50 million nodes is insane
2025-07-24 21:42:02 +01:00 by Mark Smith
There is a pretty eye popping table in this Tom‘s Hardware article that shows the Nvidia enterprise GPU roadmap. It crazily parabolic.
From a not too shabby 3 TB/s memory bandwidth to a truly ridiculous 32 TB/s, in only 3 years. For context, consumer memory bandwidth for DDR5 is still around 35 GB/s! Some context from a different angle, a fully rendered full feature 4K resolution Hollywood movie at a VFX studio that artists work on is around 10-14 TB in size. Now remember what they have been saying recently, that these generative AIs will eventually be able to generate entire movies in an instant, and you start to realise the significance.
Oh yeah and Elon is saying he will have quite a big collection of these. Last week Sam Altman said they were on track to having 2 million of them and now Elon says he'll have 50 milion.
And then there is this: "50 million of these processors will consume 35 gigawatts (GW), which is equal to the typical power generated by 35 nuclear power plants". When I started out in the VFX industry in the mid 2000s, our render farm had around 250 nodes, that was pretty big at the time.
50 million nodes. This number is just so big. I am having trouble believing this is possible. #