I’d like to share a feature request and design consideration for future DGX Spark / workstation-class platforms, focused on desk-side multi-node connectivity.
DGX Spark is clearly positioned as a developer-centric, desk-side AI system.
For users experimenting with 2–4 systems locally, the current scaling option relies on ConnectX Ethernet/RoCE, which brings data-center-style complexity:
External switches
DAC/optical cabling
Rack-oriented networking assumptions
While this is ideal for larger deployments, it creates friction for small, local, exploratory multi-node workflows.
Rather than traditional Ethernet or full data-center fabrics, I’d like to suggest consideration of a short-range, NVIDIA-managed interconnect with characteristics similar to modern workstation I/O fabrics:
This would complement, not replace, ConnectX/RoCE — specifically targeting desk-side experimentation and development.
2–4 DGX Spark systems on a desk or lab bench
Distributed inference and model sharding
Multi-node fine-tuning before scaling to DGX H100 or cloud
Research labs and startups without data-center networking
Why This Matters
Strong alignment with DGX Spark’s developer-first positioning
Lower barrier to entry for multi-node experimentation
Encourages earlier adoption of NCCL / distributed workflows
Clear differentiation from both traditional servers and consumer workstations
Many users don’t need full rack-scale networking — they need a simple, high-bandwidth, short-distance interconnect that feels natural at the workstation level.
Questions for the Team
Is NVIDIA exploring short-range, non-Ethernet interconnects for workstation or DGX Spark-class systems?
From a systems perspective, what are the main blockers today (DMA isolation, security, OS support, protocol semantics)?
Is this class of interconnect something NVIDIA sees as valuable for future desk-side AI platforms?
Thank you for the continued innovation around DGX Spark.
I appreciate any insight into whether this direction is under consideration or being evaluated internally.
Appreciate the feedback.
One point to keep in mind is that ConnectX is compatible with standard Ethernet. That aligns with a core goal of Spark: a developer-focused platform to explore and experiment using datacenter-class infrastructure, with the expectation that what you build on Spark can be ported to full datacenter environments with minimal friction.
For 2 nodes, the direct 200 GbE interconnect would work fine. You get faster than Thunderbolt class, direct system to system connectivity and NCCL compatibility which pushes GPU to GPU ultra fast.
What I think would have potential value is a desktop 200 GbE x 4 switch. You could scale from 2 to 4 Sparks, and the lessons/principles are more translatable to the enterprise level clusters, in a way that an “extra” desktop interface isn’t which requires a whole new level of drivers and maintenance.
The challenge with 2 nodes only is that I for sure have 2 nodes running well, but I have no way to test 3 or 4 nodes trivially. I could run 10 GbE but it doesn’t let me predict the scaling of my tasks and evaluate different topologies running infiniband verbs and NCCL over Ethernet.
With Microtik offering some really competitive 200/400 rack switches, it seems like a 4 port 200 GbE desktop switch would be a meaningful feature that doesn’t disrupt the enterprise products.
Yeah, that’s the style of switch I am thinking of. The Microtik is only 100GbE though. I am looking for the full 200 GbE performance. We would be leaving a lot of performance on the table with just 100 GbE.
A 4-port (or 8-port) 200GbE-capable desktop switch would be such a niche product that in order to justify production, pricing would be extremely high.
I’d wager the chances of such a thing coming to fruition would be exceedingly small.
For performance on par with the direct interconnects, you’re looking at enterprise-class switches, which won’t be exactly suitable for a desktop, and (depending on your environment) may or may not be out of your budget unless you source them used from an auction site.
Yeah, it’s just a wishlist item. I was surprised that we even got 200 GbE on the Spark in the first place. The cost is less important than the noise. Maybe the right move is a SFF PC with a pair of dual port 200GbE NICs and setting up software switching.