I've heard a rumor the main reason for Nvidia's reluctance to open source their driver is that they use it to enable/disable features. So the same board can be GeForce or Quadro, depending on what the driver tells it to be. Is this true?
AFAIK there’re only 4 differences between GeForce and Tesla/Grid: more and faster VRAM, more FP64 GFlops on most GPUs, unlimited count of nvenc encoding streams versus 2 on GeForce, and support for virtualization. Only the first two are hardware differences, the other 2 are software.
Another thing about virtualization, it sounds ridiculous but nVidia charges non-trivial per-user fee to allow people to use the hardware they already own. Specifically, for virtual workstation they want $450 per concurrent user. That’s on top of the $3000-$12000 purchase price of these GPUs.
IBM does that with mainframes, but, at least, it’s hardware they own that they shipped in the box for you to rent out when needed or to buy and enable.
I feel a lot of this comes down to them using closed source software that they either don't have the code to, or aren't privileged to distribute. Libraries are a common spot for this to appear
Of course, I'd hate that in normal times, but if they now made me an offer to "upgrade" my 1060 (to moderately higher performance, but still lower end), I'd probably take it.
Also I'd be skeptical to some extend on this. Maybe back in time when a chip fail ratio was lower was a case. But now, there is a quite measurable amount of chips not passing requirements for a high tier cards, but can be used in the lower models (with limited amount of cores etc or worse thermal efficiency).
"A developer driver inadvertently included code used for internal development which removes the hash rate limiter on RTX 3060 in some configurations," the company said in a statement. "The driver has been removed."[1]
So it's hard to tell when it's binning, and when it's hobbling.
[1] https://www.theverge.com/2021/3/16/22333544/nvidia-rtx-3060-...
“It’s not just a driver thing,” said Bryan Del Rizzo, Nvidia’s head of communications, last month. “There is a secure handshake between the driver, the RTX 3060 silicon, and the BIOS (firmware) that prevents removal of the hash rate limiter.”
Of course, distributing a "cracked" driver would breach licensing agreements and violate copyright (and possibly other laws), so it's out of the question for any kind of professional use. However, if Nvidia uses it to differentiate services, wouldn't there be a significant risk of private consumers (gamers, hobbyists, etc) using such a driver to "upgrade" their cards for personal use?
The firmware itself truly is signed and so on
Kepler could literally have all its cores enabled with a simple bridge on one laser cut. That's fun but very hit or miss, the more cores are disabled the higher the likelihood they're actually defective so it was binned lower.
Maxwell can have their Hwid modded but performance will be lower than actual Quadros.
Newer cards have an even bigger performance difference, more hardware differences I think.