[Videocardz Rumor] AMD to launch new flagship Radeon graphics card this Summer

[Cookie Monster]

I really wish they improve their blower reference cooler. Aftermarket is all great, but im one of those that need a blower type cooler.

 

[3DVagabond]

I really wish they improve their blower reference cooler. Aftermarket is all great, but im one of those that need a blower type cooler.

I would be really surprised if they didn't. There have been reports of someone at AMD (sorry, don't recall who right now) when talking about the 295x2's cooler that they learned and won't make that mistake again. Here's hoping they meant it. :fingerscrossed:

 

[RussianSensation]

-- Roadmap fixed and removal of codenames(since they always change) --
28nm TSMC Q4 2013 : High-end GPU, VI 1.0
28nm GlobalFoundries Q4 2014 : Mid-end GPU and Low-end GPU, VI 2.0
28nm GlobalFoundries Q2 2015 : High-end GPU, VI 2.0
20nm GlobalFoundries Q4 2015 : Mid-end GPU and Low-end GPU, PI 1.0
20nm GlobalFoundries Q2 2016 : High-end GPU, PI 1.0
14nm GlobalFoundries Q4 2016 : Mid-GPU and Low-GPU, PI 2.0
14nm GlobalFoundries Q2 2017 : High-GPU, PI 2.0

Are you sure you didn't make a mistake when you translated those code names?

First, you have no high-end AMD GPU from now until Q2 2015! How is AMD going to compete with a supposed 28nm GM204 that's slated by end of this year? Or are they going to drop prices on R9 290X to stay competitive until Q2 2015?

Second, if NV moves to 20nm GM210 in Q2-3 2015, how is AMD going to compete with that chip using a 28nm VI 2.0? And AMD will not release a 20nm high-end until Q2 2016, almost a year after NV? Doesn't sound plausible.

Third, since HD7970 was released December 22, 2011 and almost all of those overclocked to 7970Ghz speeds easily, and yet, we won't see a high-end GPU, VI 2.0 until Q2 2015? Very disappointing if true. I mean frankly Hawaii was underwhelming for bringing just a 26% increase at 1080P and 29% at 1600P for a "next generation" part. If such a time frame is true, I sure hope this Q2 2015 part is at least 40% faster than R9 290X!

 

[Wild Thing]

295x sound about right
or
290x GHz+ be fine too.

chances are it will be nothing more than a 290x that is overclocked with an aio water cooling slap on.

Nope,the new card will feature enhanced GCN architecture,it wont be just a clock speed and voltage increase...in my opinion.
28nm is well understood now,the process is delivering low wafer errors and is considered mature.There's room for uarch tweaks and enhancements before the shift to the 20nm fab line.

 

[Z15CAM]

Under a Water Block a Reference 290X with crummy Elpida vRam will easily run at 1250 x1500Mhz with a + 0.152mV off-set and not throttle.

 

[Mondozei]

Two quick observations.

1. Notice that AMD dumped TSMC for their future roadmap. Still rumors but plausible. NV must feel the same urge.

2. 2015 will indeed be the year of high end 20 nm GPUs. In other words: 20 nm Maxwell is coming. Now can people please stop being silly and suggest otherwise?

Overall, the leaks look credible. 20 nm does indeed look like a breezer compared to the zombie-like 28 nm.

A bit disappointed if we see no flagship 14 nm in 2016 and only midrange. Hopefully Intels iGPU gains can put pressure on both NV/AMD to prevent that from happening in that time, by putting pressure on their margins in the low end segment.

 

[Mondozei]

20nm is short lived if they are launching it in late next year, I mean what's the point? People will know its going to be obsolete by 14/16nm stuff coming not long after since 20nm for dGPU is just so delayed. The gap will probably be less than 1 year. I wouldn't pay top $$ for 20nm stuff.

Arguments like these should remind us how stagnating this past half year or so has been as 20 nm GPUs died a quiet death in 2014.

It used to be a norm that you'd see a strong improvements from year to year in the GPU market. That is a good thing. The lack of that this year has infused a kind of battered wife syndrome into the debate where having slower improvements is somehow a good thing. That kind of argument is puzzling and a bit sad.

 

[Borealis7]

2. 2015 will indeed be the year of high end 20 nm GPUs. In other words: 20 nm Maxwell is coming.

20 nm GPUs died a quiet death in 2014.

is it 2015 and no one told me??

we are still in 2014, maxwell isn't coming (at least not on 20nm), and 28nm made a death rattle with the 295X2 but will draw it's final breath with a dual GK110 if that ever materializes.

2014 is going to be a very boring year in hardware, GPU- and CPU-wise.

 

[R0H1T]

20nm in Q4 2015? That's about half a year before Intel introduces 10nm with next-gen successor of Tri-Gate (a new material will replace silicon, probably germanium). With the advantage of 2 important technologies and 2 nodes ahead, Intel's IGPs might eat into the performance territory of dGPUs all the way up to the high-end, if they launch Gen10 with GT5 Cannonlake. And closer in time, 14nm Broadwell could disrupt the midrange market if it competes for a year against 28nm.

Keep your uninformed (Intel) PR spin to yourself, 14nm for desktop/laptop isn't out yet & you'll be lucky to see 10nm PC parts anytime before 1H of 2017, funny how you continually push Intel's (gen x) graphics as the next big thing when in fact they'll always be 4th grade, behind Nvidia/AMD/ARM D:

 

[Mondozei]

is it 2015 and no one told me??

we are still in 2014, maxwell isn't coming (at least not on 20nm), and 28nm made a death rattle with the 295X2 but will draw it's final breath with a dual GK110 if that ever materializes.

2014 is going to be a very boring year in hardware, GPU- and CPU-wise.

You haven't kept up with the argument, Borealis. The issue isn't whether Maxwell would be coming on 20 nm in 2014. Everyone knows that it won't and that has been known for a long time.
A lot of folks, however, have been saying we won't see Maxwell on 20 nm at all, period, that NV will just skip the node.

I'm surprised how someone can be so out of the loop of the arguments made.

As for your assertion that 2014 is going to be a boring year hardware-wise... yeah, pretty much. But if we're seeing not insigifnicant price reductions on the current generation of GPUs this autumn with the new GPU releases; such as a R9-290 for the price of a 280X today or so, then it could be pretty good, at least for those of us who want the raw GPU to power a 4K-monitor at decent framerates. (Read: Cheap new and capable GPU for crossfire setup).

 

[Despoiler]

So, maybe a test run for HBM memory? Slap it on Hawaii see how goes and tweak it for 20nm. There was also that talk of Hawaii having 3072 SPs, so maybe a fully enabled chip?

Could be. Nvidia was touting 3D memory in their keynote, which is tech they don't have yet. AMD invented HBM along with Hynix. I think AMD wants to be, and should be, first to market on something they helped create.

http://electroiq.com/blog/2013/12/amd-and-hynix-announce-joint-development-of-hbm-memory-stacks/

 

[witeken]

if it will be affordable maybe. iris is too pricey.

Iris for LGA doesn't exist yet, so you can't know that. And when they shrink to 14 and 10nm, Intel can put 2 and 4 times as much EUs in their SKUs without a die area penalty (so the same 1xx mm² as today), and they could offer a more expensive GTn+1 with 2xx mm² die area. At 10nm, Intel will be almost 2 nodes ahead in terms of density against TSMC's 16nm, so they could put around 3 times as much EUs within the same die area, meaning an Intel Core CPU + (16nm equivalent of) GTX Titan (550mm²) APU could fit within just ~230mm², with better power and performance too.

 

[witeken]

Keep your uninformed (Intel) PR spin to yourself, 14nm for desktop/laptop isn't out yet & you'll be lucky to see 10nm PC parts anytime before 1H of 2017, funny how you continually push Intel's (gen x) graphics as the next big thing when in fact they'll always be 4th grade, behind Nvidia/AMD/ARM D:

This isn't uninformed PR spin. This is primarily just simple physics: there are a lot of things you can do with 2 process nodes and 2 crucial technologies ahead (which will be the case for probably quite some time at 10nm). It's just a matter of how Intel will exploit this advantage. We'll have to see, but the possibility is there.

Secondly, Intel announced 10nm volume production in (Q4) 2015, so the means we should see products in mid-2016, which, even if it's 1 or 2 quarters later, is still a long time before TSMC has a real 14nm node to shrink the disadvantage back to 1 node (before Intel shortly after that obviously releases 7nm to push it back to 2 nodes). Those are reasonable assumptions based on current knowledge.

Lastly, this is not "funny", but rather rational reasoning. Do note that this just remains speculation, though. But statements like "Intel will always be 4th grade" without foundation are pretty much valueless.

 

[Creig]

Maybe a binned, high clock 290? R9 290XTX would be my guess.

 

[Sohaltang]

Maybe a binned, high clock 290? R9 290XTX would be my guess.

Even on water how much gain would that net? Beat a OC'd 780Ti on air?

 

[Granseth]

Iris for LGA doesn't exist yet, so you can't know that. And when they shrink to 14 and 10nm, Intel can put 2 and 4 times as much EUs in their SKUs without a die area penalty (so the same 1xx mm² as today), and they could offer a more expensive GTn+1 with 2xx mm² die area. At 10nm, Intel will be almost 2 nodes ahead in terms of density against TSMC's 16nm, so they could put around 3 times as much EUs within the same die area, meaning an Intel Core CPU + (16nm equivalent of) GTX Titan (550mm²) APU could fit within just ~230mm², with better power and performance too.

You keep forgetting other important parts of the GPU performance. It would be incredible complex and pricy to add the memory for a GPU like GK100 in an APU. I think we are far longer into the future before that will happen.
And then Intel have to take the cost of the driver team to start competing in high performance GPU arena.

I think if you are realistic Intel will move one step forward at each opportunity and not be leaping ahead and take the risk of it all going wrong.

 

[NostaSeronx]

First, you have no high-end AMD GPU from now until Q2 2015! How is AMD going to compete with a supposed 28nm GM204 that's slated by end of this year? Or are they going to drop prices on R9 290X to stay competitive until Q2 2015?

The reason for the potential delay for the high-end GPU. Is that peak full volume for HBM isn't expected till 1H 2015. So, the smaller designs being first which use less HBM stacks are given priority.

I also have no idea about the specifications for the low-end, mid-end, or high-end GPU. So, the stacks could be in;
1, 2, 4 configs
2, 4, 8 configs.
2, 4, 6 configs.

etc.

Low-end will probably be in the proximity of Tahiti LE.
Mid-end will probably be in the proximity of Hawaii Pro.
High-end will be above Hawaii XT.

---
What could also ruin my roadmap is that I misjudged how fast the shrink ramp is;

New:

28nm TSMC Q4 2013 : High-end GPU, VI 1.0
28nm GlobalFoundries Q4 2014 : Mid-end GPU and Low-end GPU, VI 2.0
20nm GlobalFoundries Q2 2015 : High-end GPU, VI 2.0
20nm GlobalFoundries Q4 2015 : Mid-end GPU and Low-end GPU, PI 1.0
14nm GlobalFoundries Q2 2016 : High-end GPU, PI 1.0
14nm GlobalFoundries Q4 2016 : Mid-GPU and Low-GPU, PI 2.0
10nm GlobalFoundries Q2 2017 : High-GPU, PI 2.0

vs

Old:

28nm TSMC Q4 2013 : High-end GPU, VI 1.0
28nm GlobalFoundries Q4 2014 : Mid-end GPU and Low-end GPU, VI 2.0
28nm GlobalFoundries Q2 2015 : High-end GPU, VI 2.0
20nm GlobalFoundries Q4 2015 : Mid-end GPU and Low-end GPU, PI 1.0
20nm GlobalFoundries Q2 2016 : High-end GPU, PI 1.0
14nm GlobalFoundries Q4 2016 : Mid-GPU and Low-GPU, PI 2.0
14nm GlobalFoundries Q2 2017 : High-GPU, PI 2.0

--
The reason for the change is the positioning of the APU/CPU roadmap;
2015 - 20nm
2016 - 14nm
2017 - 10nm

I was accidentally out of date with the node timeline. There is no reason for the GPU timeline to be any different from the APU/CPU timeline.

 

[DiogoDX]

Maybe a binned, high clock 290? R9 290XTX would be my guess.

AMD would need a hawaii respin like nvidia did with GF110. Make some tweaks, reduce leakage, bump clocks, add a 7Ghz memory and call it 295X.

 

[rtsurfer]

Even on water how much gain would that net? Beat a OC'd 780Ti on air?

A high end 290X can even Today give a 780Ti a run for its money on Air.
A respin would probably surpass the 780Ti everywhere if it brings good enough improvements.

AMD would need a hawaii respin like nvidia did with GF110. Make some tweaks, reduce leakage, bump clocks, add a 7Ghz memory and call it 295X.

I think AMD's memory is already fast enough, they have a wider BUS, they don't need their memory to clock like Nvidia's.

Crappy Elpida RAM on reference cards, is a different issue.

 

[youshotwhointhe]

Seems like Carrizo is a good candidate for this technology. Finally solves the memory bandwidth problems of existing APUs and in mobile they could even potentially forgo any additional system RAM.

I could also see them updating their mobile GPU line up on GF's 28nm process. Beema and Mullins saw significant leakage reduction in the GPU portion after being ported over from TSMC. Add in the power savings from HBM and they could have a real advantage in this segment. Would also allow them to try out some architectural improvements for GCN before moving to 20nm for high end GPUs next year.

 

[UaVaj]

Crappy Elpida RAM on reference cards, is a different issue.

anyone over clock only memory to see if 290x will continue to scale with increased memory bandwidth? save me the attempt.

 

[Gloomy]

anyone over clock only memory to see if 290x will continue to scale with increased memory bandwidth? save me the attempt.

Only tessellation performance improves, typically. If you're not tessellation-bound in the game you're playing you'll see very small (~5%) gains. This is on a stock clock (947MHz) 290, at 1080p and 1440p.

 

[rtsurfer]

Only tessellation performance improves, typically. If you're not tessellation-bound in the game you're playing you'll see very small (~5%) gains. This is on a stock clock (947MHz) 290, at 1080p and 1440p.

anyone over clock only memory to see if 290x will continue to scale with increased memory bandwidth? save me the attempt.

Correct me if I am wrong.
But it would scale only if it was Bandwidth starving in the first place. If it isn't, then you aren't going to see much improvements.

While we are at it, also if somebody is willing to provide some data on how Nvidia cards scale with memory clock in games. I don't have one, & when I did I only tested Valley scaling, which loves any memory OC any way.

 

[Gloomy]

Correct me if I am wrong.
But it would scale only if it was Bandwidth starving in the first place. If it isn't, then you aren't going to see much improvements.

While we are at it, also if somebody is willing to provide some data on how Nvidia cards scale with memory clock in games. I don't have one, & when I did I only tested Valley scaling, which loves any memory OC any way.

Tessellation on AMD cards is always bandwidth-limited at high factors.

 

[NostaSeronx]

Seems like Carrizo is a good candidate for this technology.

HBM is a graphics and networking spec.

Carrizo will be using quad-channel DDR4 which is a system spec.