson 2

DOCA & the isolation model

Two ideas that make DPU security possible: DOCA (how you program the DPU) and isolation (why a guard that lives outside the house can't be bribed by a burglar inside).

In Lesson 1 you saw BlueField-3 is a real computer on the network card. Two things turn that hardware into a security platform: the DOCA software framework, and the isolation that comes from running outside the host. Get these and Check Point's product (Lesson 3) is obvious.

1 · DOCA — the "CUDA for DPUs"

You don't get BlueField's value by plugging it in; you get it by programming it. DOCA is NVIDIA's software framework/SDK for BlueField — the way vendors like Check Point build services that run on the DPU and drive its accelerators.doca

Security software (e.g. Check Point AI Factory Firewall) firewall · IPS · host introspection
▼ built on DOCA APIs & services
NVIDIA DOCA — framework / SDK / services networking storage security (incl. Argus)
▼ drives the silicon
BlueField-3 hardware (Lesson 1) Arm cores ConnectX-7 crypto / RegEx accelerators
DOCA is the bridge: it lets software use the DPU's accelerators at line rate. "CUDA unlocked the GPU; DOCA unlocks the DPU."

2 · The isolation argument

Here's the security insight that makes DPUs special. A traditional host agent runs inside the machine it protects — so if an attacker takes over the host, they can often disable or blind the agent. The DPU runs the guard on separate silicon with its own OS, outside the host's reach.

Traditional: agent inside the host host OS
security agent (same OS)
⚠ host compromise can disable the guard
DPU: guard outside the host host OS + GPUs
│ PCIe boundary
security on the DPU (own OS)
✓ stays standing if the host falls
Same job, different vantage point. Outside-the-host = harder to tamper with, and it costs the host zero CPU/GPU.
Two wins from one placement

(1) Tamper resistance — the guard survives a host compromise because it lives on separate silicon. (2) Zero performance tax — inspection runs on the DPU's own cores/accelerators, so no GPU or host-CPU cycles are stolen from the AI workload.brief

3 · DOCA Argus — seeing into the host from outside

Isolation raises a question: if the guard is outside the host, how does it see what's running inside? Answer: DOCA Argus — a DOCA service that uses the DPU's direct memory access to read the host's memory and enumerate every running process, with no agent installed on the host.cp

Host (AI node) — no agent installed process A process B malicious process?
▲ DOCA Argus reads memory directly (agentless)
DPU — DOCA Argus enumerate all processes · spot runtime attacks
Agentless host introspection from the DPU: full process visibility an in-host attacker can't easily hide from. This is the engine behind Check Point's host-level layer.
Keep the layers straight (you'll be quizzed)

DOCA = the framework you program the DPU with. DOCA Argus = the specific DOCA service for agentless host process visibility via direct memory access. Check Point = the vendor whose firewall + threat prevention is built on DOCA and uses Argus. Don't collapse these into one word.

Check yourself

What is NVIDIA DOCA?
Why is DPU-hosted security harder for an attacker to disable?
How does DOCA Argus see processes on the host?
Primary source — read this next

The NVIDIA DOCA documentation (framework + services, including Argus), and Check Point's AI Factory Firewall solution brief for how these ideas are used in the product.