Anyone pull the plug on DGX Station GB300?

Posted by segmond@reddit | LocalLLaMA | View on Reddit | 8 comments

I want to see what the damn numbers look like!@#$@ pp & tg

I know we got a few big pocket players in here.

[-]

DataGOGO@reddit

I doubt anyone is buying these to just run local inference.

[-]

E8@reddit

I think the best we got is fairydreaming's GH200 experiments.

W a lil imagination w wattages and bandwidth specs, I think it's a bump faster. Goog's summ sugg gb300 is 1.5x faster than gh200 (q="gh200 vs gb300"). So scale accordingly. Maybe sub 10% to acct for nv reality distortion field (aka bs).

[-]

segmond@reddit (OP)

Not at all. I'm getting more than half that performance with 3090 > 8tk/sec while offloading more than half to system ram. So I'm 100% certain full GPU load with 3090s will crush that experiment. Going full GPU often leads to me seeing 3x performance so if I should expect 3090s to give about 24 tk/sec at the very least.

I used a single B300 to approximate a GB300 to a single B300, and based on my comparison to existing GPUs, I'm guessing a 4.5-6.5x boost. So that DeepSeek should run at 100-160tk/sec at the very least. But the optimist in me expects and hopes for me.

[-]

megadonkeyx@reddit

more than 80k bananas, cargo pants players.

[-]

segmond@reddit (OP)

I got quotes from a few vendors, cheapest quote I got was $89k. One was almost $120k. But plenty of them own businesses and it will be a line item in business expense ..

[-]