[彻夜难眠●不敢相信] Nvidia, Super Micro Computer all 惨重大跌.... whats happening to AI theme?

L

Luqknight

Senior Member
Joined
Jan 23, 2015
Messages
2,361
Reaction score
779
myopenid said:
Bros I’ve held pltr from $20 to $5 and all the way to $80. This is nothing.
Click to expand...
Depends rah. If all in sure die cock stand. If only 5% cash then like losing GST mahney for nation building.

covid i play $plug, amd, gme, pltr and other memes. Now all dead in water
 
L

Lifeinmotion

Senior Member
Joined
Aug 21, 2019
Messages
2,022
Reaction score
1,235
myopenid said:
Bros I’ve held pltr from $20 to $5 and all the way to $80. This is nothing.
Click to expand...
It's nothing if it's a wide market sell off. It's changing the economics of the company it's another thing. Would you hold if pltr competitor has a better product?
 
zeebjii

zeebjii

Master Member
Joined
Mar 27, 2004
Messages
4,250
Reaction score
894
Sylar22 said:
Wah u so zhun, exited almost at the top. Now reenter can earn again. Can share 4 numbers pls ?
Click to expand...
He reminds my of my ex football punter colleague. Always boasts of his winnings. When we pointed out bets that he lost, he always said he double up to take the opposite side during half time, so in the end won again!

So if one day NVDA goes back to say $30, this guy will say yes he bought the dip, but then reverse to buy 100x $30 put options to make a killing instead of losing.
 
8

808180579

Arch-Supremacy Member
Joined
Oct 17, 2015
Messages
13,223
Reaction score
3,244
the prerequisite of fp8 quantization and knowledge distillation is that you must first have a larger model trained in higher precision, which entails requiring larger gpu memory and processing speeds. you then use these larger models to guide training of smaller models by distillation and further reduce their computational costs by quantization.

you can workaround memory requirements by coding your own project to efficiently parrellelize parts of computation that are bottlenecked by memory across multiple smaller gpus instead of a single colossal computing unit if you cannot biy the single colossal one for whatever reason. however not all computationally intensive operations can be parallelized this way, depending on the AI model architecture. nvidia also utilizes high bandwidth memory and direct access between cpu and gpu in their enterprise computing unit, so there is definitely better performance and cost effectiveness versus distributed computing.

modding existing gpu hardware with larger memory chips is also not unheard of but the applicability per card model varies i.e. so far only heard this done on Ampere generation. unofficial mods may introduce additional issues and would also require patching hardware drivers or AI tech stack to resolve compatibility issues.

bearing in mind that these are workarounds, if the majority of users aren’t using such workarounds, the overall ecosystem be it closed sourced or open sourced would not bother maintaining compatibility for such workarounds but develop according to whatever the most cost effective hardware is available to them. the additional burden would be on the smaller group of users to maintain compatibility with no guarantees that similar workarounds may still be possible on future hardware or software.

you may be able to avoid the requirement of training a larger teacher model first for distillation if you can somehow gain access to one, perhaps by chatgpt api. you would have to decouple parts of the training process so you don’t get hard bottlenecked by the api limits during training, so you have to find some way to extract data out of chatgpt that can be used for offline training in parallel across hundreds of gpu.

for the layman, what this means is at the forefront of AI, hardware performance is still a deciding factor. if you do not have a concrete idea of exactly where to go, you cannot take an optimized shortcut to that destination. optimization should be done last. however, there can’t be too many market leaders as the cost is enormous so usually is one leader at the top like TSMC, followed by many smaller market players that do not need that EUV or DUV for their product or services.

not everyone sees this or agree with it but this AI revolution isnt just propped up by consumer or industrial demand and dictated by market forces solely. in some ways the AI race between global superpowers mimics the space race during the cold war or the nuclear arms race even further back. in this case, the funding and demand for performant hardware then isnt solely driven by consumers or enterprises. for NVDA of course this also brings into the possibility of competitors borne out of necessity even in America, though at this point in time NVIDIA still holds a very large moat in their software ecosystem.
 
M

Mechafanboy

Arch-Supremacy Member
Joined
Jan 31, 2014
Messages
12,307
Reaction score
9,226
The irony is, wall Street likes its monopolies and price gouging, which can be antithesis to the principles of efficiency.
 
charsiewfan

charsiewfan

Arch-Supremacy Member
Joined
Sep 7, 2010
Messages
14,633
Reaction score
4,996
I think it’s just an excuse to correct the market so that the players can come in and get discount
 
dereth

dereth

Supremacy Member
Joined
Jan 1, 2000
Messages
18,712
Reaction score
5,163
- S$888K... 發發發

REALLY HUAT HUAT HUAT AH... 🥲

UyvKUcR.jpeg
 
LocalRider

LocalRider

Master Member
Joined
Aug 7, 2022
Messages
3,973
Reaction score
2,408
Last edited:
You must log in or register to reply here.
Important Forum Advisory Note
This forum is moderated by volunteer moderators who will react only to members' feedback on posts. Moderators are not employees or representatives of HWZ. Forum members and moderators are responsible for their own posts.

Please refer to our Community Guidelines and Standards, Terms of Service and Member T&Cs for more information.
Top