Is there a way to pair a DGX Spark with an RTX 6000 PRO Blackwell installed on my desktop?

[s0ne]

This might be a stupid question, but I’m asking just in case. A colleague of mine is also curious about this.

I have an RTX 6000 PRO Blackwell MAX-Q Workstation on my Z790 Intel desktop. I’m wondering if it’s possible to install a ConnectX-7 card in this system and connect it to a DGX Spark to expand its compute power. I don’t think it’s possible, but I think NVIDIA officials would be the best ones to know, so I’m leaving this question here.

[elsaco]

According to this post How many DGX Sparks can be interconnected using ConnectX? Ethernet is the underlying protocol, so it looks doable!

[s0ne]

I asked a question because I was wondering if ConnectX-7 would normally recognize the RTX 6000 PRO Blackwell if I installed a ConnectX-7 card in the remaining PCI Express 5.0 x16 slot.

[eggman]

You can do it with WiFi, much less Ethernet when you have two computers.

ConnectX-7 gives you really fast bandwidth.

[NVES]

RoCE and NCCL work across x86 and arm64. Both protocols are NIC/driver config dependent, not CPU-ISA dependent. If your workload is MPI launched, then mixed arch is fine as long as each node uses its native binaries, avoid features that assume homogenious nodes.

[s0ne]

thanks. I’ll try once I receive DGX spark this week.

[alan.dang]

@NVES I made the post in the other thread, but to confirm, I could take a desktop PC running Windows and WSL Ubuntu with a RTX Pro 6000 Blackwell, and then put in:

MCX753436MC-HEAB

And then connect the PC to the Spark via

MCP1650-H001E30

Those are the correct part numbers? I assume the encryption ready CX7 works fine with the Spark, even if encryption is disabled?

Edit: and is the CUDA Compute Capability of 12.0 vs 12.1 vs 12.2 of any concern?

[elsaco]

The script at dgx-spark-playbooks/nvidia/connect-two-sparks/assets/discover-sparks at main · NVIDIA/dgx-spark-playbooks · GitHub can be run on any system with a smart NIC.