Hacker News new | ask | show | jobs
by nottorp 506 days ago
It's targeted at the "AI" gold rushers isn't it? Not at gamers.
5 comments

It's it a good rush to want to run open models on your own computer? This isn't like Bitcoin mining where there is some direct monetary reward for hoarding compute.
Either gold rush or fear of missing out if you ask me.

I've succesfully run (not trained) local models * on my mac mini that cost less than a single video card anyway.

* That fit in my ram. They were probably slower than the FOMO hardware but good enough.

it is a gold rush from nvidia's perspective in that they're selling shovels
I got a 2080Ti and was looking to upgrade, and I also enjoy running local AI models. Given that the card only has 16GB memory I don't see it as a huge upgrade over my 2080Ti, I can only get marginally larger models in memory on it. And if the model is in memory then the 2080Ti is fast enough.
Not really, given that it doesn't increase the amount of RAM compared to the old 4080 Super. If you want to do 'modern' AI on a (relative) budget you should be looking at a 4090 or 5090. This seems to be the card targeted most squarely at gamers.
I heard nvidia is gimping consumer-grade cards to not be good at LLM training, is this true? If so are they gimped only for training or also for running LLMs?

I guess the limited amount of RAM is also a way to limit the cards.

Many Nvidia "gaming" SKUs are already at the point where memory is often the biggest likely limitation on their gaming use case, and they'll be noticeably better products for the consumer with a small cost increase by adding more memory.

So I'd say there's good evidence that something outside cost and value to the gaming use case is why they don't have higher memory SKUs, and eating into "professional" priced AI SKUs is an obvious possibility.

I doubt anyone outside Nvidia itself knows "for sure", but it's a pretty big indication.

At least Mistral 7B for its 128 token text generation is 58% faster with 5090 compared to 4090. https://www.phoronix.com/review/nvidia-rtx5090-llama-cpp/3
Nvidia's Digits is a better deal: 128GB/$3000.
There are plenty of deranged gamers spending 5-10k every few years don't worry