5 ms·

My guess would be 'artificial scarcity for the purpose of market segmentation', because people probably wouldn't buy that many of the expensive professional car

by mckirk 1y ago

My guess would be 'artificial scarcity for the purpose of market segmentation', because people probably wouldn't buy that many of the expensive professional cards if the consumer cards had a ton of VRAM.

amelius 1y ago
Ok, time to start supporting another brand folks.
cubefox 1y ago
Nvidia indeed does this with the *60 cards, which are limited to 8 GB. They probably copied this upselling strategy from Apple laptops.
- nacs 1y ago
  Except now, Apple with it's shared VRAM/RAM model now has better deals especially past 24GB of VRAM than you get with Nvidia now (for inference at least). A Macbook or Mac Mini with 32GB as a whole system is now cheaper than a 24GB Nvidia card.
  - cubefox 1y ago
    That's an interesting point about unified memory. But I assume most people will use their graphics card for video games rather than machine learning inference. And most non-console games are programmed for Windows and non-unified memory.
YetAnotherNick 1y ago
HBM are in very limited supply and NVidia tries to buy all the stock it could find at any price[1][2]. So the memory literally couldn't be increased. [1]: https://www.nextplatform.com/2024/02/27/he-who-can-pay-top-dollar-for-hbm-memory-controls-ai-training/ https://www.nextplatform.com/2024/02/27/he-who-can-pay-top-d... [2]: https://www.reuters.com/technology/nvidia-clears-samsungs-hbm3-chips-use-china-market-processor-sources-say-2024-07-23/#:~:text=In%20contrast%20to%20Samsung%2C%20SK,will%20supply%20Nvidia%20with%20HBM3E. https://www.reuters.com/technology/nvidia-clears-samsungs-hb...
- WithinReason 1y ago
  how about GDDR?
  - YetAnotherNick 1y ago
    If bandwidth is not the issue, you could directly use system memory via PCIe[1]. No need for on chip memory. [1]: https://developer.nvidia.com/gpudirect https://developer.nvidia.com/gpudirect
    - WithinReason 1y ago
      bandwidth is almost always is an issue, but not enough to be worth buying dedicated HW for it.
karmakaze 1y ago
I'm surprised the RTX cards don't have a Terms of Use that prohibits running CUDA on them. They already removed NVLink from the 40-series onward. Maybe running 8k VR could use the 32GB on the 5090 but I can't imagine much else that's not compute. I'm looking forward to newer APUs with onboard 'discrete' GPUs and quad or more channel LPDDR5X+ and 128GB+ unified memory that costs less than an M3 Ultra.

by mckirk 1y ago

amelius 1y ago

Ok, time to start supporting another brand folks.

cubefox 1y ago

Nvidia indeed does this with the *60 cards, which are limited to 8 GB. They probably copied this upselling strategy from Apple laptops.

nacs 1y ago
Except now, Apple with it's shared VRAM/RAM model now has better deals especially past 24GB of VRAM than you get with Nvidia now (for inference at least). A Macbook or Mac Mini with 32GB as a whole system is now cheaper than a 24GB Nvidia card.
- cubefox 1y ago
  That's an interesting point about unified memory. But I assume most people will use their graphics card for video games rather than machine learning inference. And most non-console games are programmed for Windows and non-unified memory.

YetAnotherNick 1y ago

HBM are in very limited supply and NVidia tries to buy all the stock it could find at any price[1][2]. So the memory literally couldn't be increased. [1]: https://www.nextplatform.com/2024/02/27/he-who-can-pay-top-dollar-for-hbm-memory-controls-ai-training/ https://www.nextplatform.com/2024/02/27/he-who-can-pay-top-d... [2]: https://www.reuters.com/technology/nvidia-clears-samsungs-hbm3-chips-use-china-market-processor-sources-say-2024-07-23/#:~:text=In%20contrast%20to%20Samsung%2C%20SK,will%20supply%20Nvidia%20with%20HBM3E. https://www.reuters.com/technology/nvidia-clears-samsungs-hb...

WithinReason 1y ago
how about GDDR?
- YetAnotherNick 1y ago
  If bandwidth is not the issue, you could directly use system memory via PCIe[1]. No need for on chip memory. [1]: https://developer.nvidia.com/gpudirect https://developer.nvidia.com/gpudirect
  - WithinReason 1y ago
    bandwidth is almost always is an issue, but not enough to be worth buying dedicated HW for it.

karmakaze 1y ago

I'm surprised the RTX cards don't have a Terms of Use that prohibits running CUDA on them. They already removed NVLink from the 40-series onward. Maybe running 8k VR could use the 32GB on the 5090 but I can't imagine much else that's not compute. I'm looking forward to newer APUs with onboard 'discrete' GPUs and quad or more channel LPDDR5X+ and 128GB+ unified memory that costs less than an M3 Ultra.