The story in one line
A DPU (BlueField-3) is a computer on the network card; Check Point AI Factory
Firewall runs on it via DOCA to secure AI-factory GPU servers inline — with
zero GPU/CPU impact.
BlueField-3 hardware
| Part | Spec (up to / typical) |
| Arm cores | Up to 16 × Armv8.2+ (Cortex-A78); 8 MB L2 + 16 MB LLC |
| ConnectX-7 NIC | 400 Gb/s, Ethernet or InfiniBand — the data path |
| Datapath Accelerator | 16 cores / 256 threads, programmable |
| HW accelerators | Crypto (IPsec/TLS), RegEx, compression/dedupe, storage |
| Memory / storage | DDR5 5600, ~32 GB onboard · ~128 GB SSD |
| Programmed via | NVIDIA DOCA framework |
| Successor | BlueField-4 (~6× compute, ~2× throughput) — roadmap |
Figures are "up to" and vary by SKU. Source: NVIDIA BlueField-3 datasheet.
CPU vs GPU vs DPU
| Socket | Job |
| CPU | General-purpose logic; runs the OS and apps |
| GPU | Parallel math; trains and runs AI models |
| DPU | Moves & guards data: networking, storage, security |
DOCA vocabulary
| Term | Is |
| DOCA | NVIDIA's SDK/framework to program BlueField ("CUDA for DPUs") |
| DOCA Argus | Agentless host introspection via direct memory access — sees all host processes |
Why security on the DPU
| Win | Because |
| Tamper resistance | Guard runs on separate silicon / own OS — survives a host compromise |
| Zero performance tax | Inspection uses DPU compute, not host CPU/GPU cycles |
| East-west coverage | Inline at every node → sees node-to-node traffic, not just the edge |
Check Point AI Factory Firewall — two layers
| Layer | Capabilities | Built on |
| 1 · Network | NGFW, identity-based access, IPS, virtual patching, anti-malware (inline, N-S + E-W) | DPU data path (ConnectX-7 + accelerators) |
| 2 · Host | All-process visibility; stops local code exec, priv-esc, malicious LLM processes | DOCA Argus (direct memory access) |
Threats it stops
| Threat | Layer |
| Data poisoning | Network |
| Model theft / exfiltration | Network |
| Model inversion · DoS · unauthorized access | Network |
| Malicious process in downloaded LLM | Host |
| Local code exec / privilege escalation | Host |
| Lateral movement between GPU nodes | Network (E-W) |
Deployment & naming
| Fact | Detail |
| Current name | AI Factory Firewall (was AI Cloud Protect, renamed Oct 2025) |
| Runs on | NVIDIA BlueField-3 DPUs; validated on NVIDIA RTX PRO Servers |
| Packaging | NGC container nvidia/doca-checkpoint/aicp |
| Management | Multi-tenanted, predefined AI security policies; standard orchestration |
| Status | GA, on-prem enterprise |