AMD Radeon R9 Fury X 4 GB

[Lord Nemesis]

Personally, the main problem that I see with this card is the use of HBM Memory. HBM with its current technical limitation (4GB) is not ready for the big time on a flagship card. AMD pitched this as the pinnacle of cards for UHD, but did nothing to support that. While lack of HDMI 2.0 port is one of those, the more important thing is the 4 GB of HBM that they put on this card.

What is the actual driving force for desiring UHD gaming? It is not just the screen resolution, When we have 4 times the pixel density of FHD, we would like to use those extra pixels to present more detail and that comes from higher and more detailed textures. What we are seeing as UHD gaming today is just the games being run at UHD resolution with the same textures that were designed for FHD. In place of one pixel in FHD, we have four pixels with the same shade in UHD which is quite pointless. The goal of UHD will be realized when we start getting more detailed textures to use for UHD and such textures are going to occupy a lot of VRAM.

No amount of PR about driver optimizations, texture compression and the high bandwidth offered by HBM would be able to side track the fact that 4GB is not going to be enough in the long run when higher resolution textures come into the picture. It maybe true that VRAM is not getting utilized efficiently today and that there might be ways to optimize the drivers to make the allocations better, but applies only for the present day situation when there is head room for such optimizations. Once true UHD optimized games with higher resolution textures start coming out, they simply will not fit in the 4GB VRAM and once swapping from main memory starts, you all know that the bottleneck is going to kill the performance.

Either AMD is banking on the fact that most multi platform games may not yet offer super high resolution textures for UHD in the PC versions or that for the games that do so, they can just compromise the texture quality at the driver or ask the developer to fallback to lower resolution textures for these cards.

Even if the card has the raw compute power for UHD, they have crippled this card by pairing it with 4GB HBM. I would have much rather preferred to have them pair this with 8 GB GDDR instead. They should have waited on the HBM till they could come out with 8 GB modules. Pair it with 8GB GDDR, replace the water cooling with the regular air cooled designs and market it at $500 or even $550 and this would have killed the 980 and 980Ti using the value for money tag. This was never a UHD ready card to begin with. I consider the 390X, to be more of a UHD ready card than Fury X.

There never was any need for them to be king of the hill in terms of performance, they could have claimed it in terms of value for money has they have done in the past. My last 3 GPU purchases and most of my overall GPU purchases were AMD for this reason, but this time, I went for GTX 980 after 10 years of not using an nVidia card because it offered more value for my money when I bought it. Personally I never card about Physx or any of the nvidia specific stuff, but the overall value justified the purchase.

Can Fury X compete with 980 Ti at the same price point? Not unless they beat the 980 Ti in 95% of the games with at least a 7.5~10% better performance margin which currently is not the case. 74% of the gaming GPU market share is currently owned by nVidia and a vast majority of AAA titles are using Gameworks features that add additional value for nVidia GPU users and it doesn't help that AMD fails to optimize their drivers and the game for their own GPUs and resulting in inferior performance. Personally I think all this talk about gameworks being some sort of cheating or blocking out AMD is nothing more than nonsense or a case of sour grapes. AMD should be proactive and resonsible for working with game developers to tweak their games for their GPUs. Further with 6GB of VRAM and a HDMI 2.0 port, it is somewhat more more UHD ready than Fury X.

AMD should drop $100 on the price to compete and maybe even make a Fury X version without the water cooling (even more preferable would be to drop 4 GB HBM in favour of 8GB GDDR).

 

[Aquinus]

Personally, the main problem that I see with this card is the use of HBM Memory. HBM with its current technical limitation (4GB) is not ready for the big time on a flagship card. AMD pitched this as the pinnacle of cards for UHD, but did nothing to support that. While lack of HDMI 2.0 port is one of those, the more important thing is the 4 GB of HBM that they put on this card.

What is the actual driving force for desiring UHD gaming? It is not just the screen resolution, When we have 4 times the pixel density of FHD, we would like to use those extra pixels to present more detail and that comes from higher and more detailed textures. What we are seeing as UHD gaming today is just the games being run at UHD resolution with the same textures that were designed for FHD. In place of one pixel in FHD, we have four pixels with the same shade in UHD which is quite pointless. The goal of UHD will be realized when we start getting more detailed textures to use for UHD and such textures are going to occupy a lot of VRAM.

No amount of PR about driver optimizations, texture compression and the high bandwidth offered by HBM would be able to side track the fact that 4GB is not going to be enough in the long run when higher resolution textures come into the picture. It maybe true that VRAM is not getting utilized efficiently today and that there might be ways to optimize the drivers to make the allocations better, but applies only for the present day situation when there is head room for such optimizations. Once true UHD optimized games with higher resolution textures start coming out, they simply will not fit in the 4GB VRAM and once swapping from main memory starts, you all know that the bottleneck is going to kill the performance.

Either AMD is banking on the fact that most multi platform games may not yet offer super high resolution textures for UHD in the PC versions or that for the games that do so, they can just compromise the texture quality at the driver or ask the developer to fallback to lower resolution textures for these cards.

Even if the card has the raw compute power for UHD, they have crippled this card by pairing it with 4GB HBM. I would have much rather preferred to have them pair this with 8 GB GDDR instead. They should have waited on the HBM till they could come out with 8 GB modules. Pair it with 8GB GDDR, replace the water cooling with the regular air cooled designs and market it at $500 or even $550 and this would have killed the 980 and 980Ti using the value for money tag. This was never a UHD ready card to begin with. I consider the 390X, to be more of a UHD ready card than Fury X.

There never was any need for them to be king of the hill in terms of performance, they could have claimed it in terms of value for money has they have done in the past. My last 3 GPU purchases and most of my overall GPU purchases were AMD for this reason, but this time, I went for GTX 980 after 10 years of not using an nVidia card because it offered more value for my money when I bought it. Personally I never card about Physx or any of the nvidia specific stuff, but the overall value justified the purchase.

Can Fury X compete with 980 Ti at the same price point? Not unless they beat the 980 Ti in 95% of the games with at least a 7.5~10% better performance margin which currently is not the case. 74% of the gaming GPU market share is currently owned by nVidia and a vast majority of AAA titles are using Gameworks features that add additional value for nVidia GPU users and it doesn't help that AMD fails to optimize their drivers and the game for their own GPUs and resulting in inferior performance. Personally I think all this talk about gameworks being some sort of cheating or blocking out AMD is nothing more than nonsense or a case of sour grapes. AMD should be proactive and resonsible for working with game developers to tweak their games for their GPUs. Further with 6GB of VRAM and a HDMI 2.0 port, it is somewhat more more UHD ready than Fury X.

AMD should drop $100 on the price to compete and maybe even make a Fury X version without the water cooling (even more preferable would be to drop 4 GB HBM in favour of 8GB GDDR).

I suspect memory bandwidth has very little to do with why it turned out the way it did. I think because of the performance numbers, it's not unlikely to say that the R9 Fury X is starved for ROPs.

 

[Lord Nemesis]

^^ Yes, the raw performance would definitely benefit from more ROP's. But my argument was specifically about the quantity of memory they put in because they forced themselves to use HBM on it. They have essentially crippled the card by choosing to go with HBM with its current technical limitation of 4GB instead of higher amount of GDDR.

HBM would actually make sense when games have vast amount of textures and other data to deal with where having higher bandwidths will help move or manipulate it faster. Having to reduce the quantity because of opting to go for HBM kind of defeated the purpose for having all that bandwidth in the first place. They should have reserved its use for the next refresh when they would have have larger sized modules of HBM like 8GB or 12GB.

I highly suspect if the current performance numbers for sub FHD or QHD would have been any different if GDDR had been used in place of HBM.

 

[HumanSmoke]

^^ Yes, the raw performance would definitely benefit from more ROP's. But my argument was specifically about the quantity of memory they put in because they forced themselves to use HBM on it. They have essentially crippled the card by choosing to go with HBM with its current technical limitation of 4GB instead of higher amount of GDDR.

The only problem with that scenario is that Fiji would be a totally different GPU. The only way AMD could get 4096 ALUs into the chip was because a large amount of the die space usually reserved for GDDR5 memory controllers and I/O is now able to be devoted to the core(s). IF Fiji was GDDR5, it would have just made sacrifices elsewhere. The increased power demand of GDDR5 would have meant the GPU would be clocked lower to compensate. A single Tonga chip has a 384-bit bus feeding 2048 cores/128 TAU/32 ROP. Doubling the core components but only increasing the bus width by a third (to 512-bit) starves the GPU of bandwidth, while going any larger puts the GPU die well outside manufacturability due to size limits of the lithography.

HBM would actually make sense when games have vast amount of textures and other data to deal with where having higher bandwidths will help move or manipulate it faster.

Well, that's just AMD being ahead of the curve. There's always a price to pay when a new technology arrives and its first iteration is not appreciably better than the incumbent. Put it down to the price of progress. AMD needed high bandwidth and low latency for HSA and to compete with Intel's own eDRAM and HMC roadmap. There is way more at stake here than just a consumer graphics card and AMD had to commit to HBM to accelerate its development.

 

[EarthDog]

1. [/quote]They have essentially crippled the card by choosing to go with HBM with its current technical limitation of 4GB instead of higher amount of GDDR.[/quote]No, they did not. On the cover, it looks like it, but, like NVIDIA with their "ZOMG 256 bit bus wth are you doing", they have better compression algo's. Not to mention they looked very deep at vRAM allocation and found, according to AMD, that 70% of vRAM is not used efficiently... so, its fine. You can see that it tends to pull away from 980ti at a higher res compared to low res too. If vRAM was a limit, that wouldnt be happening. Look at the 780ti vs 290x for example.

 

[Lord Nemesis]

^^ That is because none of these games have properly hit the VRAM limit so far. We currently don't have many games that max out 4 GB of VRAM, but 6 months down the line, there will that be new games that will do just that.

Further, this issue of inefficient usage of VRAM applies only to an extent. Are they seriously trying to tell us that each and every game wastes 70% of the memory that it allocates. Its more than likely the worst case scenario. Also, is AMD really going to hand tweak VRAM allocations for each individual game in their drivers. We all know how well AMD is handling their game specific driver tweaking currently. If AMD is so confident that their optimizations will allow the Fury X to run with just 4 GB where another card would require 5 or 6 GB, why didn't they release the 390X with 4GB and use those optimizations instead of putting 8GB VRAM on those cards. Nothing in their allegations of inefficiency and about their strategy for optimization suggests that it has to be memory technology specific. if inefficient memory usage is a issue and they have solution to optimize it at driver level, they should be able to do it for GDDR base cards as well.

Lastly, what about the case when a games compressed textures and other data go above the 4 GB limit beyond the walls for any optimization ? It should not be that difficult if a game has very high resolution textures to be used along with UHD resolutions.

One of the reviewers tested with GTA V @ UHD after tweaking the settings so VRAM usage goes over 4 GB and the Fury X dropped performance severely compared to 980 Ti at same settings. Once he tweaked the settings to make VRAM go above 6 GB, 980 Ti also dropped in performance severely.

 

[Vlada011]

Fury X is competitive against 980 Ti at very high resolution. Remember, AMD hasn't enabled DX11 MT drivers in Windows 8.1.

I rather see benchmarks done on Windows 10 i.e. Project Cars has frame rate uplift on R9-280 on Windows 10. Windows 10 forces DX11 MT.

I can't agree with that and most customers will not believe in that. Simply they want on paper more video memory... I afraid they will not believe to AMD that 4GB is enough for future.
I imagine someone who plan to buy 1440p monitor or example ASUS 3800R... 4GB is suicide. Than better to wait one year more when AMD offer Fury X with 8GB. They always make rebrand.
Chance of AMD is R9-390X 8GB. Not as single option, but for multi GPU that configuration become better than GTX980 SLI. Maybe is GTX980 better if someone have 600W PSU and 1080p and no plans for more.
But for anything more Hawaii CF is better. Only is problem price. If 4GB is not enough for Fury X than situation is even worse with GTX980. And that's segment hold biggest part of gaming community. People who pay under 500$. Better to say 300-400$ for graphic card because they will want more than 4GB. That's chance of AMD. NVIDIA can offer them only 600-650$ if someone want over 4GB. But someone to believe that Fury X will resist on higher resolution without fps drops next 2 years is very bad. Special because AMD owners keep graphic card longer than NVIDIA customers usually. Situation is even worse because NVIDIA dictate to developers what to do and 980Ti have 6GB, they plan 8GB for Pascal they will force games with more video memory to push people on upgrade and on TITAN X. NVIDIA 100% have plans to offer efficient Pascal with HBM, 8GB of video memory only little stronger than TITAN X but very expensive, and 8GB will be reason for upgrade from GTX980Ti, what people to do with 4GB. New cards from AMD will come for 18 months.

 

[btarunr]

This GPU with 8 GB 512-bit GDDR5 would've been 375-400W typical board power (educated guess). GCN 1.1 to GCN 1.2 wasn't as big a perf/Watt leap as Kepler to Maxwell. That's probably why this whole HBM adventure was unavoidable.

Edit. Now I'm really curious to know what this GPU would have been like with 8 GB 512-bit GDDR5.

 

[EarthDog]

^^ That is because none of these games have properly hit the VRAM limit so far. We currently don't have many games that max out 4 GB of VRAM, but 6 months down the line, there will that be new games that will do just that.

Further, this issue of inefficient usage of VRAM applies only to an extent. Are they seriously trying to tell us that each and every game wastes 70% of the memory that it allocates. Its more than likely the worst case scenario. Also, is AMD really going to hand tweak VRAM allocations for each individual game in their drivers. We all know how well AMD is handling their game specific driver tweaking currently. If AMD is so confident that their optimizations will allow the Fury X to run with just 4 GB where another card would require 5 or 6 GB, why didn't they release the 390X with 4GB and use those optimizations instead of putting 8GB VRAM on those cards. Nothing in their allegations of inefficiency and about their strategy for optimization suggests that it has to be memory technology specific. if inefficient memory usage is a issue and they have solution to optimize it at driver level, they should be able to do it for GDDR base cards as well.

Lastly, what about the case when a games compressed textures and other data go above the 4 GB limit beyond the walls for any optimization ? It should not be that difficult if a game has very high resolution textures to be used along with UHD resolutions.

One of the reviewers tested with GTA V @ UHD after tweaking the settings so VRAM usage goes over 4 GB and the Fury X dropped performance severely compared to 980 Ti at same settings. Once he tweaked the settings to make VRAM go above 6 GB, 980 Ti also dropped in performance severely.

There are plenty of games that will smash 4GB and 4K resolutions.

The issue isn't at the game level, it is at the API level. So it is game agnostic from what they said. As far as the second part of that paragraph, the 390x is a 290x with higher clocks and 8GB. IT doens't have the new algo's for compression AFAIK. I was at the release in L.A and talked with them about this.

What about compressed textures... it would do the same thing any other card does I would imagine in that it 'pages out' to system ram. ANd yes, that performance drop after running out of vRAM is normal... I believe I am missing your point there.

Edit. Now I'm really curious to know what this GPU would have been like with 8 GB 512-bit GDDR5.

Remarkably similar. Memory bandwidth isn't really a limiting factor until 4K +. Power draw would have been a bit higher though. Not as much as the estimates above guessed.

 

[CrAsHnBuRnXp]

Im not about to click through 13 pages, but has @xfia tried to come in and defend AMD yet?

In all seriousness though, for as much as I like nVIDIA for my video cards (not a fanboy just I like their features like Shadowplay and drivers are just all around better and more plentiful and I will go either side of the fence if performance is there), when will people realize that AMD just isnt what they were 10-12 years ago? They just cannot spank Intel/nVIDIA like they used to.

 

[Aquinus]

Im not about to click through 13 pages, but has @xfia tried to come in and defend AMD yet?

In all seriousness though, for as much as I like nVIDIA for my video cards (not a fanboy just I like their features like Shadowplay and drivers are just all around better and more plentiful and I will go either side of the fence if performance is there), when will people realize that AMD just isnt what they were 10-12 years ago? They just cannot spank Intel/nVIDIA like they used to.

Stop trying to start an argument, you're baiting him and you know it.

 

[Lord Nemesis]

My point is that paging from system RAM has more likelihood of happening sooner with a card equipped with 4 GB VRAM than one with 6 GB, 8 GB or 12 GB regardless of compression techniques and optimizations. Optimizing resource utilization is definitely important and I don't think nVidia ignores it either, but there would a limit for how much can be achieved through sheer optimizations. Optimizations are not always a replacement for having sufficient amount of resources.

Console games are often heavily optimized for the hardware they run on. But the lower quantity of resources like VRAM mean that after a point, they also end up having to make compromises that impact the final quality.

 

[CrAsHnBuRnXp]

Stop trying to start an argument, you're baiting him and you know it.

Im actually not trying to bait anyone. It is my legitimate viewpoint.

 

[Ikaruga]

3DMark's API overhead benchmark also test GPU's command I/O ports and pathways

Yes, but how is that relevant or argue with what I said? Command ports and pathways are needed to test the overhead of the api.

As I said, AMD cards will see more speed increase from dx12 (obviously since Microsoft makes it, who has not one but two consoles on the market with AMD GPUs), but we just don't know yet (or those who do are still under NDA) how much that speed increase will transfer to real life performance increases in games (so not talking about benchmarks here but actual game performances)... and - I believe - those who think that dx12 will magically make their GPU twice as fast gonna get a rough wake-up call.

 

[Caring1]

Im not about to click through 13 pages, but has @xfia tried to come in and defend AMD yet?

I believe he/ or she was given a little holiday recently

 

[EarthDog]

My point is that paging from system RAM has more likelihood of happening sooner with a card equipped with 4 GB VRAM than one with 6 GB, 8 GB or 12 GB regardless of compression techniques and optimizations. Optimizing resource utilization is definitely important and I don't think nVidia ignores it either, but there would a limit for how much can be achieved through sheer optimizations. Optimizations are not always a replacement for having sufficient amount of resources.

Console games are often heavily optimized for the hardware they run on. But the lower quantity of resources like VRAM mean that after a point, they also end up having to make compromises that impact the final quality.

And my point is that with their optimizations, it doesn't arrive as fast as you seem to expect. After a point, you are correct, but because of their optimizations, that point is further down the line than their previous generations.

 

[Aquinus]

My point is that paging from system RAM has more likelihood of happening sooner with a card equipped with 4 GB VRAM than one with 6 GB, 8 GB or 12 GB regardless of compression techniques and optimizations. Optimizing resource utilization is definitely important and I don't think nVidia ignores it either, but there would a limit for how much can be achieved through sheer optimizations. Optimizations are not always a replacement for having sufficient amount of resources.

Console games are often heavily optimized for the hardware they run on. But the lower quantity of resources like VRAM mean that after a point, they also end up having to make compromises that impact the final quality.

That's why I only recently started having issues with having only 1GB on my 6870s at 1080p, right? While I think you're right, I also think you're wrong. More VRAM is most definitely getting used and going into system memory for textures is a little different, because if the textures in memory aren't accessed often, you may be running at 60FPS (like I am in Elite Dangerous,) but occassionally get a blip because it was either used or it got paged in and something else paged out. It's still playable, but it's sometimes annoying if it happens at just the wrong time. It also depends on a game as well.

Either way, when push comes to shove, 4GB is plenty now and probably will stay that way for at least a couple years unless you're one of those people who shoves AA as high as it can go on your shiny new 4k display, in which case, I can't say you're the typical gamer.

 

[Initialised]

I wanted to want this card, but I think AMD should have had two SKUs, one with the CLLC and one with a full cover, single slot waterblock.

How well the card competes with the 12GB Titan-X at 4K shows that (for now at least) the extra 8GB are just for show.

 

[Xzibit]

 

[Bansaku]

And still my HD7950 in CFX performs better than the Fury X......

 

[Xzibit]

 

[tacosRcool]

I know Fury X is slower than the GTX 980 Ti but for the price you do get a water cooler. I say that isn't that bad taken that into account

 

[ensabrenoir]

not talking about any electric bill either..... that thing is border line evil.....gotta know what kinda numbers it puts out........

 

[Xzibit]

@TheMattB81

 

[Sihastru]

These days, 3D Mark doesn't say anything anymore. In actual games I think the fourth card usually means worse performance then three, and sometimes even then two cards!