Do we have any speculation on how many GB10 nodes it will take to run Nemotron 3 Ultra effectively? I’m migrating a small dev team to using on prem resources. We have Super running very well for our workloads (supporting devs, not vibe coding) on 2 Sparks. I’m trying to plan for capacity needs and hoping that 3 Sparks will be effective, but I’m very new to this world. Also hoping Ultra is “smart” enough to handle most of what we need, even if it needs some fine tuning to get there.
It is supposed to be 500B parameters, approximately. It will either require 3 or 4 Sparks to run at NVFP4. Since Nemotron 3 Super takes up about 80GB just for the weights, and about 90GB including context, I think a reasonable estimate places Ultra right on the line between 3 and 4. 4 seems like a safe bet.
Yeah, that’s my feeling as well, mostly hoping for 3 to work well so I don’t need to get a switch.