[Rumor/Speculation] GTX Titan X 12GB vs R9 390X 4GB vs Unknown GM200 GPU

[3DVagabond]

I wanted to check the scaling potential of the last GCN version and compared 285 vs 290 with the help of hardware.fr's benchmark : http://www.hardware.fr/articles/926-1/amd-radeon-r9-285-tonga-sapphire-dual-x-oc-test.html

Their conclusion was that the 285 was 30.2% slower then the 290 in average, and with 30% less CPUs (28 vs 40) this gives us an almost perfect scaling. The 285 consumes also +- 30% less power then a 290.

If we suppose that the 390X is "just" a big Tonga without other optimization and that there's no bottleneck not allowing the optimal usage of all CU's, this would place the 390X more or less on the same level then the TitanX (~30% from 980). It would also be around 350W.

What evidence do we have to support that thinking? While we don't know if it's better or not we already know it's not simply an enlarged Tonga chip. HBM, if nothing else shows that.

 

[therealnickdanger]

What evidence do we have to support that thinking? While we don't know if it's better or not we already know it's not simply an enlarged Tonga chip. HBM, if nothing else shows that.

I think that's part of his point. If all the 390X was was a "big Tonga", it would be near the performance of a Titan X. The fact that it is something "more than" a big Tonga makes it entirely possible that it will outperform Titan X by a discernable margin.

Speculation + speculation = speculation.

 

[Elfear]

I don't believe it is at odds at all. The Titan technology is used in many sectors other than just gaming and is well within normal power levels. The 295X is a specialty card with two GPUs. Upping power levels on single GPUs year over year is not sustainable. Maxwell is precisely the right way to go. Much more efficient and a large improvement over the last generation while staying in a widely deployable power range.

I get what you're saying as the efficiency improvements with Maxwell were very impressive. AMD can't sit on their laurels with regards to efficiency because after a few generations the power requirements would just be ludicrous and everything below the halo card would suffer.

I think what others are trying to point out is that there is a large number of people who don't want to sacrifice performance in the name of efficiency. If the 390X has a 170W TDP but is only 5% faster than a 980, that doesn't get people very excited. If we stick with a 170-200W limit, we'll see smaller performance increments each gen.

I for one would love for Nvidia or AMD to come out with a 300-350W monster with 290X+80-100% type performance. I would much rather have that than continue to go SLI/Crossfire each gen. Multi-GPU trade-offs, while reasonable, get old sometimes.

 

[Subyman]

I get what you're saying as the efficiency improvements with Maxwell were very impressive. AMD can't sit on their laurels with regards to efficiency because after a few generations the power requirements would just be ludicrous and everything below the halo card would suffer.

I think what others are trying to point out is that there is a large number of people who don't want to sacrifice performance in the name of efficiency. If the 390X has a 170W TDP but is only 5% faster than a 980, that doesn't get people very excited. If we stick with a 170-200W limit, we'll see smaller performance increments each gen.

I for one would love for Nvidia or AMD to come out with a 300-350W monster with 290X+80-100% type performance. I would much rather have that than continue to go SLI/Crossfire each gen. Multi-GPU trade-offs, while reasonable, get old sometimes.

I see, but isn't that a one shot deal? Then we'd be back to 20% or so increases unless they made another jump in power. Which is my point since I've been talking long term.

A good point was the long node we are currently in.

 

[ocre]

I think that's part of his point. If all the 390X was was a "big Tonga", it would be near the performance of a Titan X. The fact that it is something "more than" a big Tonga makes it entirely possible that it will outperform Titan X by a discernable margin.

Speculation + speculation = speculation.

This

Its kind of exciting if you look at it that way. Surely the 390x will have cores at least as capable as tonga. Surely there has been some improvements to efficiency. Looking at the core count alone, there really is potential. This is exciting.

But I hate to get hyped up on future cards, I think it is really risky. A great improvement can be seen as a dud if we have expectations that exceed it.

As for the closed loop, give it a break already. We don't have anything to go on so there is nothing really to debate here. I personally believe that the AMD went with a respin and that is why we aren't seeing cards now. This is wild speculations but if AMD is respinning then they expect to end up with a better all around chip. Perhaps this AIO cooler was just an effort to work with what they had. After the respin, they may not even need it. We just don't know anything right now.

Look how long ago that cooler picture leaked. Way back before the beginning of the year. I think AMD was trying to launch as soon as they could. This cooler probably was considered but they didn't launch straight away. I think a respin would have been a hard consideration, you know leaving the market to the gm204 but in the end it might have been the best route they felt they should take. Ere is no way tha AMD would have let nvidia up have all this time with maxwell if they could counter. And I think they really really wanted to.

A lot could have changed between the first leaks of the AIO and when Fiji actually launches. And perhaps there still might be closed loop cooling but we really know nothing about the situation at all.

 

[JDG1980]

As for the closed loop, give it a break already. We don't have anything to go on so there is nothing really to debate here. I personally believe that the AMD went with a respin and that is why we aren't seeing cards now. This is wild speculations but if AMD is respinning then they expect to end up with a better all around chip. Perhaps this AIO cooler was just an effort to work with what they had. After the respin, they may not even need it. We just don't know anything right now.

Look how long ago that cooler picture leaked. Way back before the beginning of the year. I think AMD was trying to launch as soon as they could. This cooler probably was considered but they didn't launch straight away. I think a respin would have been a hard consideration, you know leaving the market to the gm204 but in the end it might have been the best route they felt they should take. Ere is no way tha AMD would have let nvidia up have all this time with maxwell if they could counter. And I think they really really wanted to.

A lot could have changed between the first leaks of the AIO and when Fiji actually launches. And perhaps there still might be closed loop cooling but we really know nothing about the situation at all.

Even if Fiji is no more power-hungry than Hawaii, I think it is still going to ship with a CLC watercooler as the reference solution. AMD got beaten up badly in the reviews of R9 290/290X because of the terrible reference blower (this Anandtech review is typical of the genre). Worst of all, even though later reviews of AIB custom cards were much more favorable, the original reference scores were the ones that get recycled in later comparison charts. These have worse performance (throttling due to excessive heat), worse power consumption (Hawaii uses more power when it's run hot), and of course louder noise by a substantial margin.

For AMD to include a closed loop cooler is a defensive move against bad reviews, nothing more. Their R&D budget is stretched thin as it is, so they can't afford to engineer a good blower like the one on the Titan, and OEMs won't let them pick an open-air solution that dumps the heat into the case. So that leaves all-in-one watercooling as the only remaining option.

 

[Silverforce11]

Even if Fiji is no more power-hungry than Hawaii, I think it is still going to ship with a CLC watercooler as the reference solution. AMD got beaten up badly in the reviews of R9 290/290X because of the terrible reference blower (this Anandtech review is typical of the genre). Worst of all, even though later reviews of AIB custom cards were much more favorable, the original reference scores were the ones that get recycled in later comparison charts. These have worse performance (throttling due to excessive heat), worse power consumption (Hawaii uses more power when it's run hot), and of course louder noise by a substantial margin.

For AMD to include a closed loop cooler is a defensive move against bad reviews, nothing more. Their R&D budget is stretched thin as it is, so they can't afford to engineer a good blower like the one on the Titan, and OEMs won't let them pick an open-air solution that dumps the heat into the case. So that leaves all-in-one watercooling as the only remaining option.

How hard do you think making a vapor chamber heatsink is? All the AIBs manage to do it just fine with their own vapor chamber & heatpipe designs.

Even if I had the 780ti with the Titan cooler, I would put water on it. Blowers are noisier when ambients are high (Australian summers are brutal), and it'll prevent the top tier boost clocks.

What is not to love?

 

[AtenRa]

250-300W is already the very top of the market though. Most "daily driver" GPUs are in the 10-50W range.

To be clear, my point is that going up the power ladder has very, very clear disadvantages and a terminal limit that is reached quickly. Efficiency is all advantage, the hard part is that it is much tougher to make something more efficient rather than simply use more power. When the competition is extracting more efficiency then you must be as well to survive long term.

You can have a 400W GPU to be more efficient (perf/watt) than a 200W GPU.
Example, your 200W GPU may only produce 25 fps at 4K when the 400W GPU could produce 55fps. Im sure nobody will care about the elevated power if they could have that performance at 4K

But even if your 400W GPU doesn't have higher efficiency than your 200W GPU. If the performance of the 200W GPU is inadequate for your requirements (example 50-60fps at 4K) then you dont care if the 400W GPU that will meet your requirements has lower efficiency.

 

[Subyman]

You can have a 400W GPU to be more efficient (perf/watt) than a 200W GPU.
Example, your 200W GPU may only produce 25 fps at 4K when the 400W GPU could produce 55fps. Im sure nobody will care about the elevated power if they could have that performance at 4K

I don't see this having anything to do with what I said. I never said they are mutually exclusive.

But even if your 400W GPU doesn't have higher efficiency than your 200W GPU. If the performance of the 200W GPU is inadequate for your requirements (example 50-60fps at 4K) then you dont care if the 400W GPU that will meet your requirements has lower efficiency.

Again, I'm talking long term as I've stated many times. Sure, jump up to 400W GPUs, but then what? Go to 600W for the next big increase? As you can see, major gains must be made in efficiency to keep moving forward in performance.

 

[3DVagabond]

The rumors are claiming ~50% performance improvement for about the same power usage. Why the panic that efficiency isn't improving?

 

[Subyman]

The rumors are claiming ~50% performance improvement for about the same power usage. Why the panic that efficiency isn't improving?

No panic, here is where it started:

AIO are being used to move to 300W+. I don't see that being sustainable. It may work this gen to edge out the competition, but we can't keep moving up the power ladder to extract more performance.

Just an observation that some found up for discussion.

 

[alcoholbob]

It would be interesting if Nvidia actually launches this at $999.

One reason why it seems unlikely is because eVGA is apparently going forward with the launch of the GTX 980 Kingpin will come out at $800.

A stock Titan X would be a hell of a lot faster than an overclocked GTX 980 Kingpin, you need to hit around ~1900MHz on a GTX 980 just to match a stock Titan X...therefore eVGA must already know the Titan X is not going to be $999.

But who knows? We might be pleasantly surprised.

 

[ocre]

So....

There are way to many fake leaks of the 390x going around.

One thing is sure now.......

AMD will know exactly where the titanX performance lands and they will have plenty of time to tweak out the 390x performance.

I am convinced that this leak is a total fake but I do think AMD has a chance to do something here. Would it be so bad if e 390x was faster than the titanX, cheaper than the Titan x, but used +300 watts of power? This AIO cooler gives them an opportunity to clock this chip as high as they need to. That's what I am thinking anyway. They can clock it high and may be able to beat the Titan X. If e power consumption is high, will it matter much if the card stays cool and is quiet?

Being real now, the AIO cooler just won't be for everyone. But if AMD offers a more powerful beast that is cheaper than nvidia's, it will be an option for many who buy in the top brackets.

I think that having to stick with 28nm for another round is terrible but both nvidia and AMD are forging ahead now. AMD doesn't have to win in every metric, just don't think they have to. This leak seems to good to be true and I think it is risky for people to believe in it. But if it turns out the case, all the better when the GPU blast past expectations.

My true concerns are a little off topic. They are of glofo silicon. Is the 28nm really up to the task of such a large chip. Have they passed tsmc grade 28nm?
Another confusing thing to me is how everyone's 20nm HP can be broken. If tsmc fails, okay but now it seems glofo and Samsung.....all of them. No one was able to produce 2onm HP worth a crap. Then there is this terrible reality that intel has ran into issues with their nodes with their high performance chips. Their 32nm was the last major leap. 22nm didn't move the bar at all. The improvements have come from the design of the chip, haswell architecture. Now look at broadwell, the shrink is only able to move them forward in the low end. Sky lake will be a new design. Its really scary to see this. Intel is way ahead with their nodes and it is not looking good at all for the high performance chips. These smaller nodes are improvements but they are mainly driving down TDP without the benefit of raising the bar in ultimate performance. Nothing like the days of the past.
There is this wall. It is real. Intel forged ahead and I think they are showing us what is to come. Gpus have a rocky road ahead now. I am fearful 16nm will bring little when it comes to absolute performance. The path forward is really not node shrinks at all. Its architectural changes only. This is serious and means that................
Gains will be smaller from generation to generation. I hate to think that the GPU will become just as dull as the x86 CPUs have. I am really afraid that this is quickly becoming our future. And I can't stand the thought.

 

[Cloudfire777]

R9 390X
HBM
4096 shaders
http://videocardz.com/55124/amd-radeon-r9-390x-wce-could-this-be-real

 

[ShintaiDK]

Someone being creative with Photoshop? Why would AMD write such rubbish as DX12_Tier 3, specially without specifying what feature they refer to?

Not to mention the slide is dated in the future.

 

[HurleyBird]

These "slide leaks" seem to be legit about 2/3rds of the time (and a lot of the fakes are obvious), so I'm hoping that this is one of those times due to 8GB and the implication of architectural advances beyond GCN 1.2.

 

[Silverforce11]

Up to 8GB HBM, damn, our wish got granted!

Note the "Special enthusiast water-cooled edition". Special indeed.

 

[StereoPixel]

event 16 March?
"next iteration..." is GCN 1.3 or 2.0

 

[Cloudfire777]

Always questions about everything thats not official

"Tier3 implementation" etc could be terms AMD will explain while showing the slide.

March 16th is a weird date. We know Nvidia will show Titan X and reviews will be up on March 17th. Nvidia will be at the GPU conference and also show Quadro M6000/GM200 so it could mean AMD is trying to one up Nvidia and show their card first. Or the slide is fake.

Who knows

"DirectX12_Tier3" seems a bit sloppy job for an official slide.
"Up to 4096 shaders" and "Up to 8GB HBM" seems extremely weird for 1 card. If it was R9 390 and R9 390X I would understand it, but just 390X? Weird

 

[zornem7]

R9 390X
HBM
4096 shaders
http://videocardz.com/55124/amd-radeon-r9-390x-wce-could-this-be-real

AMD HBM 4GB and 8GB 390x leaks were true! So the next couple days are going to be what someone else said, Mayweather vs Pacman fight of the decade.

 

[StereoPixel]

"Up to 8GB HBM" seems extremely weird for 1 card. If it was R9 390 and R9 390X I would understand it, but just 390X? Weird

390X 4 GB - 4096 bit - 512 GB/s --- $600-700
390X WCE 8 GB - 8192 bit - 1 TB/s --- is expensive...

I think so...
There were two different cards from the same series: 290X 4GB and 290X 8GB

 

[Cloudfire777]

390X 4 GB - 4096 bit - 512 GB/s --- $600-700
390X WCE 8 GB - 8192 bit - 1 TB/s --- is expensive...

I think so...
There were two different cards from the same series: 290X 4GB and 290X 8GB

You make a good point about the HBM memory being expensive and all. So that makes sense I guess now that you mention it

But the same GPU having 4096 shaders and less than that. Wouldnt that make it a different GPU?

 

[Enigmoid]

Why would you label a card which is clearly defined by product name with "up to xxxx shaders"?

 

[Cloudfire777]

Same "card" but with such big difference in bandwidth and core count.

It smells fishy

 

[SPBHM]

Same "card" but with such big difference in bandwidth and core count.

It smells fishy

unless it's single vs dual GPU, like 290x vs 295x2