Right now it seems we are once again on the cusp of another round of LLM size upgrades. It appears to me that having 24gb VRAM gets you access to a lot of really great models, but 48gb VRAM really opens the door towards the impressive 70B models and allows you to nicely run the 30B models. However, im seeing more and more 100B+ models being created that push the 48 gb VRAM specs down into lower quants if they are able to run the model at all.
this is in my opinion is big, because 48gb is currently the magic number for in my opinion consumer level cards, 2x 3090’s or 2x 4090s. adding an extra 24gb to a build via consumer GPUs turns into a monumental task due to either space in the tower or capabilities of the hardware AND it would put you at 72gb VRAM putting you at the very edge of the recommended VRAM for the 120GB 4KM models.
I genuinely don’t know what i am talking about and i am just rambling, because i am trying to wrap my head around HOW to upgrade my vram to load the larger models without buying a massively overpriced workstation card. should i stuff 4 3090’s into a large tower? settle up 3 4090’s in a rig?
how can the average hobbyist make the jump from 48gb to 72gb+?
is taking the wait and see approach towards nvidia dropping new scalper priced high VRAM cards feasible? Hope and pray for some kind of technical magic that drops the required VRAM while simultaneously keeping quality?
the reason i am stressing about this and asking for advice is because the quality difference between smaller models and 70B models is astronomical. and the difference between the 70B models and the 100+B models is a HUGE jump too. from my testing it seems that the 100B+ models really turn the “humanization” of the LLM up to the next level, leaving the 70B models to sound like…well… AI.
I am very curious to see where this gets to by the end of 2024, but for sure… i won’t be seeing it on a 48gb VRAM set up.
Your wait and see approach is probably wise. The newly released GH200 chip leapfrogs the H100 by a considerable margin, which was already smoking the A100.
On the consumer side, there does not seem to be a high demand to run local LLM. However, I used a 7b model with GPT4All on my ultrabook from 2014 which has a low-tier intel 6th gen with 16gb ram and was getting about 2.5 tokens/second. It was super slow but just shows what would be possible with some optimizations on consumer hardware.
If you’re willing to spend $10k to run an esoteric 110b model, it might be worthwhile to go for the capability to train them in the first place (even if perhaps very slowly). Or, consider a mac with large amounts of memory that’s built into the soc (unified memory) which would likely run models at an acceptable rate with some optimizations. Of course, if blistering performance isn’t necessary.
Otherwise, patience will likely have some good results in the context of a solid model which works on consumer-grade components. The space seems keen on allowing general users and enabling alternatives to transmitting data to some random server elsewhere. Opinion.