NVIDIA DOCA · Lesson 1

The layered stack

You met DOCA as "the framework Check Point builds on." Now look inside: DOCA is a layered SDK — one foundation, many libraries, ready-made services on top. Get the layers and everything else has a home.

In the BlueField track you learned DOCA is "the CUDA for DPUs." This track opens it up. The first win is the simplest and most durable: DOCA's four layers, and where your code sits among them.docs

The one-sentence mental model

DOCA is built on DOCA Core, a unified foundation for the DOCA Libraries, which are packaged into ready-to-run Services and Applications — with the DPA subsystem for offloading code onto the BlueField-3 datapath processor.core

1 · The four layers

Layer 4 — DOCA Services / Applications / Tools (ready to run) containerized services reference apps & samples CLI tools
▲ built from…
Layer 3 — DOCA Libraries (the capabilities) Flow RDMA GPUNetIO Comch DMA Compress App Shield
▲ all stand on…
Layer 2 — DOCA Core (the foundation: device, memory, tasks, progress engine) one unified API for every library
▲ drives…
Layer 1 — Hardware BlueField DPU ConnectX NIC + DPA on BlueField-3
Read it bottom-up: hardware → one Core foundation → many Libraries → packaged Services/Apps. Your code usually calls the Libraries (Layer 3) on top of Core (Layer 2).

2 · What each layer is for

LayerWhat it gives you
DOCA CoreThe unified foundation: discover devices, register memory (zero-copy), submit async tasks, poll a progress engine. Same API on DPU or host.core
DOCA LibrariesThe capabilities — packet steering (Flow), RDMA, GPU networking (GPUNetIO), DMA, compression, host introspection (App Shield), and more. You call these.
Services / Apps / ToolsReady-made building blocks: containerized DOCA services, reference applications + samples to copy, and CLI tools.
DPA subsystemOffload communication-centric code onto the datapath processor on BlueField-3 (DPU-only).dpa
The line that unlocks the DGX Spark question (Lesson 3)

DOCA Core has the same API on BlueField and on a plain host — but a call returns DOCA_ERROR_NOT_SUPPORTED where the hardware can't do it.core That single design choice is exactly why "can I run DOCA without a DPU?" has a nuanced answer — hold that thought.

Check yourself

Every DOCA library is built on top of…
Which layer holds ready-to-run containerized building blocks?
The DPA subsystem offloads code onto hardware found on…
Primary source — read this next

NVIDIA's DOCA SDK Architecture page and the DOCA Overview. Match their diagram to the four layers above.