[WCCF] AMD Radeon R9 390X Pictured

[96Firebird]

Tonga is 366mm2 on TSMC. Have you not heard of Gloflo? All of the old die size calculations need to be revamped.

Which is an unknown. So to gather any information from 290X vs 780 Ti is pointless.

In any case do you buy your cards by the die size.

No, which is why I said...

I'm confused on how this matters in the slightest...

 

[maddie]

Which is an unknown. So to gather any information from 290X vs 780 Ti is pointless.



No, which is why I said...

Maybe to you. Surely someone with your knowledge of the industry should know this.

Kaveri = 2.41 billion
Carrizo = 3.1 billion

These both have the same size die so what exactly happened here?

 

[JDG1980]

Is Tonga latest gen? Pretty even with the 960, no? I realize Tonga is a cut down die, but I imagine a full die isn't that much faster than a 960.

Tonga = 366mm², 960 = 228mm².

You can cherry pick anything to fit an argument though...

A couple things to consider:

• Tonga is much bigger than you'd expect given its performance. I strongly suspect there is a lot of "junk DNA" in there, which could be trimmed for a respin. It's rather unusual among AMD chips, and I think it was rushed to market in a far different state than intended.
• Tonga has a 256-bit memory controller (there have even been persistent rumors that Tahiti's 384-bit memory controller is still intact on-die, though that has never been proven). In contrast, GM206 only has a 128-bit bus. That makes a big difference in die size. This shouldn't be an issue with Fiji, since by all accounts HBM allows the memory controller on-die to be downsized.
• The R9 285 is not only missing 256 shaders (which we know are on-die because they are enabled in the Retina iMac's R9 M295X), but it's also running at a far lower clock speed than the GM206 in the GTX 960. The R9 285 has a clock speed of only 918 MHz, while the GTX 960 boosts to 1178 MHz stock (and some AIB versions go well over 1300 MHz). If we assume that the shaders scale linearly and that clock speed scales at 50% (i.e. 10% higher clock speeds offers 5% better performance), then a fully enabled Tonga running at the same boost clock speed as GTX 960 would beat it in performance by over 30%. (28% higher clock speeds -> 14% performance boost, and 14% more shaders. 1.14 x 1.14 ~= 1.3).

 

[LTC8K6]

Maybe to you. Surely someone with your knowledge of the industry should know this.

Kaveri = 2.41 billion
Carrizo = 3.1 billion

These both have the same size die so what exactly happened here?

Thin is finally in?

 

[maddie]

Thin is finally in?

At least some of us are having fun.

Have to sleep now.

 

[JDG1980]

If AMD had a lot of funds for R&D I could see this as possible. To spend that money and only sell an elite gaming card in their present situation would be madness.

I'm not saying there will be no FirePro Fiji. Rather, I'm suggesting it might be possible that the FirePro Fiji will not be the Double Precision flagship, and instead that role might be taken by Grenada (re-engineered Hawaii). Just as an example, you might have the "FirePro W9200" with 16GB of VRAM for Double Precision users, and a specially named or labeled FirePro with Fiji aimed at Single Precision applications with modest RAM requirements.

Note that this is not much different than what Nvidia already did with Maxwell. All the Maxwell cards have Double Precision performance severely cut down; Nvidia's current Double Precision flagship is a respun big Kepler (GK210). Since Hawaii is already competitive with Kepler in DP compute applications, AMD doesn't need a massive boost there to stay in the game. However, if Apple is going to refresh their Mac Pro lineup this year, then AMD does need something better than Tahiti for the top-tier Mac Pro card with DP capability. Existing Hawaii is too hot and power-hungry, and Fiji would represent an undesirable regression in RAM capacity.

AMD obviously felt it was important to design a video card with HBM as a test bed project. For whatever reason, it looks like the card wound up being limited to 4GB in practice, even if higher configurations might be possible in theory. If AMD spent a lot of die space on Double Precision computing, they'd wind up with a "jack-of-all-trades" card, which would be hamstrung for GPGPU by its limited RAM capacity, and sacrifice performance in gaming and Single Precision computing because of the die space spent on DP math units. In contrast, whipping the Titan X by ~20% in gaming would get everyone's attention, and win AMD some much-needed mindshare. That kind of lead can't be ignored or dismissed.

 

[Cloudfire777]

Maybe to you. Surely someone with your knowledge of the industry should know this.

Kaveri = 2.41 billion
Carrizo = 3.1 billion

These both have the same size die so what exactly happened here?

Using a denser transistor design is the only way to get the 28nm SHP process benefits. Atleast on the APU side. I would guess with GPUs as well.

Which is 20% power reduction with the same clocks or same power consumption but with 10% higher clocks.

AMD could have gone this route with 390X which is maybe why they changed codename from Hawaii to Grenada although the specs otherwise is the same.

 

[Silverforce11]

Using a denser transistor design is the only way to get the 28nm SHP process benefits.

Which is 20% power reduction with the same clocks or same power consumption but with 10% higher clocks.

Yes, and power gating design as well as voltage plane in Carrizo improves efficiency further. Definitely lots of tools available for AMD to drop power use.

The other factor is performance, if they don't need to push vcore/clocks high, power use will drop a lot, similar to what users see with R290X on custom power curve, ~850mhz core ~160W is very power efficient.

So don't be hasty to discount, especially relying on 2x 8pin and assuming 375W! Certainly after AMD's Joe Macri has said publicly that "Fiji shouldn't use more power than R290X and in fact, it could end up using less".

 

[DownTheSky]

A couple things to consider:

• Tonga has a 256-bit memory controller (there have even been persistent rumors that Tahiti's 384-bit memory controller is still intact on-die, though that has never been proven).

It's been proven actually. Tonga has 6x64 bit mem controllers.

 

[soccerballtux]

Using a denser transistor design is the only way to get the 28nm SHP process benefits. Atleast on the APU side. I would guess with GPUs as well.

Which is 20% power reduction with the same clocks or same power consumption but with 10% higher clocks.

AMD could have gone this route with 390X which is maybe why they changed codename from Hawaii to Grenada although the specs otherwise is the same.

I'd take 20% and make the die 20% larger.

what's the story on GloFo anyways? AMD still required to make their CPUs with them? I thought nobody liked GloFo and nobody was using them. TBH this 'Fiji is using GloFo' seems odd to me.

I do want them to do well financially but. I just have a bad feeling about them. Do they ever get SoC orders???

 

[Silverforce11]

I'd take 20% and make the die 20% larger.

If they gone that route, they can already cram in 25-30% more transistors at the same die space. It's certainly no "rebadge" if its from GloFo.

 

[RussianSensation]

JaysTwoCents

Titan X and 980 Ti reference @ 1.4Ghz Boost = 86*C
EVGA Hybrid 980Ti @ 1.484Ghz Boost = 50*C max

https://www.youtube.com/watch?v=qtRqmzRMar8

I can't imagine going with air cooled cards for SLI/CF ever again with these results if the premiums for AIO CLC aren't too large and reliability of the pumps is good enough to last 5 years (i.e., which means we can resell the cards and the 2nd owner won't have to worry that they will break from wear).

As far as I see it, I predict the AIO CLC will slowly tart to take over air cooling in the enthusiast community with Pascal / Arctic Islands 14nm/16nm, much like AIO CLC today completely dominates the high-end air cooling CPU space. Once a lot of GPU enthusiasts try these out, it's going to be harder to go back to conventional air cooling. AIO CLC on 2 flagship 300W cards takes all the risk/hassle of building a custom water-loop away from the end user and gets one probably 90% of the performance benefits for a fraction of the cost, while exhausting most of the heat out of your case. Cleaning the radiator will be as easy as taking the fan off and using compressed air.

Air cooling has no chance to compete with that because you simply can't find any air cooler that will be able to maintain 50*C on a 300W max overclocked 980Ti and still exhaust most of the heat out of your case. Looks like AMD is going to revolutionary the GPU landscape with both HBM1 and lead the way towards reference AIO CLCs!

 

[Silverforce11]

It's not just the temps, its the noise that comes with blowers once you exceed 250W. If it sounds anywhere close to the reference R290X, its intolerable and the reason I water mod mine.

I can run my rad fans & pump on a silent profile and temps go to 65C, still well below the best air coolers, especially OC.

 

[JDG1980]

As far as I see it, I predict the AIO CLC will slowly tart to take over air cooling in the enthusiast community with Pascal / Arctic Islands 14nm/16nm, much like AIO CLC today completely dominates the high-end air cooling CPU space. Once a lot of GPU enthusiasts try these out, there will have trouble going back to conventional air. AIO CLC on 2 flagship 300W cards takes all the risk/hassle of building a custom water-loop away from the end user and gets one probably 90% of the benefits for a fraction of the cost.

OEMs want blowers, therefore blowers are here to stay in reference cards, except for the ultra high-end (and I suspect only AMD will switch to reference CLC).

The FinFET process will result in significantly better perf/watt, so there will be no particular reason for the use of CLCs, especially not with the first generation of FinFET GPUs that will probably be limited to ~350 sq. mm. at most. Nvidia's reference blower works very well on the GTX 980; in fact, it's one of the quietest coolers tested. This is for a card that peaks at ~184W in gaming (it goes to 190W in FurMark, but they didn't test the fan under that load, but I doubt the extra 6W makes any real difference). The Nvidia blower starts to run into trouble on the 250W TDP cards; TPU criticizes it for being too noisy (42 dBA). Note, however, that the R9 295 X2 posts almost the same noise level (41 dBA).

I disagree with your assertion that CLCs dominate the high-end air cooling CPU space. The best air coolers are far superior products, providing equal performance with far fewer things to go wrong. I'm not at all sure I trust these crappy little pumps to last for the long term. If you have a "wind tunnel" case like the Corsair Air 540 or the Silverstone FT02/FT04/FT05, then running a Thermalright Macho Zero or Silverstone Heligon HE02 without a fan (relying on the case fan's airflow) is going to be a much better solution.

 

[boozzer]

What does the 780 Ti have to do with what he said...?

In both instances, Nvidia was in the lead with no competition, and still released a lower priced, equally performing card 3 months later.

That is all there is to take away from that.

But we still have people who say Nvidia rushed the 980 Ti's release because they're scared? History proves otherwise...

I hope you are doing it on purpose.

 

[Cloudfire777]

Yes, and power gating design as well as voltage plane in Carrizo improves efficiency further. Definitely lots of tools available for AMD to drop power use.

The other factor is performance, if they don't need to push vcore/clocks high, power use will drop a lot, similar to what users see with R290X on custom power curve, ~850mhz core ~160W is very power efficient.

So don't be hasty to discount, especially relying on 2x 8pin and assuming 375W! Certainly after AMD's Joe Macri has said publicly that "Fiji shouldn't use more power than R290X and in fact, it could end up using less".

Thats not the only hint. I guessed you missed this:
PSU of 700W is recommended for Fury X

R9 290X have a PSU requirement of 750W.
Guess we will have to wait and see.

I'd take 20% and make the die 20% larger.

what's the story on GloFo anyways? AMD still required to make their CPUs with them? I thought nobody liked GloFo and nobody was using them. TBH this 'Fiji is using GloFo' seems odd to me.

I do want them to do well financially but. I just have a bad feeling about them. Do they ever get SoC orders???

390X most likely still have the 2816 shaders you find on Hawaii. They should be identical, just that 390X probably is more packed and have other features.

AMD announced they have manufactured chips worth $150M in Q1 in GloFo but will have manufactured a total of over $1000M in 2016 in GloFo. Its weird that all of that comes from APUs, but I dont have the details about console APUs/Carrizo etc how much they cost.

 

[Tuna-Fish]

Maybe to you. Surely someone with your knowledge of the industry should know this.

Kaveri = 2.41 billion
Carrizo = 3.1 billion

These both have the same size die so what exactly happened here?

High density synthesis libraries. They sacrificed quite a lot of transistor and interconnect performance to pack transistors tighter. This allowed to claw back a lot of the lost performance (maybe enough to end up with an actual performance gain?) because everything is closer together so interconnect distances you need to cover are shorter.

In the olden days where product clock speeds were mainly limited by just how fast they could be made to switch, this would have been a clear loss. However, these days product clock speeds are mainly limited by power budgets, and long-distance communications are a huge part power use. Reducing distances might save enough power they can now use elsewhere to be a net gain.

 

[sandorski]

I remember the days when all the cooling available on a Graphics chip was nothing at all. Then there were just heatsinks. When the Voodoo 5 5500 I have in my closet with itty bitty fans on the dual GPU's was released it was considered scandalous. Then heatsinks began using Heatpipes and even larger fans. Now AIO coolers are being used and I suspect that they will become more common.

Those aghast at AIO should get used to it as it's likely to become standard, at least on the high end.

 

[Silverforce11]

I remember the days when all the cooling available on a Graphics chip was nothing at all. Then there were just heatsinks. When the Voodoo 5 5500 I have in my closet with itty bitty fans on the dual GPU's was released it was considered scandalous. Then heatsinks began using Heatpipes and even larger fans. Now AIO coolers are being used and I suspect that they will become more common.

Those aghast at AIO should get used to it as it's likely to become standard, at least on the high end.

I remember my first GPU mod, heatsink on the (normally bare) Rendition Verite chip, get a small OC possible to achieve 30 fps on Quake GL, damn it felt smooth. It looked a lot nicer than the Voodoo 1 and it's trash texture filtering and lack of hardware AA.

People complaining about warrantied hassle free water cooling for enthusiast GPUs with such high TDPs is just not enthusiast enough.

 

[boozzer]

I remember the days when all the cooling available on a Graphics chip was nothing at all. Then there were just heatsinks. When the Voodoo 5 5500 I have in my closet with itty bitty fans on the dual GPU's was released it was considered scandalous. Then heatsinks began using Heatpipes and even larger fans. Now AIO coolers are being used and I suspect that they will become more common.

Those aghast at AIO should get used to it as it's likely to become standard, at least on the high end.

it is only a talking point because amd came out with it first. the reaction is 100% fake

 

[JDG1980]

AMD announced they have manufactured chips worth $150M in Q1 in GloFo but will have manufactured a total of over $1000M in 2016 in GloFo. Its weird that all of that comes from APUs, but I dont have the details about console APUs/Carrizo etc how much they cost.

AMD had a very bloated channel inventory in Q1 2015, so it's not surprising that their production for that quarter was low. One site claims that they stopped channel shipments entirely for that quarter.

The desktop APUs are not particularly competitive products and can't be selling well. Carrizo may do better, but we must assume that the bulk of GloFo APU sales so far have been from the consoles (PS4/XB1). Porting over discrete desktop GPUs to GloFo 28nm SHP should help AMD fulfill their "Take-or-Pay" wafer supply agreement.

 

[Tuna-Fish]

what's the story on GloFo anyways? AMD still required to make their CPUs with them? I thought nobody liked GloFo and nobody was using them.

The GloFo approach to 28nm was more ambitious than the TSMC approach. GloFo transistors should perform better (drive current is lower but device capacitance is much lower which more than compensates) and leak less current. However, TSMC bet that their approach would be much easier and faster to get to production and slightly easier to design for so they'd get most of the major design wins. In hindsight, TSMC could not have been more correct -- GloFo 28nm only really reached good production status at the very tail end of the period where 28nm is the top node, and TSMC, as usual, extracted nearly all of the profits from foundry business.

One way to think of it is that the GloFo 28nm is like a quarter or a sixth node improvement over TSMC 28nm, but it's only coming online now.

A lot of people here seem to be predicting that most of the AMD GPU line will switch over to GloFo for 300-series. I frankly don't see this happening, as making all those masks can't possibly be worth it. We might see a chip or two do the switch.

 

[RussianSensation]

OEMs want blowers, therefore blowers are here to stay in reference cards, except for the ultra high-end (and I suspect only AMD will switch to reference CLC).

I did say enthusiast community. OEMs do not even fall into that unless you are talking about boutique system builders. I never said that reference blowers will disappear. I am saying that as AIO CLCs started to slowly take over the high-end CPU building scene, AIO CLC for GPUs could also gain popularity since they are even more effective at solving the noise levels, temperatures and exhausting that heat out of a case on a GPU. It becomes a near optimal solution for SLI/CF without going full custom water-loop.

Also, if we dive back in time, in the past horizontally positioned CPU heatsinks were popular and then they were obsoleted by tower design heatsinks. Today downward blowing CPU heatsinks are only good for systems where you need a low profile CPU heatsink/fan combo. Otherwise, tower CPU heatsink designs easily outperform them. The industry moves forward to better technologies over time.

The FinFET process will result in significantly better perf/watt, so there will be no particular reason for the use of CLCs, especially not with the first generation of FinFET GPUs that will probably be limited to ~350 sq. mm. at most.

I never said anything about next gen mid-range products using AIO CLCs. What you just described are exactly those products. I said specifically flagship high-end parts, 250-300W TDP.

I disagree with your assertion that CLCs dominate the high-end air cooling CPU space.

Not from a performance point of view but what enthusiasts have been buying. Coolers like Thermalright True Spirit 140/Silver Arrow Extreme, Noctua NH-D14/15, Phanteks PH-TC14 are not very popular. 5-10 years ago, these style coolers were all the rage. If you go over to OCN or even our site, a lot of PC gamers have switched over to AIO CLCs for their high-end cooling. For example for the X99 platform, a lot of these coolers don't even work well because you have DIMM memory on both sides of the socket which makes it very difficult since it now requires specifically looking for low profile memory heatsinks.

The best air coolers are far superior products, providing equal performance with far fewer things to go wrong.

No air cooler right now for max overclocked CPUs that can beat the best AIO CLC kits like Swiftech H-220X or H-240X or Corsair H100i/110i. Noctua NH-D15 or Phanteks PH-TC14 can't do it. The best air coolers can beat H50/60/80/90 style kits but once you go 240-280mm or 360mm rads, they will lose. How may guys on OCN rocking 980Ti SLI will have high-end air coolers vs. AIO CLC kits today? A lot less compared to 5 years ago.

I think you aren't paying attention enough to where the enthusiast PC landscape is shifting.

Cases like NZXT 440, NZXT Noctis 450 and the new Fractal Design S are made specifically to cater to this new generation of hardware building where everything is shifting to AIO CLC and/or combination of customer water loops in the high-end enthusiast space. 5 years ago it would be unthinkable for a card like 980 to have AIO CLC and today EVGA sells you that card with nearly 1.4Ghz overclock and it operates at 53*C wen max overclocked to 1.6Ghz.

I'm not at all sure I trust these crappy little pumps to last for the long term. If you have a "wind tunnel" case like the Corsair Air 540 or the Silverstone FT02/FT04/FT05, then running a Thermalright Macho Zero or Silverstone Heligon HE02 without a fan (relying on the case fan's airflow) is going to be a much better solution.

This is not possible when we are talking about an enthusiast level rig. How are you going to be able to do that with an overclocked 5820-5960X or i7 4790K? I think you haven't been paying much attention to what's really happened in the CPU overclocking scene. I myself like high-end air coolers as much as you seem to do but that doesn't change market trends.

AIO CLC benefits GPUs more than CPUs, which means it's a good bet that it might take off in this area as well. It's just a matter of NV adopting it. The point you make about pumps failing is legitimate which is why I am a big advocate for all GPU makers to start offering 5-year-warranties, if not standard, then maybe for a $25-35 fee per a flagship $500-700 card.

 

[beginner99]

I have trouble with accepting that AMD will make a flagship possibly as fast as 295X2, and only put 4GB HBM maximum. It makes no difference if they can optimize the 4GB so there is no appreciable performance shortfall. They must know what Nvidia marketing will do in that situation.

I also see AMD as having a huge window of opportunity in professional/scientific markets as Nvidia dropped DP performance in Maxwell. 4GB is a huge problem in those markets.

They must have higher than 4GB models.

Or Fiji is like GM200 and has way less DP performance. Then with the huge die one could easily see how it could beat 980TI by a huge margin. That's my hope. Could not care less about DP.

We will see. It depends if more than 4 GB HBM1 is possible at all at acceptable cost. 8 GB could be limited to FirePro simply due to cost.

 

[RussianSensation]

Which is an unknown. So to gather any information from 290X vs 780 Ti is pointless.
...

Keep telling yourself that. AMD was able to create a chip that's both better at compute and gaming with a die size 78% of NV's, while managing to pack more VRAM (4GB vs. 3GB) but barely using more power than a 780Ti.

Historically speaking AMD has been able to compete with NV using much smaller die sizes. NV's performance advantage on the high end was never commensurate with the gargantuan increase in die size they had vs. AMD's smaller chips. NV would barely net 15-18% more performance despite having 30-50% larger die sizes. Today a 980 is < 10% faster than an R9 290X at 4K despite the latter being just 10% larger. But 980 is a completely new architecture and generation. That means if AMD was able to create a 600mm2 3rd generation GCN chip with HBM, it's a wide chip - meaning it has a lot more shaders, textures and geometry performance compared to the R9 290X. Right now 980 mainly looks so good because 290X was designed so long ago as a competitor to 780Ti not to a 980. AMD obviously has faced delays with Fiji XT but 290X was never designed to be competing with Maxwell's 980 cards. The fact that it isn't that far behind is remarkable if anything.

In case you need a refresher, a 438mm2 R9 290X Hawaii vs. 352mm2 original 925mhz Tahiti gave AMD a 49% increase in GPU power at 4K (GPU limited resolution gives us a good idea of the true capabilities of Hawaii chip's resources).

Digging deeper, R9 290X vs. HD7970 is:

37.5% more shaders
37.5% more TMUs
100% more ROPs
50% wider memory controller
4X the amount of Asynchronous Compute Engines (ACEs increased from 2 to 8)
True Audio DSP block

^^^ AMD accomplished all of that with an increase in die size of only 24.4%. That's nothing short of incredible.

NV needed to increase the die size 91% to move from 294mm2 680 to 561mm2 GTX780Ti to hit their flagship GPU performance targets. And yet, 780Ti doesn't even beat the 290X today at high resolution gaming. Even though hardly anyone today would call Big Daddy Kepler an engineering dud/failure but compared to Hawaii it really is now that the dust has settled. How the hell did NV manage to create a 561mm2 chip that loses both in compute and gaming to a 438mm2 chip, while barely using less power? That's straight up engineering failure right there. Kepler was a marketing success above all! Costs more to manufacture a 561mm2 die than a 438mm2 die but the chip is slower. In engineering, I am pretty sure that would be called a dud. D:

A hypothetical 575mm2 Fiji XT would be 31.2% larger than Hawaii XT. AMD doesn't even need to change the amount of ROPs from 64 because they can increase pixel fill-rate throughput 50% alone using Tonga ROPs with 40% memory bandwidth compression.

That means 45% increases in TMUs and shaders are not out of the question. Then we have doubling of geometry performance courtesy of Tonga's architecture as a foundation. If the actual die size is 600mm2, that gives the chip even more potential. The main counter-argument against this is to reduce leakage, AMD reduced transistor density significantly on a 575-600mm2 chip compared to 438mm2 R9 290X. That would explain why a 575-600mm2 chip isn't going to be a 5632 shader part or anything like that.

Unless AMD messed up badly, 45% increase in performance over 290X should be guaranteed at 4K, which already puts Fiji XT faster than the 980Ti. This includes 0 other tricks AMD might have gained with GCN 3 architecture (or 1.3 if we want to call it that).

The biggest thing holding back Fiji XT are probably going to be drivers and power usage. If it wasn't for perf/watt marketing movement, AMD could have clocked the card another 15% more while actually using 375-400W of power since the AIO CLC would be more than capable of handling it. Of course then NV marketing would have swooped in and ripped the card apart for being a perf/watt failure. Hopefully that means there is 15-20% overclocking headroom left even at the expense of 150W of extra power usage.

Or Fiji is like GM200 and has way less DP performance. Then with the huge die one could easily see how it could beat 980TI by a huge margin. That's my hope. Could not care less about DP.

That's true. In an ideal world, if AMD ditched DP like GM200 and made a 600mm2 HBM1 chip with AIO CLC, then chances are Titan X would lose the performance crown.