AMD Fury X Reviews

[Glo.]

Sorry for post under another post, but this is important, however, I don't know if this has been posted already.

http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4

Fury X somehow only gets 387 GB/s of Bandwith. Now we have answer what is blocking the performance of this GPU.

Also it is a bit symptomatic, that 20% of clock increase on memory brings only 14% more performance. Something in the GPU is blocking it.

It may be drivers, it may be the design. But it is very interesting.

 

[StereoPixel]

oclmemtest on Fury X
http://www11.pic-upload.de/28.06.15/3v87mcvjenj.jpg

 

[ShintaiDK]

512MB with PCIe speeds?

 

[Glo.]

I have to say.


WTF?

Does this even make ANY sence?

 

[Despoiler]

Sorry for post under another post, but this is important, however, I don't know if this has been posted already.

http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4

Fury X somehow only gets 387 GB/s of Bandwith. Now we have answer what is blocking the performance of this GPU.

Also it is a bit symptomatic, that 20% of clock increase on memory brings only 14% more performance. Something in the GPU is blocking it.

It may be drivers, it may be the design. But it is very interesting.

TR should have done the tests at multiple resolutions. That's the only way you can tell if something like ROP is truly limited. In other words as you decrease the resolution the numbers should come way up. What's weird is we see the opposite in benchmarks. 4k Fiji is competitive. The numbers get worse as you drop the resolution. Some games it's barely above a 980 @ 1080p.

oclmemtest on Fury X
http://www11.pic-upload.de/28.06.15/3v87mcvjenj.jpg

This is bizarre. Why would OC'ing the memory do anything? If some subsystem like ROP was already maxed it shouldn't make any difference if the memory was OC. Maxed is maxed. You know what I mean?

There is something definitely up and so far no review site has done enough testing to suggest what it is. My gut feeling it's something with the drivers. Maybe the new memory management to deal with 4GB is way too aggressive and is actually slowing everything down.

 

[Glo.]

We need way more data, than we have right now. However, if this turns out to be the case...

I really hope its because something went terrible with drivers, not with the assembly of the GPU itself.

 

[ShintaiDK]

I guess this may also be related to what sontin found.

And it works like a charm:
Modor: https://www.youtube.com/watch?v=k_wNEOviaeE
Watch Dogs: https://www.youtube.com/watch?v=QEbVMUniJLM
GGGGGGGTA5: https://www.youtube.com/watch?v=8hnuj1OZAJs

https://www.youtube.com/user/sloppywetblow/videos

They havent even cared to optimize for these games at all. The driver is doing some fancy stuff, it is just unbelievable...

There is something wrong with Fiji memory management. The driver doesnt fill all of the 4GB of the ram. It makes a cut at ~3800mb or less.

He uploaded a video of AC:Unity and Fury X shows 3800mb vram usage. On a GTX980 it is 4000mb.

So either they havent optimized for games like Modor and Watch Dogs or their optimizations dont work in certain games.

PCPer shows Fiji's problem in GTA5: http://www.pcper.com/reviews/Graphics-Cards/AMD-Radeon-R9-Fury-X-4GB-Review-Fiji-Finally-Tested/Grand-Theft-Auto-V

A 290X has less spikes.

It would explain it if the last 512MB goes via PCIe.

But lets see, it may simply be some allocation by the driver or other issue.

 

[Enigmoid]

Looks like this shows several things.

Overhead is eating up bandwidth, either physical overhead or driver overhead.

As it stands, something is wrong with the last 0.5 GB of Vram. This could be seen earlier where Fiji seems to allocate 3.5 GB and no more in a lot of reviews.

 

[railven]

Looks like this shows several things.

Overhead is eating up bandwidth, either physical overhead or driver overhead.

As it stands, something is wrong with the last 0.5 GB of Vram. This could be seen earlier where Fiji seems to allocate 3.5 GB and no more in a lot of reviews.

I guess if 3.5GBs is enough for GTX 970 and no one cared, it should be the same for Fury X.


Welps, one thing I remember when I was a proud HD 7970 owner:
"Wait for the new memory management drivers, they will boost performance by a lot."

Yeah, that driver never materialized.

 

[Glo.]

I know its premature, but...


If Fury has this performance from reviews with only 3.5 GB of RAM and 387 GB/s of bandwith, what it will do with 4 GB, and 512 GB/s of bandwith.

We already know that jump from 16K to 19K graphics score in 3dMark Firestrike is due to OC. It is really getting interesting.

Maybe, after all AMD did not lied about the Fury to be the fastest GPU on the market.

We have to see it tho...

 

[tential]

I guess if 3.5GBs is enough for GTX 970 and no one cared, it should be the same for Fury X.


Welps, one thing I remember when I was a proud HD 7970 owner:
"Wait for the new memory management drivers, they will boost performance by a lot."

Yeah, that driver never materialized.

Are you comparing a GTX 970 at $350 or so having 3.5GB of Memory compared to 4GB,

to a FLAGSHIP card, that is AMD's foray into the ultra high end to compete with Titan, at a price of $650, having only 3.5GB, AFTER THE GTX 970 FIASCO

as being the same thing?

Because it's not.....

 

[tential]

I know its premature, but...


If Fury has this performance from reviews with only 3.5 GB of RAM and 387 GB/s of bandwith, what it will do with 4 GB, and 512 GB/s of bandwith.

We already know that jump from 16K to 19K graphics score in 3dMark Firestrike is due to OC. It is really getting interesting.

Maybe, after all AMD did not lied about the Fury to be the fastest GPU on the market.

We have to see it tho...

Great.... I was just about to buy R9 290s Crossfire. But I really can't knowing that the card may have been screwed up. Hopefully AMD gets their act together because luxury high end buyers dont like ridiculous screw ups like this. AMD needs to simply look at the competition and see how they manage launches and copy pasta.... seriously. Even hire old talent that's worked there before. Even with new talent at AMD, it just isn't cutting it. Something is screwing that company's capability to launch a product without a massive screwup.

 

[Despoiler]

I guess this may also be related to what sontin found.

It would explain it if the last 512MB goes via PCIe.

But lets see, it may simply be some allocation by the driver or other issue.

That would be a really silly way to do the memory management. As everything sits today the card will start swapping over PCI-E once it goes over 4GB anyways. Why make an artificial 3.5GB limit? We see in sontin's examples the card will use 3.5-3.8GB for several games. Then you have games with superior memory management that don't use over 3GB (gold star to these devs). We know AMD is looking at memory capacity in terms of usage. My guess is they are keeping track of what goes into memory, how much it's getting used, anything not being used gets put into system memory. It probably has some prioritization of what stays and what goes based on the type of data. ie some data types are more likely to be needed over time. It's that or they prioritize certain data types from the start and less likely to be used gets put into system RAM. A possible test for my hypothesis is whoever has a Fury X needs to look at your system memory usage. We can then compare system memory usage against other 4GB AMD cards running the same game and settings. Higher system memory usage for Fury X indicates this is what's happening. It still may show nothing because we need to see activity levels between the two memory pools. I highly doubt AMD will put this to rest. They aren't going to want to give up how their secret sauce works.

 

[RampantAndroid]

That would be a really silly way to do the memory management. As everything sits today the card will start swapping over PCI-E once it goes over 4GB anyways. Why make an artificial 3.5GB limit? We see in sontin's examples the card will use 3.5-3.8GB for several games. Then you have games with superior memory management that don't use over 3GB (gold star to these devs). We know AMD is looking at memory capacity in terms of usage. My guess is they are keeping track of what goes into memory, how much it's getting used, anything not being used gets put into system memory. It probably has some prioritization of what stays and what goes based on the type of data. ie some data types are more likely to be needed over time. It's that or they prioritize certain data types from the start and less likely to be used gets put into system RAM. A possible test for my hypothesis is whoever has a Fury X needs to look at your system memory usage. We can then compare system memory usage against other 4GB AMD cards running the same game and settings. Higher system memory usage for Fury X indicates this is what's happening. It still may show nothing because we need to see activity levels between the two memory pools. I highly doubt AMD will put this to rest. They aren't going to want to give up how their secret sauce works.

I linked to MSDN's SDK a month or more ago. Memory management is not at the driver level in most cases, according to Microsoft (they quite literally say that virtually no one will be doing memory management in the driver.) Management is done by DX and the game setting priorities on resources. Moreover, are you really telling me that AMD would take things out of VRAM and put them in system RAM where they could be paged out into virtual memory on the disk? That would be such an awful system for anyone with out 16GB of RAM lying around.

 

[TerionX6]

Hello. I registered to combat any further theorizing about this VRAM test issue.

Before I begin, here is a direct Google drive link to the oclMemBench test used. Tested, safe, works on my machine, Your Mileage May Vary. On another forum a user posted that the program crashes out immediately. Myself, I have to run the x32 package and not the x64.

https://drive.google.com/file/d/0B9OoHSmkeSeNNEE3ZkpsSWlnZ1k/view

As a foreword:
I know that DirectX reserves a portion of VRAM for system use when the OS/driver is not running a program which asks for full control of the GPU.
I know that the FuryX test in question was not run headless (without loading the OS, mucking up results) from seeing the screenshot.
I know that unless run headless (in most cases) every GPU will have results like these.

Linked are my own results with an HD7850, I assume under similar conditions as this FuryX test.

If this claim were true I have to assume that every graphics card is affected in a similar way. From memory, only cards such as the 970 and 660ti have true memory segmentation access issues.

Finally, I will say that this interpretation of the results either comes from a lack of knowledge of how a typical OS-space system operates, or (let's hope not) it is intentionally slanderous.

Thanks,
TerionX6

 

[Jhruska]

I wanted to echo what TerionX6 said. If this test isn't run on a headless system with Windows Aero effects disabled, you get errata in the final memory blocks on any GPU.

http://i.imgur.com/vrlKM6W.png

That link points to a test I ran on a GTX 980 Ti *without* putting it in headless mode with Aero disabled. You can see the slowdown in the last memory blocks. I confirmed that this also happens with the GTX Titan X as well. Again, note that it's the last 512MB of RAM being affected.

This is what happens when a bona fide issue meets a crude benchmark and limited understanding. People start looking to find demons everywhere, even when the problems don't exist. Just because the GTX 970 had a problem doesn't mean that every GPU does.

Further compounding the issue is this: Programs like GPU-Z and Process Explorer don't report how much VRAM is actually in use. They report how much VRAM the card has *requested*. DX11 doesn't offer the kind of monitoring control you need to know how much VRAM the GPU is actually using and there's no way to monitor it directly.

If the GPU has more RAM it may request to use more RAM, but that doesn't mean it's actually *doing* so. This is why you don't see performance gains from many titles, even when they show more VRAM as in-use. There's not a 1:1 mapping and the GPU can change the amount of memory its requesting to use from moment to moment, which is why many games will change the RAM allocation depending on what's going on, on-screen.

 

[antihelten]

Sorry for post under another post, but this is important, however, I don't know if this has been posted already.

http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4

Fury X somehow only gets 387 GB/s of Bandwith. Now we have answer what is blocking the performance of this GPU.

Also it is a bit symptomatic, that 20% of clock increase on memory brings only 14% more performance. Something in the GPU is blocking it.

It may be drivers, it may be the design. But it is very interesting.

Fury X only getting 387 GB/s isn't really a big issue.

First of all you shouldn't really look at the black texture results, since those are all kinds of weird (a 980 gets 28% higher bandwidth than peak?!) and not really representative of a real texture anyway.

If you look at the random texture results you will see that all of the tested GPUs get less than their listed bandwidth, usually in the 70-75% range. 780 Ti is the worst here at only 60% of listed bandwidth and Fury X is slightly below average at 65%.

So Fury X is not quite as high as one might expect (maybe 10% lower than expected), but it's not really significantly out of place either.

 

[Silverforce11]

The % increase from R290X is accurate, that test, all the GPUs get lower than their paper specs.

 

[railven]

Are you comparing a GTX 970 at $350 or so having 3.5GB of Memory compared to 4GB,

to a FLAGSHIP card, that is AMD's foray into the ultra high end to compete with Titan, at a price of $650, having only 3.5GB, AFTER THE GTX 970 FIASCO

as being the same thing?

Because it's not.....

No, that part about the 970/Fury X was sarcasm.

The other part about the unicorn Memory Management drivers that were gonna give HD 7970 some kind of performance boost is not.

IE, I wouldn't put faith in AMD delivering on it.

Great.... I was just about to buy R9 290s Crossfire. But I really can't knowing that the card may have been screwed up. Hopefully AMD gets their act together because luxury high end buyers dont like ridiculous screw ups like this. AMD needs to simply look at the competition and see how they manage launches and copy pasta.... seriously. Even hire old talent that's worked there before. Even with new talent at AMD, it just isn't cutting it. Something is screwing that company's capability to launch a product without a massive screwup.

You're better off buying this version of Fury X (if it gets better, kudos) over 290(x) CFX period. At least that's my opinion. I had 290X CFX for a good 2-3 hours before ripping it out of my computer. I mean, I was going to return the cards anyways but the microstutter is still very much visible.
290 CFX:
1440p Ultra + 4xSSAA + CMAA == 60 FPS
GTX 680:
1440p Good = 0xAA == 600 FPS

And I chose to return to the 680, even though I knew I'd have the 290's for another 3-4 days, because it was smoother. Two people tested the CFX system, two people felt the 60 FPS reading as false. Felt more like 30-40FPS.

 

[Silverforce11]

@railven
What games did you tested to conclude 680 SLI was smoother than R290 CF?

I have that setup. It's smooth in every game I've played. But I don't play any GameWorks titles, except for Witcher 3, which works fine with a few tweaks (tessellation override, temporalAA off).

 

[Enigmoid]

Comparison of the old and new bandwidth benchmarks on a 660m.

Stock

Memory Underclock

Mixed OC/Underclock

Core OC

Max OC

The last 128 MB chunk looks to be hardware reserved. First bench doesn't attempt to access, second bench returns errors.

If for instance I have multiple benchmarks open I can see low bandwidth in chunk 15.

 

[Jhruska]

You have to benchmark the dGPU in headless mode in order to get accurate memory bandwidth results from it (and you need to turn Aero off).

Otherwise any card is going to return memory slowdowns at the end. AMD, NV, 6xx, 7xx, HD 7900, Voodoo 5 6000, Bit Boys, MediaGX, Mach64, doesn't matter.

 

[railven]

@railven
What games did you tested to conclude 680 SLI was smoother than R290 CF?

I have that setup. It's smooth in every game I've played. But I don't play any GameWorks titles, except for Witcher 3, which works fine with a few tweaks (tessellation override, temporalAA off).

I didn't say SLI 680, I said one 680. And the games I tested were the games I play every day at this point: WoW, ArchAge, and FFXIV.

In regards to WoW personally, I found the Microstutter worse on 290X CF than what I remembered on HD 7970 CF. I mean, I put up with HD 7970 CF for months before I snapped at the lack of profiles (in that case at the time for FFXIV) but this time around I couldn't stomach it for more than a few hours.

 

[Enigmoid]

You have to benchmark the dGPU in headless mode in order to get accurate memory bandwidth results from it (and you need to turn Aero off).

Otherwise any card is going to return memory slowdowns at the end. AMD, NV, 6xx, 7xx, HD 7900, Voodoo 5 6000, Bit Boys, MediaGX, Mach64, doesn't matter.

Yeah I know, but optimus messes with that and I have to tinker which I don't feel like doing.

Nonetheless, it provides a contrast with the other results which were not run in headless mode.

 

[Despoiler]

I linked to MSDN's SDK a month or more ago. Memory management is not at the driver level in most cases, according to Microsoft (they quite literally say that virtually no one will be doing memory management in the driver.) Management is done by DX and the game setting priorities on resources. Moreover, are you really telling me that AMD would take things out of VRAM and put them in system RAM where they could be paged out into virtual memory on the disk? That would be such an awful system for anyone with out 16GB of RAM lying around.

The GPU driver can and does do memory management. AMD has already said they are going to change their memory management specifically to deal with 4GB and that it would be transparent to the game. They are going to more intelligently manage what gets stored in dedicated memory. Yes I suggesting they could dump the lesser used data to system RAM if 4GB capacity is being approached. Why would you use a virtual disk when system RAM is still many many times faster than SSDs? You ever heard of a RAM disk? I'm not suggesting they are using a RAM disk BTW. Also, why would you dump to a virtual disk if the data is already on the disk in the game install? The GPU can already read and write to system memory even if it has to use the OS. The OS shows you how much system memory it can use for graphics related tasks if it needs to. As this memory scheme is already in use the fact that you think 16GB is necessary is clearly false. Back to sontin's observations, the new management scheme might be too aggressive or too slow (overhead) and that's slowing the memory subsystem way down.