It does work, but it’s not really fast. I upgraded to 96gb ddr4 from 32gb a year or so ago, and being able to play with the bigger models was fun, but it’s not something I could do anything productive with it was so slow.
You can have applications where wall clock tine time is not all that critical but large model size is valuable, or where a model is very sparse, so does little computation relative to the size of the model, but for the major applications, like today’s generative AI chatbots, I think that that’s correct.
It does work, but it’s not really fast. I upgraded to 96gb ddr4 from 32gb a year or so ago, and being able to play with the bigger models was fun, but it’s not something I could do anything productive with it was so slow.
Your bottle necked by memory bandwidth
You need ddr5 with lots of memory channels for it to he useful
I always thought using ddr5 average speeds with like 64gb in sticks on consumer boards is passable. Not great, but passable.
You can have applications where wall clock tine time is not all that critical but large model size is valuable, or where a model is very sparse, so does little computation relative to the size of the model, but for the major applications, like today’s generative AI chatbots, I think that that’s correct.
Ya, that’s fair. If I was doing something I didn’t care about time on, it did work. And we weren’t talking hours, it it could be many minutes though.