News for hackers. Stuff that matters.
Re:Intel A770, A750 review: We are this close to recommendin... by p1esk on Wednesday October 05, 2022 @11:14AM (#33096605)
A 16GB card for $349 is a good deal. That's more than a 3080, and same as a higher-end 4080.

What do you mean? 16GB 4080 is $899.

Re:Intel A770, A750 review: We are this close to recommendin... by neogodless on Wednesday October 05, 2022 @11:15AM (#33096622)
The parent comment is laser-focused on the VRAM capacity of the cards. They are saying 16GB is more VRAM than you find on the 10/12 GB RTX 3080, and the same as the 16GB RTX 4080.
Re:Intel A770, A750 review: We are this close to recommendin... by p1esk on Wednesday October 05, 2022 @11:21AM (#33096717)
Oh I see. My guess is it will be competing with 4060 cards as far as gamers are concerned.
Re:Intel A770, A750 review: We are this close to recommendin... by blagie on Wednesday October 05, 2022 @11:58AM (#33097213)
At this point, machine learning has seeped into a lot of my workflows. What I can do and how well is capped by video memory. This is useful enough that it's just a matter of time before this makes it into mainstream tools. There are GPGPU things other than machine learning which will likely follow.

Right now, I have a card with 16GB, which cost a bit over a grand. This means it can do Stable Diffusion at 512x512, but not higher resolutions. It can do some NLP models, but not GPT-3 scale ones. It can do awesome speech recognition.

These things are useful enough that I claim more people will want to use them as soon as they don't require knowing Python. I think it's just a matter of time before enterprising folks will write code to make some of these workflows available to everyone. I don't think it's unlikely that GPGPU will be standard at least in business laptops in the near future, for the productivity increase.

If Intel can do this for 1/2-1/3 the cost of NVidia, the impact will be huge. If Intel can democratize truly large models, like GPT-3, the impact will be even greater.

If a model takes 20 seconds instead of 10, that's not a huge difference. If it runs out of VRAM, it is.

business market >>> data center market > gamer market

This kind of stuff was Intel's sweet spot maybe two decades ago. They shipped high-quality numerical code, stable drivers, a lot of open-source, and robust integration with operating systems.