AMD Fury Series disappointment

[ultron]

I wonder the reasons of Fury Series's failure. The card has very high theoretical bandwith and crazy compute power. (But it has only 64 ROPs.) It had to be have more real life performance but it has not.

The question is:Why?

There's my assumptions about what caused this:

1/The first reason based on a fact not an assumption.

The card's bandwith is not very high in the real life applications. So maybe even this card is limited by memory performance.

But why HBM fails? The test result let AMD lovers down.

When we look at the benchmark we see a ratio about test result/theoritical bandwith ratio. It's about %75-%80. (TITAN has 288 GB/s memory power and 221 GB/s memory copy performance result---->%76, R9 390X has 384 GB/s-302 GB/s--->%78). But when we look at the result of Fury X, we see a horrible score: 357 GB/s. It's horrible because we expect more than that. The ratio is 357/512(%70)

2/The color compression doesn't work well due to unknown reasons.

3/Number of ROPs hold the potential of the card. Hawaii and GTX 980 has 64 ROPs-they aim lower performance class, TITAN X and 980 Tİ has 96. Are we sure that there's no bottleneck on the GPU and the GPU has well balanced?

4/Additional compute power helps nothing.

5/There's something wrong with the drivers.

6/Unknown, unpredictable reasons. Something limits Fury somehow.

I'm really wonder the reason. But i have not enough information to make accurate comment about this event. For example:

*I don't know which game's any frame needs how much instructions be executed, how much pixel has drawn, how much triangle to be created etc. etc. Someone profiling some games but there's no database about it in the hardware websites. I am tinkering around with some profiling tools, but i am very amateur and i am not be able to reach any useful information.

*I don't have any idea how the drivers work and how this affect the performance. Maybe the problems on the drivers hit Fury more than any other GCN cards.

Are there someone any ideas about how could so powerful card not to be powerful in the real life.

Sorry about my English.

 

[flopper]

I wonder the reasons of Fury Series's failure. The card has very high theoretical bandwith and crazy compute power. (But it has only 64 ROPs.) It had to be have more real life performance but it has not.

Are there someone any ideas about how could so powerful card not to be powerful in the real life.

Sorry about my English.

The Fury beats the 980ti in 4k so not even sure why you ask the question.
Beyond3d is normally the website you want to visit to read up on tech in the gpu cards.
You have Nano that is the world class leader in a 175w envelope.

So not sure whats a failure here, the compute addition makes Fury excellent for various tasks and the Fury also is VR ready where the Maxwell 980ti or Titanx isnt with hardware as they dont have the same fine tuned latency that is important for VR so there Nvidia failed BIG time and that AMD simply had the hardware ready for VR.

Engineers design around specific targets and if you compare agaisnt different uses of a card like a car for example, if you buy a beach car that looks nice and then drive it in a winter landscape you freeze to death. Is then the beach car better than a car made for winter landscape a better or worse car?
ask the dead guy.

 

[BigDaveX]

The Fury beats the 980ti in 4k so not even sure why you ask the question.

Not consistently, though. The benchmarks on this very site (link) show it as a fairly even mix of results between the Fury X and 980 Ti at 4K. And that's before you get into the fact that most gamers are still getting by with 1080P or 1440P monitors.

I think the big issue was that AMD got blindsided by the 980 Ti, as the Fury X would otherwise have fit nicely into the gap between the 980 and Titan X. Aside from that, the ROP count is probably the biggest individual issue, considering how the vanilla Fury hardly loses any performance from having 12% of its shaders hacked out.

 

[ShintaiDK]

The difference is most likely due to the differences between GDDR and HBM. Not due to Fury as a chip.

 

[DiogoDX]

Same ROPs and front end of hawaii killed the card.

3800sps - 96ROPs - 6 raster/geometry would be a much faster card.

 

[el etro]

If only Fiji was really a GM200 killer, it would not a "failure", even if efficiency was worse.

Anyway i don't think is a failure, it serves as test-drive for future High-End HBM Videocards, Trails GM200 performance by a bit, achieving it only via small architectural improvements and still maintaining competitive performance/mm² versus Maxwell, this last one being a real all new architecture. And the Nano catches up to Maxwell in performance per watt.

Really, Fiji hold the line very well.



If 300 series was not most made of refreshs/rebrands, it would deliver a far better impression to consumers and for the market.

 

[Stormflux]

I think the entire world got side-swiped by the great 980Ti. If we compare the FuryX to the TitanX at say a speculative price point of $700/$800 to $1000 for 1440p+ gaming how does it look?

The 980Ti came out 3 months after the TitanX and completely blew away that card. Custom cooling, overclocking potential, right amount of ram, for $350 less? It completely made both the TitanX and 980 irrelevant and questionable buys. It is that good of a card. A month later the FuryX enters the arena with the crowd still singing the praise of the ungodly 980Ti.

AMDs whole line up was late to a party it forgot. AMD's plans for 22nm went up in smoke. AMDs marketing/mindshare was/is terrible. AMD just could not execute.

Soft reset incoming with 14/16nm. I say soft because AMD or RTG still has to battle against the mindshare it has collected and learn from the lessons of long in the tooth 28nm generation.

 

[KingFatty]

Is it because you've just manufactured a narrative in your mind, and you are looking to justify it somehow?

Instead of deciding on your answer ahead of time, start with the facts and see where they lead. You could call anything a failure if you believe it in your heart, and then selectively pick and choose evidence to build up your mistaken belief. But that is not helpful at getting at the truth.

 

[moonbogg]

I think the entire world got side-swiped by the great 980Ti. If we compare the FuryX to the TitanX at say a speculative price point of $700/$800 to $1000 for 1440p+ gaming how does it look?

The 980Ti came out 3 months after the TitanX and completely blew away that card. Custom cooling, overclocking potential, right amount of ram, for $350 less? It completely made both the TitanX and 980 irrelevant and questionable buys. It is that good of a card. A month later the FuryX enters the arena with the crowd still singing the praise of the ungodly 980Ti.

AMDs whole line up was late to a party it forgot. AMD's plans for 22nm went up in smoke. AMDs marketing/mindshare was/is terrible. AMD just could not execute.

Soft reset incoming with 14/16nm. I say soft because AMD or RTG still has to battle against the mindshare it has collected and learn from the lessons of long in the tooth 28nm generation.

Fully agree. The 980ti was a sudden ninja kick to the neck! Watching the timing of the 980ti release, the cost, performance and all the extra good stuff about it, it was like watching a football game where Nvidia sacked the quarterback HARD and then scored a touchdown to win the game before the other team could even finish a play. It was beautiful because it was strategic and the perceived value compared to the Titan X was so extreme, people couldn't wait to get their hands on a 980TI (or two, or three, or four).
14/16 is the start of a whole new ball game.

 

[boozzer]

Fully agree. The 980ti was a sudden ninja kick to the neck! Watching the timing of the 980ti release, the cost, performance and all the extra good stuff about it, it was like watching a football game where Nvidia sacked the quarterback HARD and then scored a touchdown to win the game before the other team could even finish a play. It was beautiful because it was strategic and the perceived value compared to the Titan X was so extreme, people couldn't wait to get their hands on a 980TI (or two, or three, or four).
14/16 is the start of a whole new ball game.

perfect plan by nv really some posters here even called it. they were absolutely correct. titan x was there just to give perceive value to 980 ti

 

[3DVagabond]

I wonder the reasons of Fury Series's failure. The card has very high theoretical bandwith and crazy compute power. (But it has only 64 ROPs.) It had to be have more real life performance but it has not.

The question is:Why?

There's my assumptions about what caused this:

1/The first reason based on a fact not an assumption.

The card's bandwith is not very high in the real life applications. So maybe even this card is limited by memory performance.

But why HBM fails? The test result let AMD lovers down.

When we look at the benchmark we see a ratio about test result/theoritical bandwith ratio. It's about %75-%80. (TITAN has 288 GB/s memory power and 221 GB/s memory copy performance result---->%76, R9 390X has 384 GB/s-302 GB/s--->%78). But when we look at the result of Fury X, we see a horrible score: 357 GB/s. It's horrible because we expect more than that. The ratio is 357/512(%70)

2/The color compression doesn't work well due to unknown reasons.

3/Number of ROPs hold the potential of the card. Hawaii and GTX 980 has 64 ROPs-they aim lower performance class, TITAN X and 980 Tİ has 96. Are we sure that there's no bottleneck on the GPU and the GPU has well balanced?

4/Additional compute power helps nothing.

5/There's something wrong with the drivers.

6/Unknown, unpredictable reasons. Something limits Fury somehow.

I'm really wonder the reason. But i have not enough information to make accurate comment about this event. For example:

*I don't know which game's any frame needs how much instructions be executed, how much pixel has drawn, how much triangle to be created etc. etc. Someone profiling some games but there's no database about it in the hardware websites. I am tinkering around with some profiling tools, but i am very amateur and i am not be able to reach any useful information.

*I don't have any idea how the drivers work and how this affect the performance. Maybe the problems on the drivers hit Fury more than any other GCN cards.

Are there someone any ideas about how could so powerful card not to be powerful in the real life.

Sorry about my English.

Two things. If you want to not simply look like a troll then start off by defining "failure". It's a very strong word to use with no qualification. 2nd, You need to supply facts to back your assumptions.

 

[ultron]

Guys you are missing very big point:

As i mentioned the first message, there's very strange correllation about this card theoritical power/real world performance. With 512 GB/s bandwith, 4096 Shaders, 256 TMUs Fury X had to overwhelm GTX 980 Ti. But it didn't. Now i ask:Why? Talking about value of card is meaningless.

ROP Count is a suspect. Why a weaker on paper GPU like 980 Ti has 96 ROPs?

And i am looking for an explanation for this benchmark result:

I have no idea the reason behind this disappointing result.

I can expect more performance. It has %33 theoretical, %18 real life according to AIDA64 bench. more bandwith power than R9 390X. And since GCN 1.2 the new color compressions came in. So real life memory performance has to at least (in the most pesimist assumption)%25 higher than Hawaii.

Not the mention great compute power. You can say modern games are not shader bound, so why AMD made a lot useless things and enlarged the die size? They aimed GPGPU market? They aimed the future and sacrificed the present(Yes, i can describe that as sacrifice because bigger die size means bigger cost, bigger cost means lower gross margin and that means "failure" as the financial point of view)

I searched internet but no one explained the reason of AIDA64 memory benchmark surprise.

@el etro: competitive performance/mm² but HBM has higher cost than GDDR5. Plus liquid cooling is not cheap. i don't know the yields but I know wafer costs and i am definetly sure that the chip is not cheap to manufacture itself and in addition HBM is an expensive tech.

edit: @3DVagabond I saw your message now. I already defined "failure" ib this message. I can't prove my assumptions because i don't really know which is the point that game performance limited, and i am asking the users that any idea about it. This is the reason why i started this post.

I am very sad about looking like a troll by the way.

 

[tviceman]

The Fury beats the 980ti in 4k

The 980 TI is a cut down chip, is still more efficient, and GM200 has way, way more OC headroom. If you want to compare price points then go for it but note that 980 TI's can be had for $50+ cheaper than Fury X and have big OC's out of the box making those 980 TI's faster than Fury X at 4k.

From a technical perspective, it's better to compare Titan X to Fury X - both fully functional chips of the same size, same generation, and same node.

 

[Leadbox]

The 980 TI is a cut down chip, is still more efficient, and GM200 has way, way more OC headroom. If you want to compare price points then go for it but note that 980 TI's can be had for $50+ cheaper than Fury X and have big OC's out of the box making those 980 TI's faster than Fury X at 4k.

From a technical perspective, it's better to compare Titan X to Fury X - both fully functional chips of the same size, same generation, and same node.

980Ti is cut down were? 6GB memory instead of 12? Does it not spot the same # of CCs, rops etc as the Titan X?

 

[el etro]

I think the Fiji chip alone yields at 95% or more(No way yeilds are bad, is a four year old process!), and 9 of 10 fine Fiji chips going successful from the HBM-interposer implant.

 

[parvadomus]

Guys you are missing very big point:

As i mentioned the first message, there's very strange correllation about this card theoritical power/real world performance. With 512 GB/s bandwith, 4096 Shaders, 256 TMUs Fury X had to overwhelm GTX 980 Ti. But it didn't. Now i ask:Why? Talking about value of card is meaningless.

ROP Count is a suspect. Why a weaker on paper GPU like 980 Ti has 96 ROPs?

And i am looking for an explanation for this benchmark result:

I have no idea the reason behind this disappointing result.

I can expect more performance. It has %33 theoretical, %18 real life according to AIDA64 bench. more bandwith power than R9 390X. And since GCN 1.2 the new color compressions came in. So real life memory performance has to at least (in the most pesimist assumption)%25 higher than Hawaii.

Not the mention great compute power. You can say modern games are not shader bound, so why AMD made a lot useless things and enlarged the die size? They aimed GPGPU market? They aimed the future and sacrificed the present(Yes, i can describe that as sacrifice because bigger die size means bigger cost, bigger cost means lower gross margin and that means "failure" as the financial point of view)

I searched internet but no one explained the reason of AIDA64 memory benchmark surprise.

@el etro: competitive performance/mm² but HBM has higher cost than GDDR5. Plus liquid cooling is not cheap. i don't know the yields but I know wafer costs and i am definetly sure that the chip is not cheap to manufacture itself and in addition HBM is an expensive tech.

edit: @3DVagabond I saw your message now. I already defined "failure" ib this message. I can't prove my assumptions because i don't really know which is the point that game performance limited, and i am asking the users that any idea about it. This is the reason why i started this post.

I am very sad about looking like a troll by the way.

1 - The Fiji chip does overwhelm GM200 for certain tasks. Its compute performance is way higher. I would expect it to perform significantly faster at gpgpu stuff.
2 - GM200 on the other way, its much better made for gaming. It has a bigger backend with its 96Rops, its a lot less denser (which helps with clock speeds), and it has a way better memory compression technology than the Fiji GPU (look at beyond3d texturing bandwidth measurement):
http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4
This means that Fiji memory compression is on par with kepler (they probably reverse engineered it or something).
3 - Another theory is that AMD killed fiji performance with too much power gating, that added internal latencies to the chip itself, otherwise it would perform at least 2x the performance of the 380X and thats not the case.

Maybe a smaller die with 64 or 96rops/3584shaders but with no gating could have performed better.

 

[tviceman]

980Ti is cut down were? 6GB memory instead of 12? Does it not spot the same # of CCs, rops etc as the Titan X?

Nope, 22 active SMM's instead of Titan X's 24. And, of course but inconsequential, the vram size.

 

[Glo.]

How big affect on Fury X/Nano would have 96 ROPs?

Is it possible to estimate it?

 

[master_shake_]

Not consistently, though. The benchmarks on this very site (link) show it as a fairly even mix of results between the Fury X and 980 Ti at 4K. And that's before you get into the fact that most gamers are still getting by with 1080P or 1440P monitors.

I think the big issue was that AMD got blindsided by the 980 Ti, as the Fury X would otherwise have fit nicely into the gap between the 980 and Titan X. Aside from that, the ROP count is probably the biggest individual issue, considering how the vanilla Fury hardly loses any performance from having 12% of its shaders hacked out.

11 out of 15 times the fury x beats the 980ti and sometimes the titan x at 4k.

how is that not consistent?

http://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_Waterforce/1.html

 

[dogen1]

How big affect on Fury X/Nano would have 96 ROPs?

Is it possible to estimate it?

It would only benefit ROP bound parts of rendering. So, simple pixel shaders(most will be ALU or TMU bound), probably shadows, g-buffer creation I think. Most games use compute shaders for a lot, including lighting, nowadays, and those don't touch ROPs at all. Some devs go as far to say modern engines should use compute shaders for everything except rasterizing triangles(can't wait to see games actually doing this, should be interesting). Fury would probably be a monster with that kind of workload.


Essentially, the older the game, the more the benefit would be. Probably.

 

[Silverforce11]

11 out of 15 times the fury x beats the 980ti and sometimes the titan x at 4k.

how is that not consistent?

http://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_Waterforce/1.html

The problem for Fury X has always been custom 980Ti models, those boost to 1.4ghz out of the box and users can further OC to 1.5ghz (or some models even hit around that out of the box).

The reference is pretty tame, ~1.2ghz boost.

When you look on the market and buy cards, custom 980Ti is plain faster even at 4K.

Fury X doesn't get to beat it until we get multi-GPU scenarios. There, it is more than competitive.

As to why Fury X seems lacking in performance with all those shaders? Same front end as Tonga/Hawaii is made to drive many more shaders.

 

[2is]

perfect plan by nv really some posters here even called it. they were absolutely correct. titan x was there just to give perceive value to 980 ti

I think even Titan X buyers knew it, that's exactly what happened when the 780Ti launched

 

[tential]

For me, it was seeing the 980Ti blow the Fury X away at the SAME price point that sold the deal for me on the 980Ti.

 

[desprado]

The Fury beats the 980ti in 4k so not even sure why you ask the question.
Beyond3d is normally the website you want to visit to read up on tech in the gpu cards.
You have Nano that is the world class leader in a 175w envelope.

So not sure whats a failure here, the compute addition makes Fury excellent for various tasks and the Fury also is VR ready where the Maxwell 980ti or Titanx isnt with hardware as they dont have the same fine tuned latency that is important for VR so there Nvidia failed BIG time and that AMD simply had the hardware ready for VR.

Engineers design around specific targets and if you compare agaisnt different uses of a card like a car for example, if you buy a beach car that looks nice and then drive it in a winter landscape you freeze to death. Is then the beach car better than a car made for winter landscape a better or worse car?
ask the dead guy.

Wrong. Check the custom GTX 980 Ti models. However, wait i post it here for you.

http://www.hardocp.com/article/2015...0_ti_lightning_video_card_review#.VpIJV1m8SwY
http://www.guru3d.com/articles-pages/msi-geforce-gtx-980-ti-lightning-review,1.html
https://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_Waterforce/
http://hexus.net/tech/reviews/graphics/88325-gigabyte-geforce-gtx-980-ti-xtreme-gaming-waterforce/
http://www.overclock3d.net/reviews/gpu_displays/corsair_gtx980_ti_hydro_gfx_graphics_card_review/14

 

[Gikaseixas]

Fury is a great card but lack of OC, wrong pricing dictated it's fate. It should be priced like this:

$579.99 - Fury X
$549.99 - Nano
$499.00 - Fury

I went with a GTX 980ti the minute i saw how awesome it was compared against Titan and Fury cards