Hacker News Reader

Yep that's what I'm thinking as well. I was going to buy a 5090 mainly to play around with LLM code generation, but this is a worthy option for roughly the same price as building a new PC with a 5090.

View on HN

It has 128 GB of unified RAM. It will not be as fast as the 32 GB VRAM of the 5090, but what gamer cards have always lacked was memory.

Plus you have fast interconnects, if you want to stack them.

I was somewhat attracted by the Jetson AGX Orin with 64 GB RAM, but this one is a no-brainer for me, as long as idle power is reasonable.

Having your main pc as an LLM rig also really sucks for multitasking, since if you want to keep a model loaded to use it when needed, it means you have zero resources left to do anything else. GPU memory maxed out, most of the RAM used. Having a dedicated machine even if it's slower is a lot more practical imo, since you can actually do other things while it generates instead of having to sit there and wait, not being able to do anything else.

Reply on news.ycombinator.com