Why doesn't AMD or nvidia make dual gpu gpu's anymore?

Why doesn't AMD or nvidia make dual gpu gpu's anymore?

Other urls found in this thread:

anandtech.com/show/11294/amd-announces-radeon-pro-duo-polaris
youtube.com/watch?v=0ktLeS4Fwlw
youtube.com/watch?v=aSYBO1BrB1I
twitter.com/SFWRedditVideos

Because of SLI / Crossfire

AMD does
anandtech.com/show/11294/amd-announces-radeon-pro-duo-polaris

>Why doesn't AMD or nvidia make dual gpu gpu's anymore?
they are/were shit

Here's the rub: CF/SLI can easily double the performance because of how parallel game compute generally is. This is NOT conducive for the gpu companies to sell you a new gpu every 1-3 years. So they intentionally gimp the software end of things so you have to buy a 1080 Ti instead of two 1060's. And then instead of buying two 1080 Ti's you have to wait until the 3080 Ti to get a card twice as powerful, ensuring that you keep their profits up in the years to come.

As for dual gpu cards, they basically cram more shit into less space and therefore can't cool as well. Look at a 295x2 vs 2 290x's, the two cards can clock higher and cool better whereas the dual card is lower clocked from the factory. Also consider the 75 watts from the pcie slot, with two cards you get that on both, with a dual gpu you have to add more from 8/6 pin connectors to make up. Which really isn't a big deal.

Yeah but wouldn't a dual gpu card in SLI be absolute beast?

>AFR has no disadvantages whatsoever
Get fucked. Why do people have to make conspiracies out of everything?

There is rumoured to be a dual GPU version of Vega, but likely only for professional use.
20Teraflops on a single card.

If you enjoy stutter and abysmal efficiency, sure. The ~5 people in the market for that can just as easily buy a mainboard with 4 PCIex16@8lanes slots though.

Dual GPU is the future just like dual core CPUs were the future. It's just so much cheaper to produce two midrange GPUs than one high end one.

I didn't mention that did I?

Companies wanting you to pay for performance IS conspiring you fuckwit. It's just legal conspiracy.

Shit scaling, too much heat, very high power consumption.

You drastically oversimplify the topic. No, you can't easily double (gayming) performance by using more GPUs in various modes. They've been trying that since 3dfx was around. It has huge drawbacks. And don't tell me nvidia and AMD haven't been jumping at each others throat trying to do it better than their competitor. That's the reason why tools like FCAT came out to begin with.

explain how near 200% scaling is a thing then?

because it is.

So is 200% scaling a thing or do you have to buy 1080ti instead of two 1060?

depends on the software involved.

It is in some isolated cases and then you still have the frametime issues. SFR isn't a thing yet because no one bothers with vulkan/dx12. Even I'd it eventually gets used it won't scale as good.

Absolutely false.

The technology we're talking about here is two individual dies talking across a communications channel while utilizing the same on-board resources. This is not a case of dual gpu = dual core cpu, because the dual core cpu has both cores located on the same die, removing a large portion of communication equipment between. This is more directly parallel to a twin CPU motherboard. The problem is that the equipment between the two die or physical processor housings is much less efficient in transmitting information. While it is cheaper to put the chips on a single board, it is much better performance to simply offload the communication onto separate hardware designed for it, and let the GPU hardware do what it is designed for.

> CF/SLI can easily double the performance
It can't, BTW. Tried BTC mining in the days of HD5xxx/HD6xxx and Fermi. When CF is enables, hashrate drops. When it's disabled, both cards are loaded at 100%.

> because the dual core cpu has both cores located on the same die, removing a large portion of communication equipment between
Remember Core2 Quad? Two Core2 Duo slapped together.

>two 1060
I wish

I'm going to say no, because no matter how good the hardware or software are, you will always lose some scaling in the connections. Currently, scaling hovers around 85%, reliably. In other words, if you double the frames you get from 1 card, you can reliably get 80-85% of that.

On the same fabric, in a time where most tasks just had each thread do its own thing.

It will go into that direction, with interposers coming up and companies already investing in technology to puzzle hardware together from pre designed blocks (AMDs infinity fabric that connects the two 4 core clusters in ryzen is also useable for die to die communication for example).

But that doesn't mean its easy to do or comes without performance penalty. If we go back to the topic you have GPUs with cache and memory systems designed to offer ridiculous bandwidths and low latencys. I should be obvious that pairing two of these things together over the slow, high latency PCIe bus can't make them work like one big chip.

Two GPUs on one card already counts as SLI/Crossfire. Buying a dual-GPU card is like buying two shitty cards so you get microstutter in games and applications that aren't set up for SLI/Crossfire can only use one GPU off your card.

I expected dual GPU cards to be something different, like a master-slave setup where the computer sees one virtual GPU which offloads computations to its second core, but instead it's just two SLI/Crossfire GPUs in one slot.

Dual GPU cards were a fucking mistake.

AMD master plan
youtube.com/watch?v=0ktLeS4Fwlw
youtube.com/watch?v=aSYBO1BrB1I

IIRC it was "put shit dual GPU's on consoles so dev's adapt and use cheaper dual GPU's instead of more expensive single GPU's to kick ass"

Still on the same die, even if seperated. You're thinking if you took two Core2 Duos and put them onto a dual CPU motherboard.

No, it was two separate dies on the same fabric. Google it

no. core2quads were made up of two dies. Same thing is going to happen to GPUs.

Probably, but that won't be comparable to SLI/CF and their rendering modes. I'd they build up GPUs (or high performance chips in general) from individual dies they'll be on the same fabric/interposer connected by a high speed low latency interconnect.