The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Nvidia ASHAMED of its current GPU lineup, cancels all reviews
Poast new message in this thread
Date: May 11th, 2025 12:10 PM Author: impertinent carnelian telephone https://youtu.be/4n_J7jclifM?si=Ndp-VEWuRXlOJIe_ (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48921540) |
Date: May 12th, 2025 4:33 PM Author: Vermilion stirring jap the days of NVIDIA caring about the consumer gaming market are over. it's all about AI Data Center for them now, anything else is scraps. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924790) |
 |
Date: May 12th, 2025 4:36 PM Author: impertinent carnelian telephone Google is doing fine without Nvidiashit. Apple too. I wonder how long it will take others to notice. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924793) |
|
Date: May 12th, 2025 4:37 PM Author: Vermilion stirring jap Google is running NVIDIA GB200 NVL72s champ. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924796) |
|
Date: May 12th, 2025 4:39 PM Author: impertinent carnelian telephone I got things mixed up. Apple did it without Nvidia, using Google chips:
https://www.reuters.com/technology/apple-says-it-uses-no-nvidia-gpus-train-its-ai-models-2024-07-29/ |
|
Date: May 12th, 2025 4:43 PM Author: Vermilion stirring jap Apple is behind in "AI" and Google only sells processing units through their GCP cloud platform. Not remarkable - Apple is a hardware company first and foremost. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924808) |
|
Date: May 12th, 2025 4:44 PM Author: impertinent carnelian telephone Apple doesn't seem "behind" in any meaningful way. What opportunities do you think they're missing here? (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924813) |
|
Date: May 12th, 2025 4:45 PM Author: Vermilion stirring jap Apple Intelligence is a Joke https://cybernews.com/tech/iphone-users-disable-apple-intelligence/ (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924815) |
|
Date: May 12th, 2025 4:51 PM Author: impertinent carnelian telephone Oh wow an AI gives inaccurate responses? Shoulda used Nvidia I guess. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924835) |
|
Date: May 12th, 2025 4:47 PM Author: Cerebral chocolate shrine hairy legs It seems that TPUs provide most of their training and inference capacity and that will only become more true as time goes on. Companies like Microsoft going in a similar direction is not a positive thing. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924818) |
|
Date: May 12th, 2025 4:48 PM Author: Vermilion stirring jap ? (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924820) |
|
Date: May 12th, 2025 4:50 PM Author: Cerebral chocolate shrine hairy legs Their nvidia GPUs are primarily for their cloud services. Major AI companies using their own chips cuts Nvidia out of the market as time goes on. They will not retain the market share they have now. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924832) |
|
Date: May 12th, 2025 4:52 PM Author: impertinent carnelian telephone cr, their edge was never in hardware but in developer support. They could dump more money into developer support than anyone else, but it was always only a matter of time until developers started doing their own thing. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924836) |
|
Date: May 12th, 2025 4:58 PM Author: Cerebral chocolate shrine hairy legs Right. The advances in AI coding will also cause this to happen sooner than it would otherwise. Even some of the local LLm market will likely migrate to AMD or Intel GPUs as the software improves (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924854) |
|
Date: May 12th, 2025 5:03 PM Author: Vermilion stirring jap now is the time for Intel to really flood the GPU market if they want to escape Ignominy, I don't know much about their ARC shit but I hear it's ok. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924866) |
|
Date: May 12th, 2025 5:06 PM Author: impertinent carnelian telephone They are working on dedicated AI chips and aren't pushing GPUs for that purpose, although they let you run local LLMs on them now using their own proprietary software. It's not bad. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924877) |
|
Date: May 12th, 2025 5:08 PM Author: Vermilion stirring jap are you talking about NPUs? they already have those in their latest chips afaik. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924880) |
|
Date: May 12th, 2025 5:10 PM Author: impertinent carnelian telephone Yeah they are going that way. No one thinks Intel chips will be used to train models, but running local inference should be easy. I think the cloudshit is going to become less relevant as performance of local LLMs improves. You can already get models that are <1gb that run great. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924885) |
|
Date: May 12th, 2025 6:41 PM Author: Cerebral chocolate shrine hairy legs Yes, and that’s still using traditional models where all the weights are in the GPU. For many specialized tasks, the model likely only needs a small amount of additional information in order to predict a particular token well. The neural networks that play particular games, for example, are quite small and could be quickly transferred from an NVME solid state drive into GPU memory if it was necessary for predicting the next token. I can imagine mixture of expert type models that retrieve from a database of thousands of expert modules based on the current context. No need for hardware with tons of memory next to the tensor cores. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48925085) |
|
Date: May 12th, 2025 6:46 PM Author: impertinent carnelian telephone tyft (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48925095) |
|
Date: May 12th, 2025 6:51 PM Author: Cerebral chocolate shrine hairy legs I am blown away by even current local LLMs. Gemma 3 is great. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48925110) |
|
Date: May 12th, 2025 6:52 PM Author: impertinent carnelian telephone Gemma is the only one I mess with anymore. I might do some coding at some point but I have no urge to use LLMs for that rn. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48925113) |
|
Date: May 12th, 2025 5:03 PM Author: impertinent carnelian telephone The two people I know who work in the biz (one at Meta) run local LLMs on Macbooks with unified RAM. There's nothing you can buy for $5k that gives you better performance than a M4 ultra, or even the M2 ultra with 64-128gb unified RAM (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48924868) |
|
Date: May 12th, 2025 6:53 PM Author: Beady-eyed Swollen Turdskin explainw hy it needs 64 gig ram wtf? there are only 1 billion parameters (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48925119) |
|
Date: May 12th, 2025 6:55 PM Author: impertinent carnelian telephone You don't, 32 is fine. However some people think it's nbd to drop $5k on a Macbook, and the models with high capacity are in short supply so people grab them as they become available. Mac Minis are really the best value right now if you want to run M4 Max or M2 Ultra. You need at least the Max to run LLMs, and if you're getting more than 32gb you really want the Ultra
EDIT oh shit there's a M3 Ultra now. 180 |
|
Date: May 12th, 2025 7:05 PM Author: Vermilion stirring jap I just ordered a Core Ultra 9 285K and a 5070ti. Rate me. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48925151) |
|
Date: May 12th, 2025 7:06 PM Author: impertinent carnelian telephone If you got a z890 chipset you won the gen. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48925155) |
|
Date: May 12th, 2025 7:07 PM Author: Vermilion stirring jap I did but it's one of the mid-range ones, not the crazy ASUS "gamerz" one or whatever
https://pcpartpicker.com/product/xPtLrH/msi-mpg-z890-carbon-wifi-atx-lga1851-motherboard-mpg-z890-carbon-wifi |
|
Date: May 12th, 2025 7:10 PM Author: impertinent carnelian telephone That looks fine. (http://www.autoadmit.com/thread.php?thread_id=5723631&forum_id=2#48925167) |