22 comments

[ 4.4 ms ] story [ 183 ms ] thread
buried:

> none of the above benchmarks benefit from Nvidia's professional drivers — Siemens NX for example typically sees massive performance gains compared to GeForce cards

Fascinating.

Working in the ecosystem around Siemens NX, I think it's just all of dev work in the field being very use-case specific. Development happens in tandem to big customer projects, so I'd guess non-pro drivers were not just not optimized for, but were simply never on the devs machine in the first place. Maybe they use some feature, that is locked away in non-pro drivers, emulated in standard drivers.

Stuff like 64-bit double precision. I don't even know, is that still locked away behind pro drivers?

I can't really buy 4090s for rack mount servers though. Power/cooling can be a pain.
Not to mention the cuda licensing issue with running consumer GPUs in a data center
Can anyone explain to me why it performs so "bad" compared to the 4090, although the specs are way better?
4090 has a higher power limit so it can clock the cores higher to make up for the difference. Unless you need more than 24GB working set in vram it will do just fine.
4090 also doesn’t have ECC RAM, which makes its memory slightly faster as long as you don’t care about correct results…
I see an option for ECC in the nvidia control panel for my card…
A workload that only needs 24GB of VRAM will be faster on the 4090. It has higher TDP, same die as the RTX 6000 Ada and nearly the same number of cores (the difference being irrelevant).

The price difference is for the VRAM.

> However, if you don't need a blower GPU and use applications that aren't specifically de-optimized by Nvidia's GeForce drivers, the RTX 4090 ($1,599 MSRP) can lend you its firepower for 76% less than the cost of an RTX 6000 Ada.

Does nVidia specifically de-optimize (read: handicap) GeForce drivers for certain applications?

Somewhat, yes. Eg the consumer GPUs limit the fp16 accumulation rate to limit how well that scales for training models.
That is only possible because of effective monopoly. How is that not getting sued to bones I don't know.
Because monopolies aren't illegal.
Product segmentation isn't illegal... as much as it sucks for us consumers.

It's a similar thing to AMD sometimes (especially later into a process node's maturity) taking higher quality chiplets with more functional cores and turning off the cores to sell in a lower cost higher demand product or Intel doing similar with their chips.

It'd be illegal if they were lying about it.

Artificially segmenting when there is no ongoing cost seems like borderline fraud. Subscription payments for heated seats or dark cores strike me as destroying the environment and fleecing consumers for short-term gains.
If I wanted to buy myself a computer that I can stick that 4090 into, what's the minimum amount that I would spend such that I don't end up bottlenecking the GPU.

Not for gaming, if that makes any difference.

Depends on the workload really. Some workflows involve copying large amount of stuff from disk to RAM and then over to VRAM on the GPU, some workflows are really inefficient while others require basically no VRAM at all.

For example, for fast 3D renders of small scenes you just care about processing speed, large scenes need good memory bandwidth too, while with LLM training you'd care more about available VRAM, and so on.

But, just to throw out a number out there, you could probably get away with spending as little as 300-400 USD for just a motherboard, PSU, CPU, cooler and RAM, assuming you don't want things like storage, cases and whatever accessories. Basically use it as a server.