8GB 290x or 4GB Fiji?

[soccerballtux]

I have a friend and I'm trying to make up my mind on what to recommend to him.

I'm specifically interested in how some 4GB cards from the era that '2GB cards were the norm' are faring in higher-memory-usage-games (I don't know what these would be).

 

[Udgnim]

4GB Fiji because 290X 8GB isn't powerful enough to make use of 4GB+ rendering unless you crossfire and that's probably still not worth it versus single GPU Fiji 4GB

 

[LTC8K6]

Wait a bit and see what 4/8gb cards are released by AMD.

 

[Saika]

Depends where you live.In Europe the Fury card will at least cost 1000 euros (Titan is roughly 1250).You can crossfire a 8GB card for 100 to 200 euros less and profit of the 4 extra GB.Who knows, they might even get cheaper soon.
That's my plan honestly, crossfire them for 1440p and uppgrade when the second generation of 14nm comes out at finally affordable prices (if ever).1000 to 1250 euros for cards that are barely better is a joke.

 

[LTC8K6]

Also need to see if 4gb HBM has an effect on the frame buffer argument.

 

[gamervivek]

290X doesn't have dx12_1/11.3 features, so wait until Fiji comes out.

 

[LTC8K6]

290X doesn't have dx12_1/11.3 features, so wait until Fiji comes out.

DX12 support, which all remotely current cards have, should be sufficient, I would think.

 

[Spanners]

DX12 support, which all remotely current cards have, should be sufficient, I would think.

I agree, I don't recall anyone regretting not having DX11.1 for example.

 

[gamervivek]

DX12 support, which all remotely current cards have, should be sufficient, I would think.

I'm going on sebbi over at b3d saying that features like conservative rasterization and ROVs are quite helpful and he usually has good things to say about AMD. Since only 2nd gen Maxwell supports them at the moment, I'd be very surprised if they aren't put to use rather quickly unlike 11.1/2 which I think only Hawaii did or did to a satisfactory degree.

 

[LTC8K6]

AMD said that all GCN chips will have full DX12 API support, so I'm not worried about that.

Whatever specific functions don't exactly work probably won't matter much.

 

[gamervivek]

dx12 and dx12.1 are different and not all GCN will support all dx12(0) features.

They'd matter since they are good and nvidia supports them. I do hope that Fiji will support them, but lack of leaks regarding it haven't been very encouraging.

 

[Silverforce11]

Devs will continue to develop around what consoles are capable of. So thats full DX12, without FL12.1 from Maxwell 2. If NV sponsors PC ports, they can put in a few GameWorks features that uses 12.1.

 

[LTC8K6]

There is only DX12.

12.1 is a feature level.

DX12 FL12.1, for example.

Here is a thread that discusses the issue, but gets nowhere.

http://forums.anandtech.com/showthread.php?t=2425942

This is not going to matter at all, imo.

All current cards offer sufficient DX12 support so that no one needs to worry about it.

 

[gamervivek]

sebbi is a dev. and you don't necessarily need higher feature levels for just gameworks, you can use them to speed up your own gpus provided the developer plays ball. See the dx10.1 fiasco in Assassin's Creed.

And Kepler will keep falling behind since it was behind Hawaii in feature levels which is behind Maxwell in turn, so nvidia didn't have any incentive to include them earlier. It'd be interesting to probe if newer games do include them, thoug hnvidia would be making more noises if that was the case.

 

[Lonyo]

Tell him to wait until the 4GB card is out?

 

[ShintaiDK]

290X doesn't have dx12_1/11.3 features, so wait until Fiji comes out.

Fiji wont have this either if its based on GCN 1.2 or lower.

The biggest issue will be selling it as 4GB and water cooling only. Also in terms of resell value.

 

[RaulF]

I would say fiji if the performance is there. I might get one if it beats the 295X2, is water cooled (not going back to air) and the 4 GB will not be an issue.

And before someone says that my current card is only 4GB, remember that under DX12 my card will be a full 8GB since all the memory will be accessible (or so we been told).

 

[ShintaiDK]

And before someone says that my current card is only 4GB, remember that under DX12 my card will be a full 8GB since all the memory will be accessible (or so we been told).

No...Also the claim was only made to Mantle and without any examples and specifications in a single tweet. It was nothing but hot air.

 

[RaulF]

No...Also the claim was only made to Mantle and without any examples and specifications in a single tweet. It was nothing but hot air.

I would make sense that if they use SFR this would be possible, as each core is making one frame using it's own memory. As i posted and as you said just a rumor. Does not seem unrealistic to me.

 

[ShintaiDK]

I would make sense that if they use SFR this would be possible, as each core is making one frame using it's own memory. As i posted and as you said just a rumor. Does not seem unrealistic to me.

Its not possible with SFR. You still need 2x the memory. It was simply a BS statement.

The only possible way is to do post processing on 1 GPU. But then you still only have and need 4GB on the first one while the one doing post processing can get away with much less. However no sharing or any other "magic".

 

[zagitta]

No...Also the claim was only made to Mantle and without any examples and specifications in a single tweet. It was nothing but hot air.

:whiste:

 

[ShintaiDK]

You forgot the rest of the story. Each card needs duplicate textures etc. So you are not getting 8GB at all in any way. Unless you somehow imagine that one card pulls memory from the other at a staggering 8GB/sec.

Multiadapter is mainly for post processing. And it adds lag as well.

http://www.pcper.com/reviews/Graphics-Cards/BUILD-2015-Final-DirectX-12-Reveal

I am quite aware that AMD tries to upsell the "4GB experience" after they shoot themselves in the foot going with premature HBM.

 

[Noctifer616]

You can check Mantle documentation, it has a chapter on mGPU and memory sharing. You guys can read up on it if interested.

 

[zagitta]

You forgot the rest of the story. Each card needs duplicate textures etc. So you are not getting 8GB at all in any way. Unless you somehow imagine that one card pulls memory from the other at a staggering 8GB/sec.

It's entirely possible to program your way out of a slow interconnect between two memory pools, take a look at how NUMA-aware HPC processing works.

Either way I was merely pointing out your statement being blatantly false which you seem to have realized after editing it

 

[ShintaiDK]

It's entirely possible to program your way out of a slow interconnect between two memory pools, take a look at how NUMA-aware HPC processing works.

Either way I was merely pointing out your statement being blatantly false which you seem to have realized after editing it

NUMA is quite fast (~½ speed) compared to the differences between GPU memory and PCIe.