Support NeoGAF

[KillaKalz347]

One of those could be the 490x, or Big Maxwell, or Pascal too. Nobody is really doubting that it's the next thing after GDDR5 at this point since both main card makers are on-board with it. Just that we don't know the timetable of actual products.

I doubt the first revision of big maxwell will feature HBM,they would be more likely to bring the circuit to 20nm in late 2015 with HBM. than release a 28nm card with 20nm HBM.

 

[Beerman462]

I really hope AMD brings it in the GPU market. Nvidia needs competition. I used to buy AMD CPUs, but atm there is absolutely no reason to. I feel like Intel has taken advantage of it and is pushing out virtually meaningless updates.

 

[Objectively Bad Opinion]

It's said that since July. There are no commercial products using it, nor any announced. If you look at this reporting on the leaked slideshow, it mentions the same tech specs, but doesn't include the corresponding rumours that it's going to show up on Fiji in the first couple of months of next year.

It's theoretically possible, but I am not putting stock in wccftech unless we get independent confirmation.

If it has said that since July than it is more likely that we'll see it used in a product sooner rather than later.

I don't know why you're so adverse to AMD using it before Nvidia. AMD had the floor on GDDR5 for like two years before Nvidia got in on it.

 

[ThoseDeafMutes]

I don't know why you're so adverse to AMD using it before Nvidia.

I'm not? Even if it skipped 2015 entirely, Pascal (first one to use stacked 3D memory according to roadmap) is set for 2016, which would leave plenty of opportunity for AMD to move first on it.

 

[Paracelsus]

What does that mean for people looking at the gtx970 who don't want to upgrade for the next three years and four months?

 

[ThoseDeafMutes]

What does that mean for people looking at the gtx970 who don't want to upgrade for the next three years and four months?

That's a really specific amount of time that you don't want to upgrade for.

If you are in need of performance now or early next year, then the 970 is a fantastic card, and you should upgrade now. If there's nothing you're going to play that needs an upgrade between now and next year, then don't upgrade. A 970 will still be able to play games in 3 years and 4 months, but we can't predict the future with a lot of precision as to how well it will handle them. It might be the new 8800GT, able to hold out for 4+ years (although getting long in the tooth by the end of that!) or it might age very poorly.

 

[Brad Grenz]

So I've been reading this presentation linked on the first page.

Am I really reading this right that you'd need a 4096 (!) bit bus (128 bits per channel, 8 channels per stack, 4 stacks per card) to reach 512 - 1024 GB/s bandwidth when GDDR5 reaches 336 GB/s with 384 bits? Isn't a memory bus that wide going to a pretty huge hindrance in design?

No. That's one of the major benefits of these stacked memory designs. You don't have to run traces on a PCB to a bunch of memory chips. You either place the memory stack side by side with the GPU connected via a silicon interposer, or it's placed on top of the GPU and connected directly to the GPU die connected via vertical elements between the packages.

EDIT: Found the slide that shows how the interposer connects the HBM to the processor.

 

[Nikodemos]

What intrigues me is how would they implement HBM in their APU line.

They obviously can't sell a package APU+interposer+HBM 2.5D solution (it'd look like those old Slot A Athlons). The fullsize 4-HI 4x2 Gbit (1 GB total) 'sandwich' can only be used in high-end laptops, since it would not be profitable enough otherwise. However, I wonder if AMD could add a single 2 Gbit (256 MB) layer on top of a desktop Carrizo in a fully 3D config and use it as L3 cache. It would still offer considerable performance improvement.

 

[NumberThree]

No. That's one of the major benefits of these stacked memory designs. You don't have to run traces on a PCB to a bunch of memory chips. You either place the memory stack side by side with the GPU connected via a silicon interposer, or it's placed on top of the GPU and connected directly to the GPU die connected via vertical elements between the packages.

EDIT: Found the slide that shows how the interposer connects the HBM to the processor.

That's interesting, thanks! Interposer led me to a bunch of other cool stuff on Google. At least once it'll look like I'm browsing work related stuff

I'm assuming being placed on top of the GPU is out of the question with modern cooling solutions, but that's an interesting solution nonetheless.

 

[Inuhanyou]

Sony and MS done goofed for not going with high end cards... Unless of course PS4 and xbox one have like a 3-4 year life cycle. It would make perfect sense considering Sony went for x86 architecture most likely for backwards compatibility.

Its a good thing nobody buys consoles specifically for the highest end graphics or performance in general, or PS3 and 360 would have been dead 2 years after they launched.

With die shrinks(slim revisions), price cuts and exclusive games, i think they can be competitive and hold on for the average 5 to 6 year cycle.

If there is a next gen after that, by the time it rolls around, newer architectures, better processes and better technologies like this RAM breakthrough will be in a mature stage...and the absolute high end today will be the low end budget option very viable for a console.

I like to think of it like, the smaller the upgrade was from 7th gen to 8th gen, the bigger the upgrade from 8th to 9th.

 

[rezuth]

oh my god....


I seriously do not want to buy an intel CPU.

AMD pls

Dont worry we will switch to ARM in a few years.

 

[Sentenza]

wccftech has also mentioned that AMD's Sea of Tranquility fabrication plant will be coming online in the near future and that AMD is confident that it won't encounter any serious yield issues with their new 7 nm process.

Well, time to wait for 7nm GPUs and skip everything else then.