AMD Radeon RX Vega 64 and 56 Reviews [*UPDATED* Aug 28]

[Olikan]

I think it was mentioned above that it was the ROP's that do MSAA but all the hardware stuff is mostly over my head. This is why i focus on reviews and benchmarks that laymen can understand. Real world performance i understand.

I mean, it makes sence, if the geometry units are underdelivering, then vega rops are losing cicles waiting for data.

 

[Headfoot]

TXAA is just as bad as FXAA, if not worse:
https://www.gamespot.com/forums/pc-mac-linux-society-1000004/txaalarge-sheet-of-plastic-29319415/

These post process AA modes "fix" AA in everything, because they treat the screen like bitmap image and just blur everything.

There are essentially various flavor of the month blur filters. Cheap to implement, so often used instead of the superior but more expensive MSAA.

No.
There are many different SMAA/TAA/TXAA (temporal anti-aliasing) implementations, you can't paint them all with the same brush. It's not a universally similar algorithm like MSAA. Some TAA implementations are better than MSAA having seen both, and especially if you do cost benefit. Like i've already said - Cryengine 3 (Ryse) and The Division, amongst others. No doubt there are bad implementations of it, I have seen some myself, but there are very good implementations of it too. TAA can have better temporal accuracy than MSAA as well as de-alias surfaces/objects when MSAA couldn't.

See: https://de45xmedrsdbp.cloudfront.net/Resources/files/TemporalAA_small-59732822.pdf
See also: http://www.crytek.com/download/Sousa_Graphics_Gems_CryENGINE3.pdf

And an older reference that talks about all the methods generally http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.306.8449&rep=rep1&type=pdf

 

[PeterScott]

https://www.hardocp.com/article/2013/03/12/crysis_3_video_card_performance_iq_review/9

This last screenshot demonstrates up close how detail is lost in the high resolution textures due to the blurring. The stone wall simply has less detail with TXAA enabled. Also, the vegetation here is extremely blurry with TXAA enabled.

Overall, TXAA should be avoided. We don't know how this quality of TXAA got implemented into the game and left there. It seems like somebody took the day off when quality control of TXAA image quality rolled around that day. TXAA destroys the gameplay experience in Crysis 3.

I am specifically talking about TXAA/FXAA.

BTW, you had Crysis 3 as a example of good implementation before you changed your post.

 

[Headfoot]

https://www.hardocp.com/article/2013/03/12/crysis_3_video_card_performance_iq_review/9


I am specifically talking about TXAA/FXAA.

BTW, you had Crysis 3 as a example of good implementation before you changed your post.

Crysis 3 does have a good implementation of SMAA, so I'm not sure what you're talking about. nVidia's canned TXAA is a totally different algorithm

edit: one of the links above is broken. Here is the whitepaper about crysis 3's (cryengine 3) SMAA, which is defintively one of the best implementations to date. https://iryoku.com/smaa/downloads/SMAA-Enhanced-Subpixel-Morphological-Antialiasing.pdf

 

[Headfoot]

To the degree your point is "Post AA always bad" you're wrong. The reality is there are good implementation and bad ones. You can't say they are all bad and be correct.

 

[wilds]

I've seen great, efficient AA methods, and bad ones.

I do agree that The Division has great low overhead AA, while Rainbow 6, another Ubisoft published title, has the worst AA implementation I have ever seen!

Elite: Dangerous is another game that does an excellent job at providing very unobtrusive antialiasing that does not 'blur' or degrade image quality that much.

All in all, the best form of ridding aliasing I have seen is to simply game on a higher native monitor resolution unfortunately...

 

[Hitman928]

Hitman928 with rx 56 with Custom with -1015 mV and +50% Power score 6839 in timespy graphics score.
I get aprox 6980 on a rx 64 using same settings
With same memory clock on the rx56 its more or less precisely the same score.

I dont know if Hitman score is specifically the graphics but mine is and i guess his is too.

Yep, graphics only. If I bump the memory up to 945 MHz (the same as Vega64) I can hit ~7200 in timespy.

 

[MajinCry]

TAA and FXAA make things way too blurry. The ideal would be deferred MSAA (requires a G-Buffer aka Normals Buffer), with SMAA on top for any transparent textures.

 

[krumme]

Yep, graphics only. If I bump the memory up to 945 MHz (the same as Vega64) I can hit ~7200 in timespy.

Can there be a cpu reason for that?
I asume not and thats why i bring the numbers. And can verify with a rx 56 in same setup in a week.
There can actually be a few reasons the rx56 score ends 200 points higher.
But imo this test is interesting because it shows a serious bottleneck that we have seen before.
And its not power or temp as i can tell.

 

[Rasterizer]

Mmm... So, the "MSAA bug", might be a geometry bottleneck?

Thank you for asking this question, as it sent me off into researching how MSAA works. As it turns out, apparently MSAA literally works by performing parts of the rasterization process at a higher resolution than screen resolution, and thus appears to be rasterization rate limited:

In terms of rasterization, MSAA works in a similar manner to supersampling. The coverage and occlusion tests are both performed at higher-than-normal resolution, which is typically 2x through 8x.

The linear decline of Vega performance relative to Pascal as MSAA multiplier increases thus appears to be support the premise that Vega is being polygon throughput bottlenecked by its four geometry engines in the absence of a working implementation of primitive shaders. Theoretically at least, this should mean that if primitive shaders are able to substantially increase Vega's polygon throughput they should also improve Vega's MSAA performance.

 

[PeterScott]

Crysis 3 does have a good implementation of SMAA, so I'm not sure what you're talking about. nVidia's canned TXAA is a totally different algorithm

edit: one of the links above is broken. Here is the whitepaper about crysis 3's (cryengine 3) SMAA, which is defintively one of the best implementations to date. https://iryoku.com/smaa/downloads/SMAA-Enhanced-Subpixel-Morphological-Antialiasing.pdf

Where have I said anything about SMAA? I said FXAA was blurry crap, then someone brought up TXAA, and I said that was also blurry crap.

 

[IRobot23]

Why do you guys wants Raja to quit? Isn't he really good?

I am a gamer, so FPS is all it matters to me.
I have heard many rumors about power, but if you are an overclocker you know that GPUs from AMD are well overvolted.


1. We know that architecture is important, but software could be much more...

2. We all know that Polaris is very competitive, about power, well they could optimized their voltage and lower their clock to get that ~120W and GTX 1060 6GB (stock performance).

3.They could simply do larger chip with GCN4 architecture and GDDR5X for lower cost and higher availability and higher performance... we all know that ~500mm^2 GCN4 chip would be able to reach GTX 1080TI performance.

So, why did the take a risk with VEGA? Did they do a step forward to early? Do they want to force optimization of consoles into PC? Infinity fabric?... well maybe is for dual GPUs... Thinking to much, lets us see what are they preparing.

 

[PeterScott]

2. We all know that Polaris is very competitive, about power, well they could optimized their voltage and lower their clock to get that ~120W and GTX 1060 6GB (stock performance).

Polaris consumers power like a GTX 1070 and performs like a GTX 1060, so it is no winner on perf/watt.

You act like undervolting is a sure thing and it can only be done for AMD cards. It is neither of those things. You can undervolt NVidia cards as well, and it is hit and miss for both company products.

 

[IRobot23]

Polaris consumers power like a GTX 1070 and performs like a GTX 1060, so it is no winner on perf/watt.

You act like undervolting is a sure thing and it can only be done for AMD cards. It is neither of those things. You can undervolt NVidia cards as well, and it is hit and miss for both company products.

Not that much, believe me.. stock RX 480 is running 1266MHz at 1.15V, I can RUN furmark at 1.115V at 1300MHz and for gaming usually 1266MHz around 0.975-1.025V 1265mv. I have really bad ASIC quality. Basically AMD cards can run at stock voltages Furmark @Boost clock, if cooler is capable for it.

RX 480 runs best at 1200MHz ~950mV or less. It will use less than 110W.

 

[tamz_msc]

TAA and FXAA make things way too blurry. The ideal would be deferred MSAA (requires a G-Buffer aka Normals Buffer), with SMAA on top for any transparent textures.

Eh, it depends on the game. I've had better results with driver-enforced FXAA on Gamebryo/Creation Engine games, for example, compared to MSAA.

Slightly off topic, but I think that your draw call test should be most appropriate with the Strip in FNV. The way the performance tanks in that location compared to the rest of the game is a sight to behold.

 

[nurturedhate]

Polaris consumers power like a GTX 1070 and performs like a GTX 1060, so it is no winner on perf/watt.

You act like undervolting is a sure thing and it can only be done for AMD cards. It is neither of those things. You can undervolt NVidia cards as well, and it is hit and miss for both company products.

You criticize the poster for assuming a black/white situation (undervolting is a given, only AMD) and then follow it up with a black/white situation of your own (Nvidia undervolting = AMD undervolting). This type of debate really doesn't work past the 4th grade, nice try though. The thing is Nvidia does a much better job setting proper voltages for their chips and their voltage adjustments/clocks are far superior. On average you will gain far less from undervolting Nvidia (performance vs power saved) compared to AMD. Furthermore, you have to know this. It's pretty well known.That only leaves your post and its intentions.

 

[AtenRa]

Polaris consumers power like a GTX 1070 and performs like a GTX 1060, so it is no winner on perf/watt.

You act like undervolting is a sure thing and it can only be done for AMD cards. It is neither of those things. You can undervolt NVidia cards as well, and it is hit and miss for both company products.

He said competitive, not perf/watt winner.

 

[IRobot23]

He said competitive, not perf/watt winner.

Yeah,
I think that has something to do with $$$.

 

[Muhammed]

MSAA is still used in half the titles of DX12 as well as DX11. It's also the defacto AA in VR applications and games, and it has been a staple in graphics realm for decades. Don't understand this sudden urge to downplay it's importance now. Being a Vega apologist is one thing, and being technically obtuse is another.

 

[PeterScott]

He said competitive, not perf/watt winner.

It's not really competitive on Perf/Watt either.

 

[Arachnotronic]

So, why did the take a risk with VEGA? Did they do a step forward to early? Do they want to force optimization of consoles into PC? Infinity fabric?... well maybe is for dual GPUs... Thinking to much, lets us see what are they preparing.

Yes, great idea. People on these forums and elsewhere kept saying "Wait for Vega" for more than a year, and what did it get anybody? Just time spent NOT playing games.

Judge by the here and now, not the fantasies that some people cook up.

 

[IRobot23]

Yes, great idea. People on these forums and elsewhere kept saying "Wait for Vega" for more than a year, and what did it get anybody? Just time spent NOT playing games.

Judge by the here and now, not the fantasies that some people cook up.

You are right, but I cannot take market share from nvidia in year. We saw what happende with GCN vs Kepler. GCN was superiorly advanced yet it didn't helped them much. Nvidia did spend more $ and took time for maxwell.

 

[PeterScott]

You criticize the poster for assuming a black/white situation (undervolting is a given, only AMD) and then follow it up with a black/white situation of your own (Nvidia undervolting = AMD undervolting). This type of debate really doesn't work past the 4th grade, nice try though. The thing is Nvidia does a much better job setting proper voltages for their chips and their voltage adjustments/clocks are far superior. On average you will gain far less from undervolting Nvidia (performance vs power saved) compared to AMD. Furthermore, you have to know this. It's pretty well known.That only leaves your post and its intentions.

I wrote that they were both hit and miss, not that they were equal.

But arguing that AMD usually benefits more from undervolting than NVidia, is like arguing fat people are better at losing weight than skinny people. NVidia is very power efficient to start with, so of course the gains on that front are going to be smaller.

 

[IRobot23]

I wrote that they were both hit and miss, not that they were equal.

But arguing that AMD usually benefits more from undervolting than NVidia, is like arguing fat people are better at losing weight than skinny people. NVidia is very power efficient to start with, so of course the gains on that front are going to be smaller.

Its not like that.

 

[MajinCry]

Slightly off topic, but I think that your draw call test should be most appropriate with the Strip in FNV. The way the performance tanks in that location compared to the rest of the game is a sight to behold.

I considered doing that, but I got enough flak for using a D3D9 draw call benchmark, so I figured that Fallout 4 would be more appropriate. Shame that thread didn't get any traction, still waiting for someone to post their result with only 1 CCX being used.

Another good area would be the DC ruins in Fallout 3. I recall the game issuing 7.5k draw calls at one point.