Ok I see what you are saying but if the interposer wasn't silicon based, you have other things to look into, latencies of the signals, this is why they went with what they did, I think it was explained in one of their presentations.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Ok I see what you are saying but if the interposer wasn't silicon based, you have other things to look into, latencies of the signals, this is why they went with what they did, I think it was explained in one of their presentations.
primetime
Supreme [H]ardness
- Joined
- Aug 17, 2005
- Messages
- 7,338
This is exactly why hard core amd fans wish the current furry x wasn't selling as well as it is. We certainly don't want amd to fail, we just wanted a 650 dollar product that lived up to expectations and we didn't get it, nor will fury x ever be apparently.
D
Deleted member 93354
Guest
Since you have an EE degree
Maybe you can tell me what this circuit does that I designed myself...
http://i296.photobucket.com/albums/mm164/yellowblur07/zpsir0gkqkx_zps61tfmlp1.png
-Degrees in Comp Sci and Aerospace (and former Satellite engineer)
You can't sit there and claim it can't be done without knowing the release plans of all the suppliers...And it's entirely possible the Fury chip has all the silicon necessary already as 8GB might have been the original design goal on 20nm.
It appears to be a timer circuit for a relay.
I never said it couldnt be done, just not cost effectively in a commercially viable product. In its current form, it would be physically impossible to add 4 more stacks to the interposer.
I never said it couldnt be done, just not cost effectively in a commercially viable product. In its current form, it would be physically impossible to add 4 more stacks to the interposer.
Last edited:
The Fury chip probably wasn't aimed at 20nm, These chips were back up plans when 20nm wasn't a viable solution. The 20nm chips were probably completely different.
5700 wasn't performance competitive, it was way slower than AMD. Fury X is basically as fast as Nvidia's fastest card. Big difference. All it needs is 8GB RAM.
AMD is closer than ever to the performance crown, well at least since what 9700 days? They are now just a tad slower if at all, whereas last gen, the gen before, the gen before, the gen before, the gap was much wider in favor of Nvidia.
Although it was obvious to happen as Nvidia chips have been WAY huger than AMD's for a long time, so as AMD even got close AMD's much more die area efficient chips were bound to catch up as long as we were stuck on 28nm.
In other words it was like Nvidia...faster, 550mm chip. AMD, 20% slower, 300mm chip. Except 550mm is the reticule limit, cant get bigger than that. See how Nvidia was trapped? AMD still had room to grow, that is why Fury X is now as fast as anything Nvidia has according to benchmarks. It will probably continue that way if AMD is aggressive on the next node, if it ever comes. On 28nm AMD will be faster than Nvidia next chip if there is one.
Last edited:
AMD is closer to the performance crown than they've been in many generations (basically tied now) and this is a failure? Funny...
Derfnofred
Gawd
- Joined
- Dec 11, 2009
- Messages
- 606
Citation/reference/link? Not to be antagonistic--just nerd out on computer architecture. (EE Ph.D type, after all)
TaintedSquirrel
[H]F Junkie
- Joined
- Aug 5, 2013
- Messages
- 12,691
It's just a rumor, I think.
But it goes way back to Q3 or Q4 2014.
Tainted its not a rumor, it takes years to make a GPU, not 1 year, actual tape of Fury was probably around Sept ish last year maybe even earlier, the design was started probably in place two years before that.
There would be no way to switch the same chip from 20nm to 28 nm in the time frame they knew 20 nm wasn't viable. This goes the same for Maxwell 2.
Both AMD and nV have parallel designs in place just in case something goes wrong, Even the g80 was ready by the time the 7800's from nV were launched.
Derfnofred
Gawd
- Joined
- Dec 11, 2009
- Messages
- 606
I understand that, razor1, but that would suggest that they have major architectural changes pipelined for a smaller node that are otherwise ready to go. Given that die shrinks for graphics ASICs are about the "cleanest cut" things you can do in this field (given the huge amount of repetitive units, a large portion of a die shrink's advantage is pure surface area), I'd be very surprised if they have a drastically different chip design sitting in the wings (GCN 2.0, or heck a serious GCN 1.3), just waiting for a node shrink. A major architecture change makes as much sense on 28 nm as it does on 20 nm unless it requires a *huge* change in transistor count per functional unit. Terascale, with its various tweaks and tunings, saw several die shrinks.
It really looks like the Fury is an oversized Tonga (GCN 1.2) with a new memory controller, I don't think they had ready any kind of major architectural jumps (or even something like a terascale->terascale 2).
That said, with how hot they run, I'm surprised they haven't jumped down the the SOC-20nm node on some of their lower-performing chips.
It really looks like the Fury is an oversized Tonga (GCN 1.2) with a new memory controller, I don't think they had ready any kind of major architectural jumps (or even something like a terascale->terascale 2).
That said, with how hot they run, I'm surprised they haven't jumped down the the SOC-20nm node on some of their lower-performing chips.
seanclayton
Gawd
- Joined
- Oct 25, 2008
- Messages
- 882
Eh. The most important market is the middle, where the coilwhine GTX 970 and its jacked up VRAM dominates. Release Fury Nano with better price/performance and AMD will be fine.
I understand that, razor1, but that would suggest that they have major architectural changes pipelined for a smaller node that are otherwise ready to go. Given that die shrinks for graphics ASICs are about the "cleanest cut" things you can do in this field (given the huge amount of repetitive units, a large portion of a die shrink's advantage is pure surface area), I'd be very surprised if they have a drastically different chip design sitting in the wings (GCN 2.0, or heck a serious GCN 1.3), just waiting for a node shrink. A major architecture change makes as much sense on 28 nm as it does on 20 nm unless it requires a *huge* change in transistor count per functional unit. Terascale, with its various tweaks and tunings, saw several die shrinks.
It really looks like the Fury is an oversized Tonga (GCN 1.2) with a new memory controller, I don't think they had ready any kind of major architectural jumps (or even something like a terascale->terascale 2).
That said, with how hot they run, I'm surprised they haven't jumped down the the SOC-20nm node on some of their lower-performing chips.
thats because thats what it is
AMD bet the farm on Mantle and are now locked in GCN with little change for the next 2-3 or more years now or risk breaking what few games use it
Mantle was never meant to be open or run on any thing BUT GCN AMD has even admitted such link > https://youtu.be/DuA3T9MzNic?t=6m30s
I understand that, razor1, but that would suggest that they have major architectural changes pipelined for a smaller node that are otherwise ready to go. Given that die shrinks for graphics ASICs are about the "cleanest cut" things you can do in this field (given the huge amount of repetitive units, a large portion of a die shrink's advantage is pure surface area), I'd be very surprised if they have a drastically different chip design sitting in the wings (GCN 2.0, or heck a serious GCN 1.3), just waiting for a node shrink. A major architecture change makes as much sense on 28 nm as it does on 20 nm unless it requires a *huge* change in transistor count per functional unit. Terascale, with its various tweaks and tunings, saw several die shrinks.
It really looks like the Fury is an oversized Tonga (GCN 1.2) with a new memory controller, I don't think they had ready any kind of major architectural jumps (or even something like a terascale->terascale 2).
That said, with how hot they run, I'm surprised they haven't jumped down the the SOC-20nm node on some of their lower-performing chips.
Well we will probably never know what the 20nm chip would be like, but I would think it would be similiar architecturally, maybe with more TMU's and ROPs. The 28 nm chips that just came out, sacrifices were made, both Fury X and Titan don't have good DP performance, and of course both are near the reticular limit.
rumartinez89
Limp Gawd
- Joined
- Apr 16, 2012
- Messages
- 384
Fury X DP is more than double that of the titan x so it is still decent, the issue is which games actually leverage the compute performance. They might be able to undercut Nvidia on the facial recognition software and what ever else Nvidia is trying to do.
No these are gaming cards, not to be used for anything else. That's why DP performance is crap compared to last gen cards. DP isn't really used in games at least not yet the performance implications aren't worth it. Compute in games is usually done with SP.
i expect an hbm2 part coming as soon as hbm2 is ready
ok ill bit your troll bait explain this
http://www.extremetech.com/extreme/200286-not-dead-yet-amds-mantle-powers-new-vulkan-api-vr-efforts
durquavian
Gawd
- Joined
- Dec 13, 2014
- Messages
- 757
Wasn't DP used in Ryse: Son of Rome? That was the explanation when AMD was running so well ahead of Nvidia at launch.
Mantle as we know it right now, is dead, both the VR API and Vulkan are going to be different, do to different needs and evolution. They might use similar specs which is actually the same case for almost all API's since they are run on hardware that have been based off of specs, even in previous gens. Direct X in the past recent years have been setting the standards that even Open GL had to conform to. Way back in the day , Open Gl was doing that.
There is an interesting line in that article, select partners with specific needs. Mantle is going now be something that is customizable for developers, most likely for new types of hardware, so as an API, its going to contain many branches or subbranches, almost like a game engine. I don't see this being good at a hardware level, that means standards are going to change based on what the developer needs?
Edit, this is what also was done with Open Gl, and it became a big mess, I'm hoping this isn't the case with Mantle.
Last edited:
yes but as we know it today, its pretty not going to be used. And no development company in their right mind would want to spend extra resources unless there is a necessity for it.
most of the developers have said its dead simple to implement, so im not sure thats true. The oxide guy did it buy himself in a month in his basement.
its also wholesale in Vukan, so hopefully that will gain traction.
My guess is they will trot it out from time to time to light a fire under someones ass.
Yeah... now do it for every GPU on the market.
Dude will come out of his basement whiter than snow.
lol...
its only for AMD, so im not sure what your point it.
Razor1, the lead developer on Civ:BE said there are only two important numbers when developing multiple APIs. N, and N+1. If you have to do 2 apis, the development cost and time for each additional API is trivial.
Which makes sense, if you need more than one API, in order to re use your code, it has to be highly modular.
its only for AMD, so im not sure what your point it.
Razor1, the lead developer on Civ:BE said there are only two important numbers when developing multiple APIs. N, and N+1. If you have to do 2 apis, the development cost and time for each additional API is trivial.
Which makes sense, if you need more than one API, in order to re use your code, it has to be highly modular.
The issue there is that there are no legitimate DP-heavy features for games.
NVIDIA figured that out two generations ago.
In many cases even SP is overkill.
I agree with that, if the engine was made from the beginning with that in mind. but this isn't the same in all occasions, optimization from different api's do change, that's where things get more complex. And if you want a compelling product for something like a game engine that is something you have to look into. The base game code doesn't change much if at all.
We have seen things like unreal engine and cry engine, with its Ogl path that is much slower then the DX path.
rumartinez89
Limp Gawd
- Joined
- Apr 16, 2012
- Messages
- 384
I wonder if Pascal will be like Maxwell and have no DP ability. If it does not then I guess that means Nvidia will end up developing two architectures one for compute and one for gaming. Where amd would have one architecture for both because DP is ingrained in the GCN stream processor design. By the way what would be a DP feature in games?
Vulkan is a non-starter wont be ready for 2-3 years in the mean time DX12 will take over on PC and Apple is dropping OpenGL for there own api and wont be supporting Vulkan at all
and no one cares about linux and MS sunk steamOS free upgrades to 10
cageymaru
Fully [H]
- Joined
- Apr 10, 2003
- Messages
- 22,091
Lots of games have an OpenGL path. A lot more than you think. Having multiple paths is not hard at all for developers.
how many have it to support Apple? how many use it as the main api 2? maybe 3? in the last 5 years
openGL is dead vulkan will be stillborn DX12 will rule on PC and Metal will rule on OS X/iOS
openGL might linger on on android but thats it
Android is a pretty large platform.
I can't imagine any processing in a game that would require double precision.
I think for Pascal nvidia might do a separate architecture / chip for their Tesla line sort of like what they did with Kepler. They're also emphasizing short floats (16-bit) for the consumer stuff; I'm not sure if that will be better for games than SP or not.
rumartinez89
Limp Gawd
- Joined
- Apr 16, 2012
- Messages
- 384
That does make sense, according to the [H] review of the Fury, AMD already has float 16-bit instructions cooked in. From the looks of it AMD might have tried to do to much with one card and any type of refresh would probably be focused on the back-end of the chip.
Derfnofred
Gawd
- Joined
- Dec 11, 2009
- Messages
- 606
Okay, yeah, we're on the same page.
First iteration of Pacsal will have a big gain in DP performance if we follow nV's roadmap and this will be used for both gaming and professional I think (rumors).
Oh want to add this in, there was talk about thermal runaway in a post that I couldn't find, probably in another thread, thermal runaway occurs close to the phase change of silicon, so anything above 125 degrees C thermal runaway, becomes an issue, not before, resistance does increase all the way up to the point thermal runaway begins at which point resistance decreases. I think a battery was used as an example, but different materials have different threshold for thermal runaway.
Last edited:
D
Deleted member 93354
Guest
You are correct. DP is used for things like bitcoin mining (cloud apps), image processing, and CAD CAM
NVIDIA offers improved DP performance on their professional lines of quadro cards. But those cars cost a buttload.