[ home / overboard ] [ soy / qa / raid / r ] [ int / pol ] [ a / an / asp / biz / mtv / r9k / tech / v / sude / x ] [ q / news / chive / rules / pass / bans / status ] [ wiki / booru / irc ]

A banner for soyjak.party

/tech/ - Soyence and Technology

Download more RAM for your Mac here
Catalog
Email
Subject
Comment
File
Password (For file deletion.)

File: Icelandic-ram-on-green-pas….jpg πŸ“₯︎ (583.2 KB, 1920x1275) ImgOps

 β„–35711[Quote]

Opinions on the memory shortage and what do you guys think the solution might be?

 β„–35713[Quote]

the solution is more efficient software, and the shortage will last until early 2030s at best due to STA's

 β„–35731[Quote]

>>35713
early 2030s isn't that bad

 β„–35761[Quote]

A collapse in AI demand will create the incentive to convert fabs back to DRAM, but that will take at least 2 years.

 β„–35768[Quote]

>implying there's an end to this
haha. very funny.

 β„–35770[Quote]

>>35768
LLMs do have scaling limits though. This illusion of unbounded growth will be maintained at all costs because it is the only thing causing the economy to grow right now. Maybe it will stop?

I think that it has found so much success in the corporate world because managers want to replace everyone that is not a manager with a robot to hire more managers to they can continue to play the game.

 β„–35783[Quote]

>>35711 (OP)
More efficient software (geg like nudevs are ever going to let that happen) and tiered swap devices. Think M.2 SSDs like a modern Intel optane that are <= 64GB for consumers with insane random read and write and are possibly even volatile and then put them between your regular swap device

 β„–35784[Quote]

>>35783
This and you will see systems with unified memory become more common. At the very least, you will see more kernel advancements in the direction of VRAM as swap and RAM as VRAM as has already been going on

 β„–35865[Quote]

>>>35783
VRAM as RAM is actually trivial in it’s just that NVIDIA jews block direct CPU access at the firmware level on consumer cards. You can get around it with CUDA hacks (see nbd-vram), but it’s slower and janky. On server cards, VRAM gets pinned and mapped directly into the CPU’s address space via BAR1 by the kernel

 β„–35866[Quote]

>>35865
fail reppie

 β„–35867[Quote]

put jews in work camps and force them to make new ram

 β„–35884[Quote]

>>35761
AI demand will probably not "collapse", fizzle out in the best case

 β„–35898[Quote]

Put memory in magic cubes

 β„–35901[Quote]

>>35770
You are thinking of the wrong scaling limits. LLMs cannot scale infinitely, but what about corporate greed and stupidity? I think that will scale much further

 β„–36075[Quote]

its not ai (((they))) are hoarding vital things for agenda 2030

 β„–36102[Quote]

You WILL own nothing und be HAPPY

 β„–36106[Quote]

solution is China stealing tech to make modern hardware from greedy kikes and building thrembillion factories to supply the goyim (us)



[Return][Catalog][Go to top][Post a Reply]
Delete Post [ ]
[ home / overboard ] [ soy / qa / raid / r ] [ int / pol ] [ a / an / asp / biz / mtv / r9k / tech / v / sude / x ] [ q / news / chive / rules / pass / bans / status ] [ wiki / booru / irc ]