Hacker News Reader

Comment by KeplerBoy | Hacker News Reader

That's not how this traditionally worked. A few generations ago board partners were allowed to solder whatever RAM they wanted onto the GPU board.

Boards from different OEMs were different in meaningful ways and OEMs had the chance to differentiate themselves from the competition with some actual engineering. More RAM, multiple GPUs per board, AGP to PCIe chips, you name it.

Nowadays Nvidia restricts is partners from all that fun and undercuts them with their own models. No wonder EVGA quit this game.

View on HN

The murder of SLI made me so angry. It was so wonderful to be able to buy a card, then a few years later buy another of the same card for a bargain price and boost your performance by ~60%.

That and increasing RAM down the road for the system itself were some of the few upgrades that made sense.

Usually by the time you could buy a high-end CPU replacement, it was a "better deal" to just replace the entire motherboard/CPU combo and get a whole new computer. But a second video card, some more RAM, additional SSDs, those made sense.

SLI was ultimately killed by the rise of TAA. Interleaving frames breaks the assumptions of TAA models about the presence of temporal history. That's why NVIDIA went down the path of building a better TAA model/upscaler instead.

TAA is also terrible and turns visuals into a smeary mess, particularly because devs keep building it into games in such a way as to be nearly impossible to turn off like in Halo Infinite.

Nah, you hate bad TAA. DLSS is TAA too and it's actually great, DLAA is real nice, or the cool kids are doing some DLSS+DSR thing (bigger intermediate space?). Halo infinite is the poster child for fucked-up TAA, same for RDR2, you know what works better than anything else? DLAA lol. It provides a minimum floor.

Games are just going to use temporal accumulation now. There is too much signal to be gained from re-using past samples. It makes things way cheaper and opens the door to extrapolation and non-uniform sampling etc. It's the least bad of all the options, and DLSS actually is quite good at weighting samples pretty reasonably. DLSS 2.5, 3.0, 3.5 and upwards are actually significantly better and that can be injected back into (non-anticheat) games and the bar will likely continue to be raised. It is a signal processing technique that recovers a lot of signal with very low "noise factor".

https://en.wikipedia.org/wiki/Noise_figure

The only difference between the consumer RTX cards and a Quadro card is memory (and maybe power pin location, which is better on Quadro cards for rackmounted chassis). If nvidia put more memory inside their consumer cards, professionals would just buy the relatively less expensive consumer cards rather than the Quadro cards which are twice the price.

True, but they weren't doing such a range, it was like 512MB or 1GB, 2GB or 4GB, and there wasn't such ML interest in them as there is/would be today - if there had been that market then EVGA et al. would've been lapping up that premium too, I don't think it really affects the argument who's responsible for placing the chips. (And I wasn't really aware of that shift.)

>That's not how this traditionally worked. A few generations ago board partners were allowed to solder whatever RAM they wanted onto the GPU board.

A few generations ago (1080/2080) there was no mainstream ML. Products change, product lines change.

Interestingly these changes precede what you might call mainstream ML. The crazy OEM stuff was already over by the time Maxwell arrived.

Yes, and NViDIA didn't care when it was all selling to gamers. Now it wants crypto bros (obv a little historical there) and ML people who are using it to make money to pay a higher price point. So they do it by restricting the options available at the affordable level. See also, what happens when you want to use group administration features in Windows.

NVIDIA did care when it was all selling to gamers - although at least they had the decency to use ECC memory and some other CAD relevant bits, as well as signed production-ready drivers. They've just transitioned from QUADRO to RTX as ML became the dominant professional tier.

Most crypto is Proof of Stake -- how much GPU market is there really for Nvidia?

Not much anymore (hence my historical comment)

> See also, what happens when you want to use group administration features in Windows.

Care to elaborate?

I think they're saying that Microsoft uses features like group admin as differentiating factors between Windows variant so that they can sell professional/enterprise-oriented variants for a higher price than Home. So, same sort of "pro features in the pricier Pro variants" logic.

Thank you for explaining what I meant better than I did.

Reply on news.ycombinator.com