Short answer: yes — as a DOCA-Host developer box on its ConnectX-7 — but a DGX Spark is not a BlueField DPU, so a whole slice of DOCA stays out of reach. Here's the precise line.
This is the question that opened the track. To answer it well you need one fact about the hardware and one fact about DOCA — both of which you now have.core
(1) DGX Spark has a GPU and a ConnectX-7 NIC (200 GbE) — but
no BlueField DPU.specs
(2) DOCA runs either on a DPU or on a host (DOCA-Host) with a
ConnectX NIC — same API, but DPU-only features return DOCA_ERROR_NOT_SUPPORTED
elsewhere.arch
Put them together and the answer writes itself.
| On DGX Spark | Status |
|---|---|
| DOCA-Host on ConnectX-7 (Flow, RDMA, Ethernet, DMA) | ✓ works — it's a ConnectX NIC |
| DOCA GPUNetIO (GPU + NIC on one box) | ✓ works — with a caveat (see below) |
| On-DPU services · DPA offload | ✗ needs BlueField |
| App Shield / Argus host introspection | ✗ needs BlueField (the AI Factory Firewall model) |
| GPUDirect RDMA (peer ↔ GPU memory) | ✗ not on Spark — unified-memory topology |
DGX Spark does not support GPUDirect RDMA (its unified-memory topology means
pinned device memory isn't coherently reachable by PCIe peers). DOCA GPUNetIO still runs — you use
CPU pinned memory (DOCA_GPU_MEM_TYPE_CPU_GPU) and enable
CPU proxy mode for transmits.nvidia
So you can learn and prototype GPU-networking on Spark, just not at full GPU-direct performance.
Yes — enjoy DOCA on a DGX Spark as a DOCA-Host + ConnectX-7 + GPU developer box. You get Flow, RDMA, Ethernet, DMA, and GPUNetIO (with CPU-proxy). You do not get the BlueField/DPU half — DPA, on-DPU services, and App Shield/Argus — because Spark has no DPU.
NVIDIA's answer "Is GPUDirect RDMA supported on DGX Spark?" and the DOCA GPUNetIO Installation & Setup (the CPU-proxy notes). For live confirmation, the DGX Spark GPUNetIO forum thread.