AMD Fury Series disappointment

[AtenRa]

Personally i would not even bother with Fury X, I would only introduce two cards, Fury Nano at $599 and Fury X2 at $1200.

 

[JoeRambo]

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

What sealed the deal for me was how well it OCed vs how "exploited to the max" Fury was. AMD trumpeted "built for overclocking" all day long, but under delivered disastrously in the end!

I think ANY custom 980ti can run 1.4Ghz effective clock and that is the real halo competitor for AMD 28nm gen. At least for forum crowds and trend setters

 

[Head1985]

Same ROPs and front end of hawaii killed the card.

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

This

 

[tviceman]

Same ROPs and front end of hawaii killed the card.

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

Unfortunately there is only so much die space available and there is also a power budget a GPU has to stay within when being designed. Adding in the extra ROP's and geometry engines would have ballooned both the die size and TDP, forcing other areas of the GPU to be cut down.

None of us will ever know how it could have turned out because, well, none of us here are smart enough to make GPU's.

 

[dogen1]

This

Only for some games though. Some games show a performance improvement over hawaii nearly equivalent to it's increase in CUs. AC Unity at 4k is one(difference drops somewhat at lower resolution, which also demonstrates that higher resolutions do not necessarily tax ROPs more than other parts of the GPU). This means that a 3800 ALU 96 ROP fury x would not be faster in all games.

 

[3DVagabond]

Here's what I think. AMD sucks at predicting the future. The believed that low level API's that leverage the strengths of GCN would have been proliferating in usage by now. They even went as far as writing one and handing it to everyone to show how it's done. Every OS out there adapted and adapted the idea. Surprise, surprise though... Nobody is using it. They need to be bribed, coerced, and paid off before they'll bother writing code for the new API's. IMO, this is what has held Fiji back. Also, the two prior gens.

 

[Silverforce11]

Unfortunately there is only so much die space available and there is also a power budget a GPU has to stay within when being designed. Adding in the extra ROP's and geometry engines would have ballooned both the die size and TDP, forcing other areas of the GPU to be cut down.

None of us will ever know how it could have turned out because, well, none of us here are smart enough to make GPU's.

AMD pretty much said as much when Fury X was released, they talked to Computerbase.de and they specifically said Fiji is a compromised design, that cannot reach efficient levels at lower resolutions because the front end is the same.

Though in AMD's PR, they said "Designed for 4K"... fair enough but its only because they compromised it. A stop-gap or test-bed solution.

 

[tential]

What sealed the deal for me was how well it OCed vs how "exploited to the max" Fury was. AMD trumpeted "built for overclocking" all day long, but under delivered disastrously in the end!

I think ANY custom 980ti can run 1.4Ghz effective clock and that is the real halo competitor for AMD 28nm gen. At least for forum crowds and trend setters

They brought light to overclocking specifically and got destroyed in the metric

 

[3DVagabond]

Only for some games though. Some games show a performance improvement over hawaii nearly equivalent to it's increase in CUs. AC Unity at 4k is one(difference drops somewhat at lower resolution, which also demonstrates that higher resolutions do not necessarily tax ROPs more than other parts of the GPU). This means that a 3800 ALU 96 ROP fury x would not be faster in all games.

The relative lack of performance could be caused by API overhead.

 

[tviceman]

AMD pretty much said as much when Fury X was released, they talked to Computerbase.de and they specifically said Fiji is a compromised design, that cannot reach efficient levels at lower resolutions because the front end is the same.

Though in AMD's PR, they said "Designed for 4K"... fair enough but its only because they compromised it. A stop-gap or test-bed solution.

It seems like every AMD high end GPU is stuck with a "would have been better if" slogan on the box. Cayman woulda coulda shoulda didn't. Tahiti woulda coulda shoulda did two years later. Hawaii woulda coulda shoulda power noise heat cooler. Fiji woulda coulda shoulda engineers don't know squat.

Always an excuse. At least their parts have been catching up in performance over time with Nvidia parts; the console wins are paying dividends.

 

[Silverforce11]

That's AMD. Good engineers, terrible management.

Tahiti was an OC champ... left too much performance on the table out of the box.

Hawaii was great... crippled by crap reference cooler.

Hopefully with the new management, they turn it around. We shall see soon enough, Zen & Polaris definitely needs to perform due to their dire situation.

 

[dogen1]

AMD pretty much said as much when Fury X was released, they talked to Computerbase.de and they specifically said Fiji is a compromised design, that cannot reach efficient levels at lower resolutions because the front end is the same.

Though in AMD's PR, they said "Designed for 4K"... fair enough but its only because they compromised it. A stop-gap or test-bed solution.

Oh, if that was the official statement then we have our answer.

Do you know exactly what they mean by "front-end".

Going by this diagram, I assume it was the geometry processors they were talking about.

 

[AMDisTheBEST]

I'll pick fury over Maxwell anyday.

 

[Arachnotronic]

I'll pick fury over Maxwell anyday.

Of course you would. Your signature and handle suggest that you really love AMD.

 

[RampantAndroid]

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

I would have jumped on a Fury X for that price (or a Fury IF they had included the Never Settle bundle. The lack of Never Settle absolutely was the final straw for me to go get a 980 and wait for Pascal.

 

[3DVagabond]

Oh, if that was the official statement then we have our answer.

Do you know exactly what they mean by "front-end".

Going by this diagram, I assume it was the geometry processors they were talking about.

I can't remember the name of the feature that's coming in Polaris, but it's an advanced "back face culling" technique that apparently nVidia already uses. Removes the unseen tris before they are processed.

 

[dogen1]

I can't remember the name of the feature that's coming in Polaris, but it's an advanced "back face culling" technique that apparently nVidia already uses. Removes the unseen tris before they are processed.

Yeah, the primitive discard accelerator. If geometry processing is the bottleneck then what zlatan said about it being a big improvement makes sense.

 

[mysticjbyrd]

It seems like every AMD high end GPU is stuck with a "would have been better if" slogan on the box. Cayman woulda coulda shoulda didn't. Tahiti woulda coulda shoulda did two years later. Hawaii woulda coulda shoulda power noise heat cooler. Fiji woulda coulda shoulda engineers don't know squat.

Always an excuse. At least their parts have been catching up in performance over time with Nvidia parts; the console wins are paying dividends.

Who buys a furyx or 980ti for 1080p gaming?

If anything, I would argue AMD designs their gpus for what they expect they will logically be used for.

Tessellation is a great example. Amd's gpus can handle it fine as long as you don't crank the tessellation dial to 11, which does nothing for fidelity.

 

[Arachnotronic]

If anything, I would argue AMD designs their gpus for what's logical.

Seems like that would be NVIDIA considering they are putting the boot in pretty hard in terms of market share/sales

 

[3DVagabond]

Who buys a furyx or 980ti for 1080p gaming?

If anything, I would argue AMD designs their gpus for what they expect they will logically be used for.

Tessellation is a great example. Amd's gpus can handle it fine as long as you don't crank the tessellation dial to 11, which does nothing for fidelity.

Except they actually expect the industry to up their game to maintain advancement and keeping he cutting edge sharp. Instead though the industry maintains the status quo. They only advance when someone pays them. Where are DX12 games? Even console ports, which should be easier than DX11.

 

[IntelUser2000]

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.

It's horrible because we expect more than that. The ratio is 357/512(%70)

I suspected this would happen after reading Fury benches. It happens with every new memory technology. You needed DDR2-667 to reliably beat DDR-400. DDR2-533 wasn't enough. You needed DDR3-1333 to beat DDR2-800, not DDR3-1066. Same with DDR4. You start "needing" DDR4-2400. DDR3-2133 doesn't seem too good against DDR3-1600.

The way they increase the bandwidth you lose a bit of efficiency compared to previous generations. 2x theoretical = 1.8x, but first iteration of the new memory turns out to be 1.5x theoretical = 1.4x

With dual channel it wasn't double the bandwidth either. It'll probably be same with HBM2 versus HBM.

What HBM really given AMD is lower power usage of the memory chips allowing chip to use more which contributes to the performance rather than memory bandwidth.

 

[gamervivek]

The 'failure' of Fury has been clockspeeds and the lack of voltage control to make up for it.

It's faster than a 980Ti at 4k and will most probably improve that in coming months. AMD seem to have not concentrated on GCN3 in Tonga as much as they did on Hawaii.

10% better at stock + 10-15% OC improvement would land it close to the overclocked 980Tis.

HBM overclocking does pay dividends once you've overclocked the core. hardware.fr did a good review of it.

 

[Krteq]

The 'failure' of Fury has been clockspeeds and the lack of voltage control to make up for it

Voltage control is enabled in latest MSI Afterburner (version 4.2.0)

• Added AMD Fiji graphics processors family support.
• Hardware abstraction layer architecture has been revamped to allow implementation of voltage control via direct access to GPU ondie voltage controllers (e.g. AMD Fiji SMC) in addition to previously supported external voltage controllers connected to GPU via I2C bus. Please take a note that direct access to AMD SMC from multiple simultaneously running hardware monitoring applications can be unsafe and result in collisions, so similar to I2C access synchronization we introduce global namespace synchronization mutex “Access_ATI_SMC” as SMC access synchronization standard. Other developers are strongly suggested to use it during accessing AMD GPU SMC in order to provide collision free hardware monitoring.
• Added core voltage control for reference design AMD RADEON R9 Fury / Nano series cards with on-die SMC voltage controller
• Added unofficial overclocking support for PowerPlay7 capable graphics cards (AMD Tonga and newer graphics processors family). Please take a note that unofficial overclocking mode with completely disabled PowerPlay is currently not supported for PowerPlay7 capable hardware.

 

[gamervivek]

It was there with Trixx even earlier, but it tooks months for it to be pushed out.

And of course the other problem AMD have is nvidia's software supremacy. 970 matching a Fury X and AMD cards falling behind in games with green teams' involvement.

 

[Krteq]

Oh yeah, the "GameBarelyWorks" controversy