AMD Fury X Postmortem: What Went Wrong?

[redzo]

I think price/performance is what killed this product. I wonder how cheaper can a air cooled partner board fury X go. The GTX 980 would have had a hard time competing with a similarly priced fury X. I wonder what kept AMD from positioning the fury against or as close as possible to the GTX 980. AMD knew about the 980's performance.

 

[amenx]

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.

Its not a brand new architecture. Its (a larger) Tonga with HBM. Not sure how much driver improvements will bring to an older arch that they're already quite familiar with.

The way I see it, the big failings for AMD here are:

1- Price
2- Poor overclock headroom
3- poor connectivity options.

The areas where it seem to do well is UHD and temps. But whats the point of fantastic temps with little OC headroom? And on the UHD front, you basically need a UHD monitor with DP. If existing UHD TV, good luck with that.

Bottom line, for enthusiasts, 980ti with superior OC'ability is a no-brainer at same price. AMD would need a $100 price drop to stay in the game. Would devastate their margins, but not much choice.

I think the 980ti blind-sided them. I believe Fury-x was initially designed with lower clocks and higher OC headroom, but when 980ti arrived, they frantically did last minute changes of upping clock speeds to avoid a poorer showing in reviews. The same basic strategy with bulldozer, yank up the clocks to do avoid embarrassing reviews.

 

[Silverforce11]

@amex
Not sure why you believe such strange fantasies. 100mhz OC is quite GOOD for GCN without vcore. Anyone who has tried to OC 7970, R290/X etc would agree with that.

NV OC is easy because the vcore is dynamic, if you adjust extra power limits, the GPU will reach higher clock bins at a higher vcore.

AMD OC you need access to manual vcore control otherwise 100mhz is about the limits. This is exactly as it happened with R290/X, no surprises there at all.

The leak from last yeah from ChipHell had it at 1ghz or 1050mhz. It was measured to be 51-56% faster than Reference R290X. Which is spot on. Unfortunately for them, more GW titles being included in reviews, 51-56% above R290X is just not enough.

Think about what would the summary chart look like if AMD actually fix performance in Dying Light, Project Cars. In those games the gap is massive.

 

[amenx]

I guess depends on which review/card sample are involved. Also doesnt seem to get much performance scaling out of the OC.

http://www.guru3d.com/articles_pages/amd_radeon_r9_fury_x_review,37.html

But point taken, lets see what upping vcore can bring.

 

[Flapdrol1337]

Nothing went "wrong", it's just not as fast as the 980Ti in games.

 

[ShintaiDK]

People are saying it's priced too high. It's completely sold out.

Even reviewers said it was extremely limited.

 

[Dribble]

If you look at the levels of *hype* across the board you can see it's just not going to sell - sure one or two hard core fans will take a punt, but most of the markets going to be getting 980Ti's. Given that fury is the only card that isn't a straight rebrand of a range of cards that were already not selling well it's difficult to see how AMD can survive. They are just behind on so many fronts, and don't have the time or money to catch up.

As to what went wrong with fury in particular - well you could point to the 4GB memory, lack of rops, no HDMI 2, expensive bom, high power usage, lack of dev support, but the single biggest problem is the 980Ti. If nvidia only had the $1000 titan X Fury would have had a great market. Unfortunately nvidia saw what was coming and released the right card to counter it, not that AMD can really be surprised by that, it's what competitors do, but I suspect they were hoping nvidia might not bother and give them a foothold.

 

[Erenhardt]

Ironically they would be crippling their own 980 and 970 cards if that is the case.

Then there is another reason to do so.

If that forces 970 and 980 users to upgrade, they will most likely get another nv card.

 

[Hitman928]

Its not a brand new architecture. Its (a larger) Tonga with HBM.

 

[ShintaiDK]

Try remove all the actual cross references. And there may be nothing left besides HBM. And thats not a GCN tech in the first place.

 

[Silverforce11]

I guess depends on which review/card sample are involved. Also doesnt seem to get much performance scaling out of the OC.

http://www.guru3d.com/articles_pages/amd_radeon_r9_fury_x_review,37.html

But point taken, lets see what upping vcore can bring.

That's a 75mhz OC, or ~7%, pretty weak but thats quite normal for GCN without vcore. I suspect with vcore it would do up to 1.25ghz but with a major increase in power usage (hence the water cooler being put to good use).

Btw the scaling is about 5-7%.

83 -> 89 (7%)
134 -> 141 (6%)
107 -> 112 (5%)

From a gamer PoV, $599 would be a better deal WITH the potential vcore OC. If its not possible to tweak vcore, $549. Custom 980Ti @ 699 is just much better value. Which is strange when we think about AMD vs NV.

 

[AtenRa]

Try remove all the actual cross references. And there may be nothing left besides HBM. And thats not a GCN tech in the first place.

New 16-bit float and Integer instructions
New data parallel processing instructions
Improved Compute Task Scheduling
L2 cache size increased to 2MB
New memory interface for the HBM.

Its not the same as Tonga GCN 1.2

 

[Silverforce11]

People should not talk trash about potential driver improvements. Look what happened with Omega drivers, then recently with the 300 series launch drivers.

If anything, GCN has proven itself to mature very well over time.

 

[ShintaiDK]

New 16-bit float and Integer instructions
New data parallel processing instructions
Improved Compute Task Scheduling
L2 cache size increased to 2MB
New memory interface for the HBM.

Its not the same as Tonga GCN 1.2

Updated ISA instruction set. That one can easily be split out. Just like the 2 renames for the scalar and UVD.

http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/2

This is GCN 1.2. Sounds familiar?

 

[Erenhardt]

I guess depends on which review/card sample are involved. Also doesnt seem to get much performance scaling out of the OC.

http://www.guru3d.com/articles_pages/amd_radeon_r9_fury_x_review,37.html

But point taken, lets see what upping vcore can bring.

5-7% perf increase from 7% gpu clock increase is rather good.

As silver said, amd cards work on fixed voltage that has to be manually set for overclocking/overvolting.

nv has a dynamic voltage that increases automatically as the core clock is increased. It doesn't give the user full control over the gpu.

Sadly, the tools for increasing GPU voltage were not ready for Fury Release.

 

[Enigmoid]

New 16-bit float and Integer instructions
New data parallel processing instructions
Improved Compute Task Scheduling
L2 cache size increased to 2MB
New memory interface for the HBM.

Its not the same as Tonga GCN 1.2

The L2 cache is nothing special. Tonga has 1 MB for 2048 shaders. Fiji has 2 MB for 4096 shaders. That may be an increase to 2 MB but its nothing significant or noteworthy. Its bigger so it has more cache but the cache ratios are the same as tonga.

 

[dark zero]

HBM is not the problem.. That crappy GPU is...

 

[tviceman]

That's a 75mhz OC, or ~7%, pretty weak but thats quite normal for GCN without vcore. I suspect with vcore it would do up to 1.25ghz but with a major increase in power usage (hence the water cooler being put to good use).

Btw the scaling is about 5-7%.

83 -> 89 (7%)
134 -> 141 (6%)
107 -> 112 (5%)

From a gamer PoV, $599 would be a better deal WITH the potential vcore OC. If its not possible to tweak vcore, $549. Custom 980Ti @ 699 is just much better value. Which is strange when we think about AMD vs NV.

IF it overclocks well, if drivers bring improvements, if Maxwell drivers remain stagnant, if it isn't summer time when Fury X is sucking down 550 watts, if your electric rates aren't as bad as when Fermi came out.....

How about if AMD quit shooting themselves in the foot with every new product release.

Bottom line: Fury X: runs quiet. Doesn't overclock well. Doesn't run 4k fast enough, falls behind at 1440p, gets destroyed at 1080p 120/144hz. Costs the same.

 

[cbrunny]

Is Fury X dead?

 

[Silverforce11]

IF it overclocks well, if drivers bring improvements, if Maxwell drivers remain stagnant, if it isn't summer time when Fury X is sucking down 550 watts, if your electric rates aren't as bad as when Fermi came out.....

Hyperventilating much?

Everything I posted in that quote is spot on with how GCN behaves with & without vcore OC. Found anything wrong with that assessment?

I also ended by saying it needs a price cut to be competitive.

 

[vissarix]

AMD Fury X Postmortem: What Went Wrong?

what are you talking about?

 

[amenx]

^AMDs own pre-release benches. I presume you've missed all the reviews and went to manufactures site for that. As post-release reviews show, never trust any manufacturers own testing.

http://www.techpowerup.com/reviews/AMD/R9_Fury_X/31.html

 

[Keysplayr]

Hyperventilating much?

Everything I posted in that quote is spot on with how GCN behaves with & without vcore OC. Found anything wrong with that assessment?

I also ended by saying it needs a price cut to be competitive.

I don't think he is hyperventilating. Is that how you see him while reading his responses? I see him severely tired of the "IF" word being used to somehow makes AMD's situation lighter or show hope. If if if. It's kind of enough enough enough already.
There is always room for driver improvements for any architecture, however many feel this is just a larger Hawaii GPU, and some say it's a large Tonga GPU. If that is the case, then the drivers are probably pretty mature right about now. Maybe, and I use the term loosely because it reminds of "if", better memory management in the drivers is possible as HBM is relatively new. I can see room for improvement there. Of course this could be me totally overestimating the need for bandwidth. It could have all it ever needs and all it could ever take advantage of as of right now. Don't know.

 

[exar333]

From a pure performance standpoint, I wonder if Tonga was a step backwards? If this was a tweaked Hawaii with HBM and 50% more everything, it likely would be better. We saw the 285 exhibit erratic performance and this seems to follow suit. Sometime amazing, sometimes meh. Just hypothesizing...

 

[gamervivek]

Fury's achilles heel seems to be the lower resolution performance which is why it's quite comical that nvidia introducted DSR with the 9xx series and now they get comprehensively outplayed on that front with VSR having next to no hit and also not blurring the whole image.