Vision Pro is the ultimate birth control device.
Young men will be watching porn with a monitor strapped to their faces. Or they will be watching sports when the women are around, neglecting them.
Journalism?
I donât have the habit of checking peopleâs âcredentialsâ. I just judge by the reasons in a personâs arguments.
Agree if the reasons make sense. Disagree if they donât. Use my own brain in the process.
From the profile, can figure out the motivation
ä¸ĺ¨ĺ
śä˝ďźä¸č°ĺ
śćż
Dylan Field @zoink was mentioned by Andrew Wilkinson
Above comment smell of he is recruited as an astrosurfer or at best, a religious belief about what is the correct go2market form factor. From affordable to premium or premium to affordable. From roadster to model 3/y or from model 3/y to roadster. Palmer Luckey, founder of Oculus, has quitted Meta after disagreeing with Mark over the go2market strategy. Mark chose affordable2premium while he prefers premium2affordable.
Recall that Mark has been caught doing astroturfing.
Do we play piano while playing video games on a desktop or watching a movie on a TV?
Letâs get real. You have no logic. You go by whoever supports your feelings. At this point, you trying to claim logic is laughable.
1 Like
Manchâs point was there was no better chip than nVidia for traing and Apple was not buying nVidia chips, hence, he was wondering how they would build/train their own AI model. It is a valid question.
In addition, it is also true that inferencing requires much less computing power, hence easier to build inferencing chips than training chips.
On the other hand, I donât follow your answer. Would you elaborate?
1 Like
.
Is not a valid (just to be nice) question. Take note I use content posted by him only 
15 days ago, I told @manch, he canât imply that Apple didnât buy any H100s and that training must be done by H100s because firstly,
16 days ago, @manch made a conclusion based on the chart below.
Secondly, from Zuckâs tweet, META has other GPUs that are equivalent to H100s, implying can be used for the same purpose i.e. training model.
Those companies arenât buying 15k+ chips for training. Even 7B models can be trained on 128 H100s (1,024 GPUs) with the 8x400Gb networking setup. They are buying so many GPUs to run inference on them. Apple will run inference on devices. Itâs a complete game changer for cost structure. Itâs why Apple doesnât need nearly as many GPUs for AI.
1 Like
H100 is mainly built and optimized for training. Using it for inference is a bit of waste IMHO.
People are building to use them for both. Most people donât have equal usage for inference. They use off peak for training. They are too expensive for most to dedicate them to once use.
1 Like
7B models are mostly toys for amateur enthusiasts.
GPT-4 was trained on about 25K A100s in 3 months. Despite what some companies claimed, the bigger the model the better its performance.
Only 2 companies in the world make high performance GPUâs. Nvidia and AMD.
Most companies use Nvidia because of Nvidiaâs investment in its software ecosystem. All the software tools work with CUDA out of the box, so you can get off the ground running. Meta is one of the small handful of companies that can make AMD chips work. They are the creator of PyTorch after all. I am not surprised they are investing in there.
Apple is already behind in AI. If they are fiddling with AMD drivers instead of focusing on getting their models right I am even more bearish on its prospect.
Mistral already has equal performance at 1/10th the cost. You only need a huge model if you expect it to solve every random request. Thatâs a horribly inefficient way to design it.
.
I have no insider info.
Apparently you donât get the message. âYou canât make those conclusions using those content presented.â IMHO, youâre speculating.
We donât know if itâs true yet. While the GPT models are widely tested, Mistral models arenât.
Mistral said they will release a 70B model later this year that will rival GPT-4 performance. We will see whether thatâs true or not. But note that itâs still 70B, not 7. And OpenAI is moving the ball forward as well. Itâs not like GPT-4 is the end.
This article has good summary on nVidia inference chips. Although h100 is listed in table for inference chips, usuall 100 chips are optimized for traing. I think ad104 is better option for inference than hopper given perf per power and cost. It is optimized for inference.
The L4 GPU accelerator, like its P4 and T4 predecessors, is mean to fit in a 75 watt or smaller power envelope and is meant to handle AI inference as well as graphics and video encoding workloads in the datacenter.
2 Likes