Anyone try 5090 yet
Posted by 4hometnumberonefan@reddit | LocalLLaMA | View on Reddit | 13 comments
Is the 50s series fast? Looking for people who have the numbers. I might rent and try some if interested. Shoot some tests and what models to try below.
goolissfun@reddit
QwQ32B AWQ，60TK/s
serious_minor@reddit
I have a setup with a founders card and one with a 5000 ada. My experience with the founders card when I first got it was there weren't drivers
IntrovertedFL@reddit
https://www.hardware-corner.net/dual-rtx-5090-vs-h100-for-llm/
codingworkflow@reddit
Not fully true!!! Model need to fit in Vram. H100 will be faster then by far.
Low-Opening25@reddit
I am not sure why this is a surprise. Interface is basically down to VRAM bandwidth, so newer card with faster VRAM wins. what H100 excels at is compute intense training and fine tuning performance.
Low-Opening25@reddit
since bottleneck is VRAM bandwidth, RTX50xx wont really bring any performance gain that would justify the cost vs older gens with the same amount of VRAM. considering they are expensive there is little appeal in buying them for rig dedicated solely for AI, unless of course you are also going to use it for gaming, then sure it will work great for both.
330d@reddit
32GB 1.8TB/s vs previous gen consumer flagman's 24GB 1TB/s, what same amount of VRAM are you talking about? It's a huge uplift for those who can afford it, tears through models up to 32B with more context or better quants.
Low-Opening25@reddit
24GB variant of 5090 is 1GB/s, so basically the same as 3090 and 4090 with the same amount of VRAM. sure, 5090 may be worth it for for the 36GB and bigger variants, but why not buy multiple 3090/4090 instead for even more VRAM? esp. considering gamers will be dropping them on secondary market in the coming years swapping for 5090s.
Escroto_de_morsa@reddit
Are you sure what you're saying? is not 3090/4090 a 384 bit bus while 5090 is 512?
Bandit-level-200@reddit
Only issue its such a pain to setup...I got mine today like all ai stuff I use is broken cause it only supports 12.8 cuda. so its a bunch of wacky solutions to get it to work because they don't update. I suppose LM studio works out of the box, text gen ui which I use seems to be dead, forge? dead. Comfyui? a seperate version that was packaged so thats good but you had to search for it...
Such a pain in the ass to get it all to work
LA_rent_Aficionado@reddit
They’re brand new cards and this is free open source software with tens of thousands of lines of code… give it time
Bandit-level-200@reddit
Well nvidia could've made them backwards compatible with at least cuda 12.6 or something
Low-Opening25@reddit
3090 with 24GB is 900GB/s, 4090 is 1TB/s. 5090 may be worth it for for the 36GB and bigger variants, but why not buy multiple 3090/4090 esp. considering gamers will be dropping them on secondary market in the coming years.