[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 61A0MVnu9AL.jpg (82 KB, 1024x1024)
82 KB JPG
lower your tone if you don't have 1tb ram
>>
>>109358732
use case?
>>
>>109358763
three chrome tabs
>>
>>109358763
Three virtual machines running After Effects
>>
File: ram.png (471 KB, 438x423)
471 KB PNG
>>109358732
lower your tone if you need over 8gb ram
>>
>>109358814
I thought 16 was the minimum in [insert_current_year]
>>
>>109358763
Firefox
>>
>>109358763
Gaming. Simple as.
>>
>>109358763
LLMs
>>
>>109358763
compiling some homosexual code in rust
>>
>>109358814
Dual channel 8 is probably comfy but still living on the edge.
>>109358834
Hopefully a few years because that's what I'm stuck with.
>>
File: I am fat fuck.jpg (115 KB, 800x619)
115 KB JPG
>>109358732
>8x128
Come back to me when you have 4x256
>>
>>109358996
just how slow do LLMs run on standard DDR5 like this as opposed to faster VRAM?
>>
File: 1776520666155276.jpg (73 KB, 967x1024)
73 KB JPG
>>109358763
so my fizz buzz programs run longer before they run out of ram.
>>
>>109361726
>his cpu only supports quad channel memory
point and laugh
>>
>>109358763
worse programming
>>
>>109358763
Windows 11
>>
>>109361745
My rule of thumb for LLMs is they need to read twice the size of the model in memory per token. So multiply the model size (after quantization) by two and divide by your bandwitdh to get a rough estimate of token/s.
My Ryzen Threadripper 7970X with 4 channels of 6400 MT/s will do 200 GB/s.
My RTX 4080 Super will do 736 GB/s.
So it's a pretty big difference, but of course a consumer GPU doesn't have 192 GB of memory to play around with. And another way to look at it is that you can bleed into the RAM with less of a performance penalty if you have 4 or 8 channels of high speed RAM, compared to a 2-channel consumer CPU/motherboard.
>>
>>109364362
>twice the size of the model in memory per token
"in memory ACCESS per token", you don't need 140 GB of RAM to run a 70 GB model, but you will read 140 GB of RAM in the process of running the model for one iteration.
>>
they make 128gb ddr5 dimms now? when I got my workstation in 2024 64gb was about how far you could go, maybe 96, and that's at 4800 (rdimm ofc)
im such a poorfag for only having 256gb ddr5-4800 registered... (worth 10 grand now when I bought it for like 1)
>>
>>109358763
Running a dual react native app
>>
>>109364362
Miss me with the fancy talk I just want to know if I can fuck it.
>>
>>109367041
>DDR5
>poorfag
nigga i still run DDR4 and DDR3, you call DDR5 being a poorfag??
>>
I might have 1tb across all my devices. Does that count?



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.