sirzerp
745
12
7
I know how to do it. We need to request Local AI Hardware with Wire Speed ROMs.
What are Wire Speed ROMs, They are cheaper Read Only Memory chips that mimic expenisve RAM.
Here's the bonus, chip fabs not making cutting edge RAM can make Wire Speed ROMs.
That spreads out the manufacturing to other companies; and helps us break this thirty six month RAM log jam!
What is special about Wire Speed ROM? Nothing special other than allowing increased size of Local LLM Models.
They let your GPU run models bigger than its own memory — 2x for dense models, 10× for MoE.”
Think of it as Local LLM democracy, allowing users to run big models at home.
Of course if you can run big models at home, then why do we need new data centers?
What a shame they borrowed all that money over prime, and something as simple as a pacman cartridge can change the whole damn business model.
For more https://www.pic2mag.com/Frozen_Weights_Fast_ROM_Master_Whitepaper.pdf
lightfoot2
And static and fixed LLMs. I mean maybe but meh. A good 8gb Orin seems to do well for what I need (random bullshit on demand)
kisselFL
Use "write only memory" instead.
raitchison
The 5200 port of Pac Man was ass.
BastardMan1977
So was the 2600 port
Marsupialmessiah
Llm democracy? What the fork you talking about? These data centers and all the computing power aint for shit art, you can do that with silicon graphix indigo 2 fron 30 years ago its for mass surveillance, to correlate statistically what you do, what you will do and where you are bro, the ai slop plagiarizing option is to appease the masses goving them a slice of bread and circus, without the bread cuz it used all the water and the electricity