AMD Radeon R9 Fury X Reviews

[Rys]

@kalelovil said:

How's the status on Anandtech's review?

Ryan's last tweet about the subject was 2 days ago, but the twitter feed from everyone at anandtech has been awfully quiet..

He just tweeted:
27pages,25Kwords,uploadingnow
Imagesandtablestakeabitoftime.Pleasedont break my server. The RSS feed will let you know when itsup

 

[Rys]

@3dilettante said:

As it turns out open or closed test cases have the annoying whining.

http://www.pcper.com/reviews/Graphics-Cards/Retail-AMD-Fury-X-Sound-Testing-Pump-Whine-Investigation

It also appears that AMD may have sent Golden Samples (less noisy ones) to reviewers.

The individual reports are pretty variable. The odds that a very limited number of review samples would have been the worst out of a larger batch are pretty low, and only a very few have gone through comparative sound recording analysis.
The exact nature of the sound profile between samples looks to vary as well, so there may be cards that might have less perceptible problems or users with differing hearing ranges.

 

[Rys]

@Rys said:

He just tweeted:
"27 pages, 25K words, uploading now"
"Images and tables take a bit of time. Please don't break my server. The RSS feed will let you know when it's up"

http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review

 

[Rys]

@pharma said:

Ryan,

Nice review!
One question .... did you retest the 980ti or are these benchmarks results from your original 980ti review? Just wondering about the 352.90 drivers as these aren't listed in the NV driver download site as either beta or whql. The only two drivers recommended for the 980ti are the 353.06 and the 353.30.

 

[Rys]

@Ryan Smith said:

Ryan,

Nice review!
One question .... did you retest the 980ti or are these benchmarks results from your original 980ti review? Just wondering about the 352.90 drivers as these aren't listed in the NV driver download site as either beta or whql. The only two drivers recommended for the 980ti are the 353.06 and the 353.30.

Original results. There isn't anything in those drivers that impacts the games we test, so the original results are still valid. They're all the same branch and the later drivers just include game-specific fixes.

 

[Rys]

@lanek said:

Nice to see that you still goes deep in the technical details about the architectures, not like most reviews who just put specifications, 3-4 photos of the gpu's and the benchmarks .

 

[Rys]

@Tridam said:

So as some people already pointed it out, it's actually possible to overclock the memory on the Fury X. I'm not sure what exactly is included in this clock domain though. Anyway it works after enabling Extendofficialoverclockinglimits in Afterburner. I could change the memory clock and the option actually appeared in the CCC Overdrive panel.

I couldn't push the memory higher than 550 MHz. It's 'stable' at 575 and 600 MHz but there are artifacts. I settled on 1134 / 540 Mhz to get the same 8% overclock on the memory and on the GPU (1134 MHz is just stable enough for benchmarking, the GPU prefers 1125 MHz). I've just published the results there : http://www.hardware.fr/articles/937-26/overclocking-gpu-fiji.html

 

[Rys]

@3dilettante said:

The power savings from HBM are in the ballpark of what was speculated.
It seems that Fiji does have adaptive voltage in, which was good for 5-10% elsewhere.

I didn't quite expect that level of power increase in the 40-65C temperature range. Given the non-linear behavior at increasing temps, some of the earlier power savings speculation may have been conservative.

It doesn't seem like adaptive voltage has done enough to make the GPU tolerant of voltage droop in an overclocking attempt, which I was curious about.
Perhaps only critical portions like the CUs but not all of the GPU hardware is set to downclock when it happens, hence the artifacting and instability. If the biggest consumers found their margin needs reduced, it might make some of the background hardware's sensitivities more evident. Otherwise, it may be the case that the method of operation for that scheme is calibrated to the standard ratings and doesn't work as effectively outside of them.

 

[Rys]

@pjbliverpool said:

So as some people already pointed it out, it's actually possible to overclock the memory on the Fury X. I'm not sure what exactly is included in this clock domain though. Anyway it works after enabling Extendofficialoverclockinglimits in Afterburner. I could change the memory clock and the option actually appeared in the CCC Overdrive panel.

I couldn't push the memory higher than 550 MHz. It's 'stable' at 575 and 600 MHz but there are artifacts. I settled on 1134 / 540 Mhz to get the same 8% overclock on the memory and on the GPU (1134 MHz is just stable enough for benchmarking, the GPU prefers 1125 MHz). I've just published the results there : http://www.hardware.fr/articles/937-26/overclocking-gpu-fiji.html

Interesting results, thanks. So it seems there is a small memory bandwidth bottleneck at play but the GPU is overall more core limited.

 

[Rys]

@trinibwoy said:

So as some people already pointed it out, it's actually possible to overclock the memory on the Fury X. I'm not sure what exactly is included in this clock domain though. Anyway it works after enabling Extendofficialoverclockinglimits in Afterburner. I could change the memory clock and the option actually appeared in the CCC Overdrive panel.

I couldn't push the memory higher than 550 MHz. It's 'stable' at 575 and 600 MHz but there are artifacts. I settled on 1134 / 540 Mhz to get the same 8% overclock on the memory and on the GPU (1134 MHz is just stable enough for benchmarking, the GPU prefers 1125 MHz). I've just published the results there : http://www.hardware.fr/articles/937-26/overclocking-gpu-fiji.html

The hitman, battlefield and batman results are really strange. It's almost as if Fiji is perfectly balanced in those titles. Weird.

 

[Rys]

@gamervivek said:

Meanwhile (wccftech claim that) Hallock claims HBM overclocking is doing nothing. :smile:

UPDATE : We’ve confirmed with Robert Hallock, technical PR lead at AMD, that while the GPU-Z tool is reporting an increase in memory frequency in reality the frequency did not change. As HBM’s frequency in the Radeon R9 Fury X is determined in hardware and cannot be changed through software.

And I doubt that's just broken AF on Titan X.

On a Titan X?

AMD and their mantle shenanigans.

 

[Rys]

@trinibwoy said:

So if the memory clock didn't actually increase what did Damien measure!?

 

[Rys]

@gamervivek said:

Dunno, but folks are breaking 20k on firestrike now with oc'ed HBM.

Spoiler

 

[Rys]

@CarstenS said:

On a Titan X?

Find the irony.

 

[Rys]

@Nemo said:

Fury X Mantle Numbers

http://www.golem.de/news/grafikkart...mantle-schnittstelle-flotter-1507-115005.html

 

[Rys]

@entity279 said:

Damien alluded in his post above to the fact that we don't know memory OC slider does. It increases some clock that belongs to a domain that is related somehow to the memory interface. That's all we got. His results of course prove that the slider itself impacts performance in a statistically significant way.

 

[Rys]

@fellix said:

Damien alluded in his post above to the fact that we don't know memory OC slider does. It increases some clock that belongs to a domain that is related somehow to the memory interface. That's all we got. His results of course prove that the slider itself impacts performance in a statistically significant way.

Maybe the controllers and the interface are locked to the HBM's clock domain? That would also affect the L2 performance, I guess.

 

[Rys]

@Tridam said:

I haven't spent much time playing with what is exposed as memory clock but I had a quick look at pixel rates. I could see that the peak pixel rates (RGBA8, RGBA16F, no blending) scale almost perfectly with the GPU clock while the pixel rates in bandwidth intensive modes (RGBA32F with blending) scale almost prefectly with the memory clock. Here are some fillrate results with RGBA32F and blending enabled :

Fury X ref clocks : 13.5 Gpixels/s
Fury X GPU +10% : 13.6 GPixels/s
Fury X mem +10% : 14.7 GPixels/s

Fiji acts here the same way as any other AMD GPUs. At this point I can't see any sign that this memory clock is any different than the memory clock of other AMD GPUs.

 

[Rys]

@Dave Baumann said:

Meanwhile (wccftech claim that) Hallock claims HBM overclocking is doing nothing.

Attribution to that is being chased up.

MCLK overclocking via afterburner is working, but apparently the granularity of the steps is a little coarser than the software tools will state, with the final value being rounded to the closest step.

 

[Rys]

@gamervivek said:

Dave saves!!

More overclocked benches, however not as dramatic as that firestrike increase(which might be tessellation enhanced).

http://www.overclock.net/t/1547314/...nano-x-x2-fiji-owners-club/2080#post_24123030

and 11% improvement in firestrike ultra.

http://www.3dmark.com/compare/fs/5276383/fs/5256512

http://www.overclock.net/t/1547314/...nano-x-x2-fiji-owners-club/2050#post_24121077