@Lauramaywendel
Laura Wendel
7 months
GPT4 was presumably trained for around 90 days using 25k A100 GPUs. Microsoft and Meta having reportedly bought 150k H100 GPUs each this year, can now train a GPT4 class model in only 7 days from scratch (not even including their existing GPU stockpile).
Tweet media one
29
127
577

Replies

@scottemeryy
Scott
7 months
@Lauramaywendel Nividia about to rip again
1
0
2
@Lauramaywendel
Laura Wendel
7 months
@emeryhacks they have a great money printer going atm
0
0
3
@kevinyun
SaaS Kevin
7 months
@Lauramaywendel Awesome news for consumers and developers. Basically the whole world
1
0
5
@Lauramaywendel
Laura Wendel
7 months
@kevinyun depends on how they're used, but yes! 😃
1
0
3
@raj_raj88
Raaz
7 months
@Lauramaywendel Seems Llama 3 won't be on diet.
2
0
4
@Lauramaywendel
Laura Wendel
7 months
@raj_raj88 Can't wait - I'm hearing it will also have vision capabilities! Just wish it wasn't only Meta contributing to open source, seems like all decent open LLMs are Llama derivatives.
0
0
13
@post_relativity
Mr. Architect
7 months
@Lauramaywendel Soon there will be a bottle neck and it’s not GPUs.
0
0
2
@albfresco
albs — 3/staccs
7 months
@Lauramaywendel not included with all shipments: knowing wat you’re doing
0
0
1
@SeanDoesLife
Sean Kaye
6 months
@Lauramaywendel I just want Nvidia to make affordable 4090 cards… why can’t they do that instead of making cards to train AI to reply to tweets or make shitty recipes involving eggplant! LOL.
0
0
2
@dmsimon
dmsimon
6 months
@Lauramaywendel The problem with this graph is it says H100 shipments and then adds below it was A100 as well which is misleading.
0
0
0
@hiraosamasocial
Hira | Conversion Copywriter & Funnel Strategist
6 months
@Lauramaywendel AI is only going to get better & faster. The fact that Microsoft and Meta reportedly acquired such a massive number of H100 GPUs, allowing for the potential to train a GPT-4 model in just a week, speaks volumes about the rapid progression in AI capabilities.
0
0
1
@hermanfaster
Herman Faster
7 months
@Lauramaywendel How many H100 to turn the reality around to get Biden reelected?
0
0
0
@Rufus87078959
Rufus
7 months
0
0
3
@Isaibella_LindB
Isaibella Lindberg
7 months
@Lauramaywendel Wow, so i will get even more 🤩 in the next month or so🥰
0
0
0
@DicksonPau
Dickson Pau
7 months
@Lauramaywendel Doubtful it can be done in 7 days
1
0
1
@henri_nguembi
thesouth
7 months
@Lauramaywendel Is Apple training its models on Apple silicon?
0
0
1
@raywinTX
raywin
7 months
@Lauramaywendel WOW! tiktok has more GPUs than Tesla!
0
0
0
@TomMooreHunts
Tom Moore
7 months
0
0
0
@NietzscheKant2
Nietzsche Copernicus
7 months
@Lauramaywendel A.I. itself may not "destroy the world", but phony political propaganda created with the help of A.I. certainly will.
0
0
1
@x2quark
particulate — e/acc
6 months
@Lauramaywendel Who's the lambda?
0
0
0