Meta's solution to the global memory shortage is to use DDR4 in a DDR5 server, with a custom chip making the impossible possible

https://www.pcgamer.com/hardware/memory/metas-solution-to-the-global-memory-shortage-is-to-use-ddr4-in-a-ddr5-server-with-a-custom-chip-making-the-impossible-possible/

164 points · 27 comments · view on lemmy.world

27 Comments

FrankFrankson@lemmy.world · 145 pts · 47d (11 replies)

Hurray now DDR4 prices can get even more fucked!

tal@lemmy.today · 26 pts · 47d (9 replies)

I shouldn't have thrown out all that DDR3 memory that I did.

historicaldocuments@lemmy.world · 15 pts · 47d (4 replies)

I'm going to drag out my same soapbox: a lot of systems old enough to use DDR3 RAM will have x86_64 v1 or v2 processors. Some projects have already removed support for those, the big one being the RHEL kernel as of RHEL9.

db2@lemmy.world · 8 pts · 47d (2 replies)

I miss the days when mainstream Linux distros could run on a supercomputer or a toaster and anywhere between.😭

historicaldocuments@lemmy.world · 9 pts · 47d

Most still will. Like I'm sure a lot of people are doing, I was trying to reuse old hardware for a new purpose. Perfectly good computer with 16GB of RAM with an AMD A8-3850. I'm not complaining about progress's march towards the future, but I missed the warning signs about the changes. I'm sure some other folks probably did as well.

neinhorn@lemmy.ca · 2 pts · 47d

Back in the day the anywhere between was a lot smaller.

tal@lemmy.today · 2 pts · 47d

Maybe someone will make a way to use DDR3 in a DDR4 system. :-) Make the impossible possible.

taiyang@lemmy.world · 4 pts · 47d

I still have DDR2s somewhere... pray we do not need it.

Appoxo@lemmy.dbzer0.com · 3 pts · 47d

Shouldnt have built a christmas tree with it.

::: spoiler Tap for spoiler It is mostly defective ram. At least most of the modules arent fully bricked as they are just hot glued in place. So technicslly some could still be viable. :::

SharkAttak@kbin.melroy.org · 1 pts · 47d

And where exactly did you throw it?

CosmoNova@lemmy.world · 1 pts · 46d

Even DDR2 got more expensive. We are giga screwed.

boonhet@sopuli.xyz · 1 pts · 46d

Well hopefully just server grade but still not great

SomeRandomNoob@discuss.tchncs.de · 33 pts · 47d

Fuck Meta and the Lizzard Man. Thank you for comming to my TED talk!

PattyMcB@lemmy.world · 28 pts · 47d

When someone said "oh, the ddr4 will be fine," I said it wouldn't. Here we go.

mereo@piefed.ca · 24 pts · 47d (1 reply)

Yay... It will be expensive to build an AM4 system now...

village604@adultswim.fan · 18 pts · 47d

It already is. DDR4 doubled in price.

nbsp@programming.dev · 16 pts · 47d (10 replies)

whats metas actual main usecase for llms? don't think that many ppl are using the chat thing in whatsapp? im sure there are backend systems driving recommendations and ads?

4am@lemmy.zip · 20 pts · 47d

The main use case for LLMs is to dazzle capitalists who now believe they’ve at the precipice of a revolution to eliminate human labor; by doing so the AI proprietors capture the majority of the world’s compute resources for data centers so that society, now reliant on computers and electronic communication for everyday tasks, is forced to rent cloud applications and storage from them forever, where they can monitor consumer and political sentiment, control news narratives and therefore markets, and squash dissent before it can affect their bottom line. And if you dare speak up you can be debanked without a warrant, it’s in your TOS.

Please drink a fucking god damn verification can.

Bustedknuckles@lemmy.world · 19 pts · 47d (3 replies)

They recently said they're going to start selling compute to other LLM companies. This after claiming all the build out was for internal LLM use. Looks to me like you're right and they've given up on devoting their compute to a product

Pechente@feddit.org · 6 pts · 47d (2 replies)

Renting out capacity sounds like a good opportunity for the LLM money circlejerk to make another round

a_non_monotonic_function@lemmy.world · 10 pts · 47d

It would be hilarious if no one bought in.

nbsp@programming.dev · 6 pts · 47d

i believe that's the primary source of revenue for musk's llm interests as well (and has largely given up on homegrown llm).

everyone thinks that battling aws is their out for infinity billions sunk? with zero existing enterprise IT sales... against micrsoft and google as well? 🙃

unitedwithme@lemmy.today · 6 pts · 47d

Well they're building a Polymarket competitor now, bc, god forbid something drew some attention away from Meta products for once, and Zuck can't stand not being the center of everyone's world.

Seriously, they bought IG, had to make Messenger, bought WhatsApp, tried to buy snap, bought Oculus, and according to wiki bought about 100 other companies over the years... Mark, go fuck yourself.

Cocodapuf@lemmy.world · 3 pts · 47d (3 replies)

Well they're developing llama I believe. Which (credit where credit is due), is one of the few open source gpl language models. And you can run it locally, without imposed limits or sending tons of data off to Meta servers.

nbsp@programming.dev · 3 pts · 47d (2 replies)

unfortunately llama is a long way from gpl, and largely abandoned.

you really need to look at chinese models (eg qwen and kimi) if you want osi compatible licence in a model thats actually going anywhere.

https://en.wikipedia.org/wiki/Llama_(language_model)#Licensing

Cocodapuf@lemmy.world · 3 pts · 47d (1 reply)

Yeah, that's fair. I don't really have my finger on the pulse of AI tech, but llama was a thing I remembered.

nbsp@programming.dev · 2 pts · 47d

credit where its due, it really spawned the "ecosystem" which is still reflected in a lot of tooling

eclipse7@feddit.nu · 1 pts · 47d
[ removed ]