Question - as NVIDIA seems to be lagging and is not stepping up to support this community effort by giving @Eugr access to a Spark (at least one for the nightly builds of the spark-vllm-docker project) should we as a community set up a Gofundme and buy one for him?
I’m based in Europe so I don’t really know how to setup such a campaign, but if someone would set it up I’m up to donate the first €100 or so … :)
Hi eugr, indeed, thanks a lot for your contributions! I have an 8 node with microtik setup working. If you ever want to try stuff, you’re more then welcome…
Frankly, no chance. For a whole host of reasons, this platform needs probably full optimization plus a full year of operating at that level, and then a successor would be viable to be considered.
Beyond that, Jensen was very clear that only one Vera Rubin CPU was being made.
My crystal ball suggests this device class will be on a 2-3 year cycle with the next architecture seen being Feynman.
An update: thanks to NVIDIA, I now have a third Spark and two more QSFP cables, so I can not only have a dedicated Spark for builds, but I can also now see if I can make any optimizations for a 3-node cluster.