3 ms·

I buy cards primarily on: - Machine learning / GPGPU support - Stability - RAM A 16GB card for $349 is a good deal. That's more than a 3080, and same as a h

by blagie 4y ago

I buy cards primarily on: - Machine learning / GPGPU support - Stability - RAM A 16GB card for $349 is a good deal. That's more than a 3080, and same as a higher-end 4080. If Intel is ever able to ship stable open-source drivers supporting pytorch, tensorflow, numpy, etc., this will be a game changer. What I really hope Intel or AMD decide to ship is a card designed to disrupt the higher-end NVidia products here: something with e.g. 64GB RAM for $500-$1000. That would bring voice recognition, image generation, speech synthesis, etc. to a much broader audience. Or better yet, some hybrid model which can make use of system RAM as well, so the memory is upgradeable. Yes, I know all the points about performance, things being IO-bound, and whatnot. I'll take it. For gaming, they could also include a smaller amount of very fast RAM.

p1esk 4y ago
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.
- neogodless 4y ago
  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.
  - p1esk 4y ago
    Oh I see. My guess is it will be competing with 4060 cards as far as gamers are concerned.
    - blagie 4y ago
      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.
smeagull 4y ago
Yup. Made a massive mistake getting an AMD graphics card last time around. Has never worked for any ML workload ever.

by blagie 4y ago

p1esk 4y ago

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.

neogodless 4y ago
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.
- p1esk 4y ago
  Oh I see. My guess is it will be competing with 4060 cards as far as gamers are concerned.
  - blagie 4y ago
    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.

smeagull 4y ago

Yup. Made a massive mistake getting an AMD graphics card last time around. Has never worked for any ML workload ever.