The stack (bottom-up)
| Layer | Is |
|---|
| Services / Apps / Tools | Ready-to-run containers, reference apps + samples, CLI tools |
| DOCA Libraries | The capabilities — Flow, RDMA, GPUNetIO, Comch, DMA, Compress, App Shield… |
| DOCA Core | Unified foundation: device, memory, tasks, progress engine. Same API on DPU + host |
| Hardware | BlueField DPU · ConnectX NIC · DPA (on BlueField-3) |
DOCA Core has one API for BlueField and host; unavailable calls return
DOCA_ERROR_NOT_SUPPORTED.
The five Core objects
| Object | Role |
|---|
doca_dev | Discover + query a hardware acceleration unit |
doca_mmap / doca_buf / _inventory | Zero-copy: register memory once, hand HW buffers over it |
doca_ctx | A configured, started library instance you submit tasks to |
| task → event | Async unit of work; completion returns as an event |
doca_pe (Progress Engine) | Poll it to advance tasks and receive completions |
The universal loop
// setup (once)
open doca_dev → create doca_mmap + bufs → start doca_ctx → attach doca_pe
// data path (repeat, async)
submit task → while(!done) doca_pe_progress() → handle completion event
Libraries you'll reach for
| Library | Accelerates | Where |
|---|
DOCA Flow | HW packet steering as pipes (match & act) | DPU / host+ConnectX |
DOCA RDMA | Remote direct memory access | DPU / host+ConnectX |
DOCA GPUNetIO | CUDA kernel drives the NIC (GPU-initiated net) | needs GPU + NIC |
DOCA Comch | Host ↔ DPU control channel | DPU pair |
DOCA DMA / Compress | HW copy / (de)compress | DPU / host+ConnectX |
DOCA DPA | Offload comms code to the DPA processor | BlueField-3 only |
DOCA App Shield (Argus) | Agentless host process introspection | DPU only |
DPU vs DOCA-Host
| Mode | Runs on | Extra powers |
|---|
| On the DPU | BlueField's own OS | On-DPU services, DPA, App Shield/Argus, isolation |
| DOCA-Host | A host + ConnectX NIC | NIC-accelerated libraries only (no DPU-only features) |
DGX Spark (GB10) — the verdict
| Spec | Value |
|---|
| Chip | GB10 Grace Blackwell; 20-core Arm; ~1 PFLOP FP4 |
| Memory | 128 GB unified LPDDR5x; 4 TB SSD |
| Networking | ConnectX-7, 200 GbE, 2× QSFP — no BlueField DPU |
| DOCA on Spark | Status |
|---|
| DOCA-Host: Flow, RDMA, Ethernet, DMA | ✓ works (ConnectX-7) |
| GPUNetIO | ✓ works — CPU pinned mem (CPU_GPU) + CPU proxy |
| GPUDirect RDMA | ✗ not supported (unified-memory topology) |
| DPA · on-DPU services · App Shield/Argus | ✗ needs BlueField |
Pocket answer
Yes — enjoy DOCA on DGX Spark as a DOCA-Host + ConnectX-7 + GPU dev box. No to the
BlueField/DPU half (DPA, on-DPU services, App Shield), because Spark has no DPU.