Amd Volcanic Islands details

[nossy23]

So, would people be happy with a card from AMD that has a performance level in the neighborhood of a GTX780/Titan? Sure it would be cool to get lower prices, but at the same time it's worrying. This is AMD's big upgrade, their new gen GPU, not just a GPU that is clocked a bit higher with some optimizations. Nvidias next gen will be out in, what, about half a year after AMD releases their new cards? And if Maxwell delivers even half of what is promised, AMD's next gen GPU is toast.

[Dimitriman]

If Maxwell is on 20nm then AMD will not compete with Hawaii but rather, a new 20nm card obviously.

People need to realize that all performance expectations and predictions on Maxwell are based on it being at 20nm. There will be nowhere near that level of improvement if it is made in the same 28nm process, everyone knows the challenges of improving performance on the same node yet a lot of people expect a miracle 100% improvement of performance on Maxwell at 28nm. Cumon...

Besides that, AMD and Nvidia do not have identical manufacturing schedules, just because one releases a product today, doesn't mean the other can and should release a new product equally competitive in the same period.

[Iconyu]

So, would people be happy with a card from AMD that has a performance level in the neighborhood of a GTX780/Titan? Sure it would be cool to get lower prices, but at the same time it's worrying. This is AMD's big upgrade, their new gen GPU, not just a GPU that is clocked a bit higher with some optimizations. Nvidias next gen will be out in, what, about half a year after AMD releases their new cards? And if Maxwell delivers even half of what is promised, AMD's next gen GPU is toast.

I think AMD are trying to make GPGPU much more important this generation, the 680 and 780 are weak in this area while the 79*0's are strong. For example I said a few months back, that GPU compute was going to become very important:



After seeing this, if Nvidia are going to make a Maxwell with most of it's GPU compute removed it's going to be very vulnerable next year, especially with AMD's developer ties. So it doesn't matter if the benchmarks next month say it's only on par with Titan, what really matters is have AMD got the compute balance right to make all the PS4 and XBOne ports run like silk. Engines built to take advantage of GPU compute might become the future, and for once AMD are holding the reins.

[LordEC911]

So, would people be happy with a card from AMD that has a performance level in the neighborhood of a GTX780/Titan? Sure it would be cool to get lower prices, but at the same time it's worrying. This is AMD's big upgrade, their new gen GPU, not just a GPU that is clocked a bit higher with some optimizations. Nvidias next gen will be out in, what, about half a year after AMD releases their new cards? And if Maxwell delivers even half of what is promised, AMD's next gen GPU is toast.

Why? Hawaii is meant to compete with GK110...

[PedantOne]

i am sure, if will have gaming SKU 2816 SPs, can beat Titan in performance. I hope for 512-bit mem bus, but not for width but 4 GB RAM.

[AbortRetryFail?]

So, would people be happy with a card from AMD that has a performance level in the neighborhood of a GTX780/Titan? Sure it would be cool to get lower prices, but at the same time it's worrying. This is AMD's big upgrade, their new gen GPU, not just a GPU that is clocked a bit higher with some optimizations. Nvidias next gen will be out in, what, about half a year after AMD releases their new cards? And if Maxwell delivers even half of what is promised, AMD's next gen GPU is toast.

No worries from the Red Team

My understanding: Volcanic Islands is more of a optimized *tweener* card still in the architectural range of Southern Islands, but a major preparation for the die-shrink to 20nm.

Bigger chip, expanded power-states, higher frequencies (if Bonaire, and Cypress ---> Cayman are any guide) My guess: 5.2 billion transistors, 2688 cores and 440mm2

Not quite a Titan, but Jen-Hsun Huang will be sweating bullets on those GTX780 price drops

[zalbard]

After seeing this, if Nvidia are going to make a Maxwell with most of it's GPU compute removed it's going to be very vulnerable next year, especially with AMD's developer ties.

You have to understand that cutting down GPGPU is a choice made by Nvidia. GK110 is better at GPGPU than anything AMD has to offer right now; the thing is, Nvidia makes a lot of money on professional GPU segment (unlike AMD), so they cripple gaming GPUs on purpose to get people who do need GPGPU performance to spend more.

Once there is sufficient demand for gaming GPGPU performance, I am sure they will step up.

This is very unfortunate, but there's absolutely nothing we can do about it. It's their business model.

I hope for 512-bit mem bus

Not happening, would increase die size and PCB cost too much. Not really feasible.

[To(V)bo Co(V)bo]

I'm getting 4 of the new cards as soon as they make one with 6 mini displayports :-/

It's called FirePro.

If anyone didn't know they got a killer promotion for 50% off the W7000 & W5000 till Oct 31.

[Iconyu]

You have to understand that cutting down GPGPU is a choice made by Nvidia. GK110 is better at GPGPU than anything AMD has to offer right now; the thing is, Nvidia makes a lot of money on professional GPU segment (unlike AMD), so they cripple gaming GPUs on purpose to get people who do need GPGPU performance to spend more.

Once there is sufficient demand for gaming GPGPU performance, I am sure they will step up.

This is very unfortunate, but there's absolutely nothing we can do about it. It's their business model.


Not happening, would increase die size and PCB cost too much. Not really feasible.

Aye, Nvidia's philosophy is to build the best card they can for today's games, tomorrows games will need another card.

Edit: @LordEC911: Aye that's pretty much it, when Nvidia were in Crytek's ear Crysis 2 went tessellation crazy, AMD now pretty much has every console game developer for the next 5 years under their sway, even the WiiU can exploit these features. If these particular strengths are exploitable in a way that actually makes games look better then it would be stupid to ignore them. But I do imagine that most developers won't allow it to get too out of hand and literally build games that cripple a 680.

[LordEC911]

Bigger chip, expanded power-states, higher frequencies (if Bonaire, and Cypress ---> Cayman are any guide) My guess: 5.2 billion transistors, 2688 cores and 440mm2

Think more with transistor count. I estimate, depending on the exact die size but assuming around 440mm2, around 5.5b-6b.

You have to understand that cutting down GPGPU is a choice made by Nvidia. GK110 is better at GPGPU than anything AMD has to offer right now; the thing is, Nvidia makes a lot of money on professional GPU segment (unlike AMD), so they cripple gaming GPUs on purpose to get people who do need GPGPU performance to spend more.

Once there is sufficient demand for gaming GPGPU performance, I am sure they will step up.

This is very unfortunate, but there's absolutely nothing we can do about it. It's their business model.


Not happening, would increase die size and PCB cost too much. Not really feasible.

I think, though you did mention it, he was specifically talking about gaming compute performance, possibly OpenCL?
They didn't remove any of that for GK110 dies and they still are getting beat by Tahiti.
http://www.anandtech.com/show/6973/n...-780-review/18
Possibly driver related?

512b is possible by using a more area efficient PHY like in Bonaire/Pitcairn but lower speed would easily give +300Gbps of bandwidth while decreasing the area of the PHY by a third or about 66% the size of Tahiti's total PHY. There is also the comment that Stilt made about a rectangular die which suggests they are looking to maximize the perimeter.
We already know that Hawaii has a more complex PCB than Tahiti, more layers, so that could hint at either faster memory speeds(7Gbps) or a larger bus(512b) or just more power.

[NEOAethyr]

Geforce vs quadro, on the 680's they disabled dp, the titan though has this option as a gaming card (oddly enough).
Then you also have the usual diff's between quadro and geforce, a loss of a tiny bit of detail when drawing distances with mip mapping and some speed differences.

All in all, the 680 chip still has that stuff..., but on a geforce pcb it won't.

You got the 2880 shader k6000 (pro, dp), the 2668 shader titan (dp, gaming), the 2304 shader 780 (sp, gaming), and the 1536 shader 680 (sp, gaming).
(Edit: oh and who can forget the monster of a pro card the k2, 3072 shaders and dp, equ to the gaming card the 690 which only does sp)
The 2 that stick out in terms of dp ability are the titan and the 680, the 680 was originally a k5000 pro chip.
As for the 780, maybe it has the ability, maybe there was a quadro that was the exact equ, but to my knowledge it doesn't.

The 780 was a step down in costs compared to the titan.
Both really aren't the full fledged quadro equ.
I got the 680, now as far as gaming is concerned I wouldn't mind an upgrade, but once you go quadro, you wouldn't wanna jump ship so quickly, and that's how I feel.
I'll stick with it until I can find another exact gaming vs pro equ chip.
That and I don't have the money to jump onto another vga card so soon .

But I like to see prices drop, and many people don't mind amd/ati at all, I don't think dp is a big issue right now.
I think it's more or less a quality and feature thing then a performance thing.

I think the main issue this gen and the next is vram, and texture performance, because eventually we're gonna have to get an upgrade in the tmu department (without needed a giant expensive die that is...).
We ain't gonna be pushing the same amount of textures at the same old resolution for ever...


Edit:
In a few more gens from now, they'll have to focus on 3d monitor perf, they demo'ed last year some nice non-glasses 3d screens with viewing angles for up to 20 so on people at a time.
Sooner or later those will hit mainstream.
And performance right now in that department is a little meh, it could be better is what I mean.
But then again we have some games only using one cpu core :\ ...

[hurleybird]

GK110 is better at GPGPU than anything AMD has to offer right now

It is?

GK110 seems to go back and forth between Tahiti in GPGPU benchmarks, and that's at a significant die size advantage. GCN seems like it's at least as good at general purpose tasks architecturally (a theoretical Tahiti scaled up to GK110 proportions would likely thrash GK110 in the majority of GPGPU benchmarks), though Nvidia obviously has a software advantage with CUDA.

[dasa]

wonder how the cost compares of 4g 512bit vs 6g 384bit if its close i think i would rather the extra memory
as it has the potential to make the most noticeable difference but only if games are made to make use of it

although 7970 has ample memory bandwidth and increasing memory clock speeds from 5600-7600 makes very little difference that may change if what amd has for us is significantly faster
does titan see much benefit from overclocking its memory?

[tajoh111]

It is?

GK110 seems to go back and forth between Tahiti in GPGPU benchmarks, and that's at a significant die size advantage. GCN seems like it's at least as good at general purpose tasks architecturally (a theoretical Tahiti scaled up to GK110 proportions would likely thrash GK110 in the majority of GPGPU benchmarks), though Nvidia obviously has a software advantage with CUDA.

The drivers haven't been made for quadro k6000 yet so I don't think we can judge it yet.

As far as a compute card, tahiti is just so so for a next gen card. IT performs well for consumer purposes but not so well in the professional space. The firepro w9000 cards had tahiti and quadro 6000 fermi basically trading blows in the reviews available online(tomshardware and hothardware). Considering the computer power of tahiti it should simply decimate fermi but it doesn't.

http://www.tomshardware.com/reviews/...mark,3265.html
http://hothardware.com/Reviews/AMDs-...vidias-Quadro/

If AMD puts too much of a gaming emphasis on Hawaii, it's going to screw up the gains it made with tahiti. And particularly in the professional market, AMD needs to make bigger improvements, even more so than the gaming market for its top end card.

How well AMD big chips perform in professional spaces is what's going to make them money not so much its consumer performance(at least not for the next 3 years).

[hurleybird]

As far as a compute card, tahiti is just so so for a next gen card. IT performs well for consumer purposes but not so well in the professional space. The firepro w9000 cards had tahiti and quadro 6000 fermi basically trading blows in the reviews available online(tomshardware and hothardware). Considering the computer power of tahiti it should simply decimate fermi but it doesn't.

http://www.tomshardware.com/reviews/...mark,3265.html
http://hothardware.com/Reviews/AMDs-...vidias-Quadro/

Those aren't GPGPU benchmarks, and yes AMD needs to improve their Firepro drivers.

[nossy23]

Not sure why AMD would bring out their highest end GPU with just 3 GB of ram. Sure it's more than enough today. But with next gen consoles that have a lot of memory that can be used for graphical things, and taking in to account on PC we use (higher) AA etc, I can see the memory becoming a bottleneck.

[LordEC911]

If AMD puts too much of a gaming emphasis on Hawaii, it's going to screw up the gains it made with tahiti. And particularly in the professional market, AMD needs to make bigger improvements, even more so than the gaming market for its top end card.

The things Tahiti lacks, IMO, that Hawaii will make up for is specifically in the compute department.

[zalbard]

I think, though you did mention it, he was specifically talking about gaming compute performance, possibly OpenCL?

OpenCL is ignored to get more people to use CUDA. Whether that's wise or not, time will tell.
Pretty stupid, in my opinion, since some people will keep using OpenCL as long as CUDA only runs on Nvidia hardware.
Proprietary stuff rarely wins in the long run.
Not going to happen before OpenCL development is as easy and convenient as CUDA's, though...

It is?

GK110 has a lot of GPGPU features that AMD offers lack. It's not just performance.

[aussie-revhead]

Wow so we get 4 days warning of a solid release date? Had I known this a week ago I would have held off buying my 780 Lightning, I am sure it will do the job and I know I can always sell it but I would have liked to see how these go before lashing out.

[hurleybird]

GK110 has a lot of GPGPU features that AMD offers lack. It's not just performance.

Care to elaborate?

[tbone8ty]

Ruby is so hot!

[Elmy]

On the 25th AMD will be streaming at 3 EST 12 PST on their Youtube page about their new announcement.

So excited to see what they are bringing to the table!

[jvc]

Think more with transistor count. I estimate, depending on the exact die size but assuming around 440mm2, around 5.5b-6b.

512b is possible by using a more area efficient PHY like in Bonaire/Pitcairn but lower speed would easily give +300Gbps of bandwidth while decreasing the area of the PHY by a third or about 66% the size of Tahiti's total PHY. There is also the comment that Stilt made about a rectangular die which suggests they are looking to maximize the perimeter.
We already know that Hawaii has a more complex PCB than Tahiti, more layers, so that could hint at either faster memory speeds(7Gbps) or a larger bus(512b) or just more power.

Would be great to see this happen.

So, would people be happy with a card from AMD that has a performance level in the neighborhood of a GTX780/Titan? Sure it would be cool to get lower prices, but at the same time it's worrying. This is AMD's big upgrade, their new gen GPU, not just a GPU that is clocked a bit higher with some optimizations. Nvidias next gen will be out in, what, about half a year after AMD releases their new cards? And if Maxwell delivers even half of what is promised, AMD's next gen GPU is toast.

You have to remember this generation will be a very short lived one. As soon as 20nm is feasible for high-end we'll see new SKUs from both Nvidia and AMD.

[STEvil]

maybe mid-range market feelers, but I doubt top end SKU's will be "soon"

[zalbard]

Care to elaborate?

Dynamic Parallelism and Hyper-Q, for example. AMD GPUs don't support anything like that (yet).
CUDA, obviously, and it's more popular / convenient than OpenCL...
Their virtualization features are also very neat, but I haven't worked with them, so cannot comment on it.

Right now it's actually pretty simple. If you really want OpenCL only, you buy AMD. Otherwise, for GPGPU you buy Nvidia... And no, I don't mean gaming cards.