AMD Fury X Postmortem: What Went Wrong?

[Mako88]

Didn't Intel just delay 10nm yet again?

Leaked partner roadmaps indicate Zen is supposed to debut around the middle of 2016. If they meet that timeline, they should be able to reclaim at least some market share. If things start dragging out past October 2016 and Zen is still nowhere to be seen, that's when to start worrying.

Raven Ridge has the potential to be AMD's next blockbuster product. Intel will probably still have a better x86 architecture, and Nvidia may still have a better GPU. But neither company can put a decent CPU *and* a decent GPU together on one chip, using shared HBM2 to unify them. Intel's graphics are still way too far behind, and Nvidia lacks an x86 license. 2017 could see AMD bring out a mid-range gaming PC on a chip, something only they can provide.

One thing that AMD benefits from in the CPU market is lowered expectations. In graphics, AMD is expected to be fully competitive with Nvidia, and fans are very disappointed when it falls short. On the other hand, Bulldozer and its progeny were such epic failures that if AMD can match the single-thread performance of a 2011 Sandy Bridge chip, it will be a major leap ahead and enough to satisfy a lot of buyers.

You're missing the point, we've seen this all before. Every AMD architecture preview gets this "OH BOY OH BOY OH BOY I CANT WAIT" treatment...

...and every single time for most of the last decade by the time it arrives, long delayed, it's already blown away by the Intel price point competitor.

Come on man...this isn't fanboy redacted, this is reality. AMD is on the ropes, and if they can't figure out a way to get off them, it's a KO. Game over. Done. We lose.

And then the ridiculously high GPU prices today will look "low" and "a bargain" compared to what they'll be 2-3 years from now without any competition.




No profanity in tech.


esquared
Anandtech Forum Director

 

[crisium]

Do we know that for sure, that the air version is cut down?

We don't know for sure. But 3584 shader rumour has been posted a lot, and it's perfectly in line with 7950 vs 7970. I hope it's 4096 though, because that will be a good card then.

 

[iiiankiii]

Skimping on geometry and tessellation performance was a choice that AMD made when developing GCN in general and Fury in particular. In retrospect, it was a poor choice. Yes, if they'd skimped somewhere else, then Nvidia might have targeted that instead. So they need a good all-around architecture without any major weaknesses.

That's why I said it was tough for AMD. Building a GPU or anything really, is a game of compromises. There is only so much die space, power budget, R&D, time, etc. that will hinder certain feature sets. You can't put everything in there. Heck, compute on Nvidia's card is extremely poor compared to AMDs. If AMD had the money and resource, they would have made sure feature sets would take advantage of those compute superiority. But, they can't.

Instead, Nvidia is dictating the playing field. They get to choose what hardware feature to extract from since they control the software (GameWorks). Since Nvidia is strong on tesellation, they ran with that. But, imagine if next gen, AMD decided to focus on tesellation and cut down a bit on compute (just like Maxwell). Now, AMD is up to par in tesellation but at the cost of some compute unit. But wait, Nvidia decided they need more compute unit at the cost of some tesellation units. Now, Nvidia can just modify GameWorks to use less tesellation and focus heavily on compute. Guess what? AMD is back to square one. Now, even with better tesellation, it doesn't matter. It's all about compute, now. So, in the end, AMD loses again because they don't control the software stack. They can't pick which feature sets to exploit. They're at the mercy of Nvidia. That's the power of GameWorks.

 

[Silverforce11]

The scariest part is I think during the initial unveiling comments were made by AMD that the air version would likely throttle and not perform as well as the fury x. I will look for the source.

Found it:

http://www.tomshardware.com/news/amd-fury-x-fiji-preview,29400.html

This time they won't be able to put their crap reference blower on it at least.

Custom models don't throttle.

 

[AnandThenMan]

Found it:

http://www.tomshardware.com/news/amd-fury-x-fiji-preview,29400.html

That to me is about overclocking, With Fury X it will be nigh impossible to get the card to throttle due to heat, air cooled of course that is the norm if you really push the GPU.

 

[SolMiester]

Didn't Intel just delay 10nm yet again?

Leaked partner roadmaps indicate Zen is supposed to debut around the middle of 2016. If they meet that timeline, they should be able to reclaim at least some market share. If things start dragging out past October 2016 and Zen is still nowhere to be seen, that's when to start worrying.

Raven Ridge has the potential to be AMD's next blockbuster product. Intel will probably still have a better x86 architecture, and Nvidia may still have a better GPU. But neither company can put a decent CPU *and* a decent GPU together on one chip, using shared HBM2 to unify them. Intel's graphics are still way too far behind, and Nvidia lacks an x86 license. 2017 could see AMD bring out a mid-range gaming PC on a chip, something only they can provide.

One thing that AMD benefits from in the CPU market is lowered expectations. In graphics, AMD is expected to be fully competitive with Nvidia, and fans are very disappointed when it falls short. On the other hand, Bulldozer and its progeny were such epic failures that if AMD can match the single-thread performance of a 2011 Sandy Bridge chip, it will be a major leap ahead and enough to satisfy a lot of buyers.

Intel has just killed AMDs APU advantage, you know that right? Iris pro kills AMDs best APU in both CPU and GPU benchmarks.

 

[.vodka]

390x sucks down nearly twice the power as the 980, is 10% larger die, has twice the bus width, over a billion more transistors, and has very little OC headroom vs. 980's headroom plethora, and can only manage to almost tie the 980 when neither are overclocked. The 390x is the fat fifth year senior being outclassed by the sophomore half his size.

If you call that on par, have another drink. Nvidia could go to town on 980 pricing, still clear some decent margins, and absolutely nullify any profit AMD would be able to make if forced to price match.


AMD is being clearly out engineered now and the only thing they can do to compete is price themselves out of profit.

Yes, but in spite of all that, Hawaii manages to be competitive after some polishing. Overpriced as it is, the 390x is what the 290x should've been from day one, and it'll come down in price at some point. Not a bad buy, I'd say. For 1080p it's an excellent card. Single TX/980Ti and FuryX fall short at 4k and are overkill for 1080p, anyway. 1440p on the other hand is better suited for them.

Air cooled Fiji could very well obsolete the 390x, we have to see the rest of the lineup.


Which brings us to my point...

It's better for AMD to save their resources for a full new lineup sometime in the future making use of HBM2 than to waste money on 28nm at this point. Fiji is clearly a beta product if you will, to learn the intricacies of implementing HBM after codesigning it with Hynix, and removes one potential problem when making the jump to 14/16nm next year and maybe a new architecture (the 400 series, arctic islands, are rumored to be new architecture IIRC). Something went wrong somewhere, we have 45% more SPs, a similar amount of extra memory bandwidth, tons of extra texturing power, and sometimes the difference over the 290x is less than that. Extra graphics hardware should more or less scale linearly... There's a bottleneck somewhere. Yet the efficiency is there when you go 4k, the TX loses its lead at lower resolutions or outright loses by a few %.

Meanwhile nV with Pascal has to tackle all three fronts at once again, 14/16nm + HBM2 + new architecture. I hope they do better than Fermi. If they don't, it would be an excellent chance for AMD to strike.

All in all I say bring it on! We need competition.



Now for what's left of 28nm cards, we have to see how W10+DX12 changes the landscape. AMD has some interesting features and strengths here (async shaders, the ACEs, GCN 1.1 FL12.0 baseline in the consoles, etc) while nV has theirs (FL12.1 features). Interesting times ahead.

We'll see then who was the most forward looking when designing their architectures when DX12 games start coming out.

 

[sandorski]

It's neither a failure nor is AMD behind Nvidia tech wise. It's just not a clear winner.

 

[utahraptor]

It's neither a failure nor is AMD behind Nvidia tech wise. It's just not a clear winner.

This would be my opinion if it had launched the same month as the 980 or even the 980ti.

Instead they have almost caught up to yesterday.

 

[sandorski]

This would be my opinion if it had launched the same month as the 980 or even the 980ti.

Instead they have almost caught up to yesterday.

It did launch the same month as 980Ti.

 

[Mako88]

This would be my opinion if it had launched the same month as the 980 or even the 980ti.

Instead they have almost caught up to yesterday.

Or at a $599 price point for the AIO version FX, $499 price point for the air version FX.

It seems trivial, $50 less, but the value proposition would have been screaming at that level, $1200 for Crossfired up Fury Xs on water is dead sexy compared to the $1500 it costs to mimic that setup via EVGA's hybrid 980 Ti.

 

[utahraptor]

It did launch the same month as 980Ti.

Wow I just checked and you are right. It feels like 3 months!

 

[Mako88]

Wow I just checked and you are right. It feels like 3 months!

That's Titan X, nearly three months to the day.

 

[chihlidog]

What went wrong....

- Hype. It did not live up to AMD's marketing hype. Feels like Bulldozer all over again. AMD ranted and raved about how fast it was, how it was awesome for 4k gaming, and how it was going to be an overclocking beast. Sure, it's fast, but not as fast as they were hinting at. It cant run 4k at settings and FPS that are worth playing, so that's out the window. And, at least right now, overclocking potential is basically zero. So it failed to live up to what AMD was claiming it was going to be.

- Price. Why in the world would anyone buy a Fury over a 980ti? Same price, better performance from the 980ti. Simple choice.

If it had not been hyped to the level it was, and launched at, say, 75 dollars cheaper, then the reception would have been pretty good IMO. Not the win I had hoped for from AMD, but pretty kickass nonetheless.

I think that the overarching feeling of disappointment from most AMD fans, though, stems from the fact that AMD's top end card, which is coming out so long after the Titan X, still cant take the crown from NV. They just cant do it, and so many of us want them to.

 

[geoxile]

Wow I just checked and you are right. It feels like 3 months!

You're drifting in time. Come back to us, with winning lotto numbers of course.

But on topic. If AMD can't hack on the hardware they should do it via software. I'm talking about gameworks-like partnerships, exclusive libraries like PhysX, and so forth, etc. AMD needs to put more work into their dev relations.

 

[sandorski]

Wow I just checked and you are right. It feels like 3 months!

Ya, the last few weeks seemed to crawl.

 

[Elixer]

If AMD doesn't have the R&D necessary to do this, they need to hurry up and sell the company to Samsung.

http://forums.anandtech.com/showpost.php?p=37509599&postcount=476

I don't think I need to add to that post, and no, the sky isn't falling, and while it may be depressing, it isn't just doom & gloom.
Could have it been better? Sure. Was it possible? Doubtful.

It isn't about just R&D, it is using all your resources you have available at the time the product will launch.

Just selling the company to Sammy or Qualcomm wouldn't fix everything, in fact, it could just mean the end of competition on desktop graphics cards, and EVERYONE would hate that.

 

[Abwx]

The thread title is a trollish title, doesnt surprise me given the op post history...

Postmortem is a latin saying (of Rome time) whose meaning is "after the death" (i.e after the death has occured)....

Could this title be edited.?. as it s by the definition a deffamatory title.

Edit : The right expression would be "a posteriori"....

 

[SolMiester]

I do think the performance will jump considerably with driver updates, as almost every GPU release from AMD has shown.

 

[monstercameron]

I do think the performance will jump considerably with driver updates, as almost every GPU release from AMD has shown.

you are surprisingly positive.
we can almost expect another 10-15% improvement via drivers and the dx12 landscape may be interesting.

 

[Abwx]

why do Nvidia's Maxwell cards overclock so much better than AMD's GCN offerings?

Because they paid for a custom process at TSMC, same as Altera, rest are urban legends, and dont come with fishy explanations about pipeline length, we are not at 4GHz...

 

[YBS1]

Where to begin? How about sites that test a variety of games instead of piling on crap like Project Cars, Dying Light etc..

It basically ties 980Ti, while running cooler & a lot quieter, exhausting heat out the case. ie. all the benefits of water cooling at the same price.

"In UHD scenarios this card is easily able to trade blows with the GTX 980 Ti (if it wasn’t for a lack of optimizations in Dying Light, it would actually be ahead by 2%) and gives the uber expensive TITAN X a swift kick in the pants."

And a compute beast too, destroying the competition on performance and perf/w.

As to why it doesn't shine more? Easy.

Custom 980Ti models. Great overclockers.
More GameWorks titles being tested.

I'm going to try to counterpoint this in the least combative way possible, quite frankly we just don't know the full story yet with Fury, the overclocking, possibly early drivers, etc. The problem with your argument is a lot of these so called biased games are in fact games people play. In fact quite a few of the biggest recent/upcoming games are. Second, you are trying to show the Fury in the best possible light...Now I know the 4K/multi monitor resolutions have that wow factor that catches people's attention and the Fury is pretty close there but let's get real, the vast majority of gamers (yes, including the ones buying these type of cards, even multiple cards) are running plain old 1080P. The 980Ti simply decimates it there. Now I'm not even going to pretend to know what's going to happen weeks or months from now, but that is the reality of the high end battle at this point.

 

[frozentundra123456]

WOW, this forum is a tough crowd. I dont think the Fury is a failure, as long as its competitive it gives users a choice. Perhaps the price needs reviewing.
What I found interesting from the data so far is that AMD usually widened the gap the high the resolution, but in this instance its the other way around.
I never thought GCN was memory bandwidth staved, and todays results I think show that, hence little improvement in the lower resolutions.
With that in mind, I dont think HBM did much for Fury, so AMD better come up with something more powerful for next time, cause when NV get pascal with HBM2, they will leave AMD behind.

HBM seems the killer tech for APUs if the cost can be brought down, or if the can somehow integrate say 16gb system ram with 6 or 8 gb of HBM for the gpu. But I never expected a great benefit to discrete cards, because they aren't particularly bandwidth limited to begin with.

I would not call Fury a failure. It is a decent product, but I think the price will have to come down, since performance is similar to 980Ti, and nVidia always seems to be able to command a higher price for the same performance. Fury is a halo product though. The problem is with the rest of the line-up. Old, tired, and power hungry.

 

[brandonmatic]

What's all this gloom and doom about? As far as I can tell the Fury is roughly equivalent to a stock 980Ti - although currently slightly slower on average. But it's a brand new architecture and my guess is that the Fury will end up significantly faster with driver updates.

That said, was it overhyped by AMD? Yes. Would I buy one now? No. I'm not in the market for a $650 card but if I were I'd probably buy one of the OC 980Ti's. But I think the Fury will end up performing better than the 980Ti over the long run.

I also hope for the sake of competition that the Fury is a commercial success.

 

[at80eighty]

their failure stemmed from pricing. they knew they were up against an excellent competitor in the 980Ti & should have reconsidered pricing, so it would adjust expectations. improvements will come over time & it would have made it a higher value card then. oh well.

This is a good enough card that will now lose out to the snowball hysteria of online reviews. further worries me that the GPU and (indirectly) CPU industries are headed towards effective monopolies. that should scare every leader diehard.