#Triple 3090 build, aka the spaghetti
1 messages · Page 1 of 1 (latest)
For anyone wondering, it's for AI work.
I do plan to case it one day, but for now, atleast it gets alot of airflow.
The 3rd card is on a pcie riser.
The CPU is a 7800
64gb of ddr5
4tb's of storage
so its server rack time for you then

Server boards and CPU's that support DDR5 are so expensive man... 😢
you literally have triple 3090s, thats some broke ass excuses from you

They are all from ebay, it's only like $2400
Whole machine is under $3200
If I was to buy a ddr5 compatible server/threadripper board and cpu it would add $1500
(even from ebay)
about $2000-2500 if it was brand new
and whose fault is that for going into it trying to fit 3 3090s in an actual professional build 
as said you pay for HEDT
Wouldn't call it professional per se... 
you can always go consumer if you cant afford HEDT after all but you pay for the extra performance/features of HEDT
and even if it looks jank at least the current setup works
Main thing I'd want from HEDT/server is the pcie lanes.
I've got a few routes
Get a better consumer grade motherboard that can do 16x, 4x, 4x
Get a super expensive server board with ddr5 support and can do 16x, 16x,16x
Build a seperate rig with the 3090's, and use an older server board that can do 16x, 16x,16x
I don't really want to build a separate rig, as it feels a bit redundant to not use a machine with three of the third best gaming card... for gaming
technically 5th best if going by pure gaming perf
yes, 4070ti aswell technically too
Damn... too bad nvidia loves to save the $40 bucks on the extra vram
It's also too bad, that RoCM is so far behind nvcc and tensorRT
because I would love to use the cheaper amd cards for this stuff
but yeah 4 nvidia ones are faster in gaming
also the new super cards coming out soon
It's not really about the gaming for me, CUDA, Nvcc, tensorRT
Are all research/development tools.
It's just a nice bonus that the 3090 is good at gaming
If AMD had the same tools, I'd probably be using it, or whoever was cheapest
I'm no fanboy
yeah for those tasks its 4090/4080/3090ti are the only really better ones by far in comparison
4090 especially
4090 is a fucking monster
4090 is better for research tasks, or gaming?
much larger diff between 4090 and 4080 then previous 3080 vs 3090
RTX 4090 vs RTX 3090 benchmarks to assess deep learning training performance, including training throughput/$, throughput/watt, and multi-GPU scaling.
because for AI inference, it's a waste of money.
Despite the third 3090 running at 50%, and one a pcie x1 lane. It only slows down inference by 3%
Training and inference are different, but yeah the compute speed would definitely help for training
yep you pay for the best though
thats for sure
and a lot of companies space used up is worth more then cash spent on cards
Still a waste of money though lol, if you need compute, get an a5000, or a6000, hell even a a100.
Fuck it spend 40k on a H100 and get the speed of 20 4090's in one card!
3090's and 4090's are good for their vram to price ratio, not their compute
yep but for almost every normal person that cant afford it for ai stuff it makes no sense
big businesses it makes sense to go for the bigger stuff like that ofc
if you can manage to get into the nvidia inception program you can buy GPU's for their manfacturing cost
get a 4090 for $300
A6000 for $400
Shit's ridiculous
What the fuck, seriously???
Yeah