AMD Vega (FE and RX) Benchmarks [Updated Aug 10 - RX Vega 64 Unboxing]

[eek2121]

1. Stop listening to the marketing people. Their job description is literally to influence your perception and behavior.
2. You have no proof a die shrink alone would increase clock speed.

The only way a die shrink would NOT allow for increased clock speeds is if there were a barrier of some kind in the chip itself. Since there have been a few people getting the Fury to 1.6 GHz with exotic cooling, this is not the case for the Fury-X. I've never seen a case where a die shrink has not allowed for a clock speed increase. Halving the size of the die would half the power requirements, allowing for up to twice the clock speed with the same cooling.

Why is it that Pascal enjoyed such a huge performance boost over Maxwell? The 1080 had a 43% increase in clock speeds over the 1080, 32 more TMUs, and 512 additional CUDA cores. It was more than twice as fast as the GTX 980, all of this while only consuming 15 watts of additional power.

 

[Bouowmx]

More testing: https://twitter.com/GamersNexus/status/880825325252472832

Our order finally came in. Expect content today. #VEGA

 

[Bouowmx]

Or they could make lower clocked model, let's say around 1450 MHz. It would have similar performance as GTX 1080 and similar TDP as GTX 1080. And if it has similar price, than what's the actual problem?

The problem is AMD Vega Frontier Edition already uses ~300 W at ~1.4 GHz: PC Perspective power test using Metro Last Night 4K.

 

[Bacon1]

The problem is AMD Vega Frontier Edition already uses ~300 W at ~1.4 GHz: PC Perspective power test using Metro Last Night 4K.

Because its basically an OC'd Fury without any of the new features working. None of the architectural changes appear to be enabled from the testing done so far.

 

[Bacon1]

Here is someone with a Fury. Doesn't make any sense at all that Vega with TBR looks like the same as Fiji without. Fiji driver like months ago for Vega anyone?

https://gfycat.com/InsecureEagerKingbird
https://www.reddit.com/r/Amd/comments/6kdwea/vega_fe_doesnt_seem_to_be_doing_tiled/djleqwq/

Nice find, shows that Vega:

http://i.imgur.com/1vUrQ2K.gif

And Fury:


https://gfycat.com/InsecureEagerKingbird

Are both rendering similarlly so Tile based rasterization doesn't appear to be working / enabled. That is a huge perf/watt loss. We should see much better perf/watt once enabled.

Vs Maxwell:

https://youtu.be/Nc6R1hwXhL8?t=631

 

[Maverick177]

Why don't they test Vega at Fiji clock in gaming? Surely that's possible?

 

[Elixer]

This is why people should wait for SIGGRAPH.

An AMD official has stated that the drivers are not gimped, but they are older drivers. If the drivers are old, they could be missing architectural specific changes like tile based rendering. Worst case scenario is that the drivers only have a few tweaks to support Vega...in fallback mode. AMD may even be intentionally NOT releasing optimized drivers until RX Vega rolls out. I've read many cases where people claim that drivers cannot possibly account for a 30-50% speed boost...but if everything that makes Vega great is not implemented in the driver, then it's losing a ton of performance just by virtue of not having those features. PCPer themselves may have proved that the drivers aren't in great shape. They stated there was no difference between pro mode and gaming mode except a UI change. That tells me that something is definitely wrong. Gaming mode should be orders of magnitude faster because gaming drivers are optimized for performance, while pro drivers are optimized for accuracy.

First off, "Gaming mode" is a red herring. It doesn't do anything, except for enabling/disabling a few options for the card. It isn't a magic lane that you think it is.
There is no way that AMD would intentionally gimp the card this much, it wouldn't make any sense at all to do this.

The bigger problem that most everyone is missing is, this Vega release is intended for Pros (AMD's own words), and is listed under "Workstation" cards. http://support.amd.com/en-us/kb-articles/pages/workstation-graphics-drivers.aspx
Pros demand WHQL, stable drivers.
Looking at the Vega drivers on AMD's site right now,

Known Issues

• When running ReLive Pro, the cursor is not visible. The cursor is only visible when the user has switched to gaming mode
• When switching between Pro Mode and Gaming Mode, Timeout Detection and Recovery (TDRs) have been observed with multiple display configurations
• When numerous displays are used, there is an unlikely event where the driver does not complete installation
• During sleep and resume, after multiple instances, a blue screen hang has been observed
• Specific to Linux users, infrequent system hangs have been observed with 2015 Formula One
• During sleep and resume, a user may experience a TDR when instant replay is enabled
• Intermittent stuttering has been observed with Unigine Heaven in extreme cases
• System hangs have been observed in stressful situation in some applications when in CrossFire mode
• When using 8K displays, occasional situations with no display have been observed during sleep and resume
• Users may encounter Bluray playback issues on PowerDVD after switching to Game Mode
• Infrequent hangs have been observed during wake from sleep and resume states
• When resume from sleep, OpenGL applications have experienced TDRs
• In CrossFire configurations, blue screens have been observed during stressful ACPI testing\
• When rendering using Blender cycles, users may encounter infrequent iterations of poor performance
• With 10 bit displays, users may encounter anomalies in display results
• On Z170 chipsets, users may experience infrequent hangs from sleep states. Frequency may depend on system configuration.
• In extreme cases with stressful applications, the user may experience performance throttling in thermally stressful situations
• When running express install with multiple GPUs installed, the user may experience anomalies, including unexpected system reboots

Look at all the TDR errata, and talk about throttling / sleep issues.
10 bit displays problems? WTH, are they doing, again, this is a PRO card!
And what Pro is going to run Unigine Heaven, as if that means anything to them?

The only reason they did the launch was because they told investors they would, and would rather take the lumps like this instead of a lawsuit.

There is no sugar coating this, the card is not ready for prime time.
It was the wrong call.

They royally screwed up a PRO release with flaky drivers, which could cost them millions in the long term, there is no other way to look at this.
There is no sane professional that would use this card on production machines. Sure, they may test things with these cards, but, that is about it.

The pure profit is with Pro gear, not "enthusiast" gaming cards that might bring you a few $$$ in profit.

AMD can do better than this, and sure, the Vega RX release may be better for the "enthusiast" crowd (it can't get any worse), this is just another bad mark on AMD's marketing department, not to mention RTG's ability to deliver a product on time with stable drivers.

/end rant.

On the plus side, AMD's GPUopen (http://gpuopen.com) is pumping out releases.

 

[Phynaz]

The only way a die shrink would NOT allow for increased clock speeds is if there were a barrier of some kind in the chip itself.

The days of a strait shrink are long gone, that's why nobody does them anymore. Nothing is automatic anymore when it comes to process technology.

 

[Face2Face]

Official review is up

https://www.pcper.com/reviews/Graphics-Cards/Radeon-Vega-Frontier-Edition-16GB-Air-Cooled-Review

OC Results:

• Peak clock adjustment: +7% (resulting 1682 MHz rated clock)
• Set fan to 3000 RPM fixed, kept GPU under 80C
• Resulting clock speed was about 60 MHz higher than stock settings
• Set power target to +20%, was drawing ~350 watts

 

[Phynaz]

Because its basically an OC'd Fury without any of the new features working. None of the architectural changes appear to be enabled from the testing done so far.

If the new features aren't working as you claim, then they will be power gated. Which means power consumption will go up if those features are enabled. Such as when a CPU runs AVX code.

We know what Vega is, stop trying to deny it.

 

[Peicy]

Official review is up

https://www.pcper.com/reviews/Graphics-Cards/Radeon-Vega-Frontier-Edition-16GB-Air-Cooled-Review

Thanks. Reading now.

 

[QualityTime]

That's what I'm wondering too. There are some great and welcome changes but why those do not show? On paper everything looks great but somehow that doesn't translate to real word performance. This is basically a tweaked GCN so I don't think that drivers should be a big problem. At least at the moment with this card. Or maybe they should just completely bin GCN for good and start from scratch.

There is no "on paper" at the moment. AMD haven't released anything showing layout or implementation details. And with geometry binning seemingly off by default, we can't ferret out any information on it.

Vega the architecture is not the same thing as Vega 10 the chip.
It's possible that the terrible performance observed from Vega 10 is due to bottlenecks like having only 4 shader engines (should be 8) and 64 ROPs (should be at least 96, if not 128). Raven Ridge then would not have these problems, nor would a smaller Vega 11 chip with a more balanced layout. Still, it's hard to imagine why they would not have fixed problems that were well known in 2015 when Fury X was released. Even Raja tacitly admitted (in an interview with TechReport) that there were bottlenecks, pleading the 28nm reticle limit as an excuse. Now that's gone and the transistor budget went way up, but the exact same bottlenecks remain... how does that happen?

Lets use Polaris since its the most recent known quantity. The theoretical triangle throughput at 1,200mhz is 4.8 billion per second. Enough to render a triangle per pixel ~580fps at 4K. Add in the improvements to things like geometry caching and rejection, the shader engines themselves are fine.

The problem with GCN's front end scaling and efficiency has to do with the rasterizer. It needs to be decoupled and have the one per shader engine limit removed. Additionally scale the width down from the current 16ppc.

And I'm not sure adding more ROPS is going to help much given GCN's memory usage. For example, the RX 480 is bandwidth limited and slower than the GTX 1070, but they have the same memory interface. It would be better to spend the transistor budget on improving rendering/memory locality. Perhaps implementing a binning/tiling system, but who knows how that would turn out.

AMD is now more than a full generation behind Nvidia and they seem to be regressing. This is Bulldozer-level worrying - they need to purge the R&D team and start over from a clean sheet of paper.

Great idea. Spend at least 2-3 years to rebuild the team and get a new design into testing. During that time they'll have nothing shipping. Maybe hit tapeout after just losing the next-gen console contracts.

 

[Cloudfire777]

OH MY GOD what a complete failure Vega is.

PCPer review showed it was between GTX 1070 and 1080. You heard right...Jesus!
To top it all, power consumption of Vega is 280W while GTX 1080 is 180W.

This is another bulldozer. This is what they have been working on for 2-3 years.
They freakin did it again


https://www.pcper.com/reviews/Graph...B-Air-Cooled-Review/Clock-Speeds-and-Power-Co

 

[EXCellR8]

still comparing the FE big vega to gaming cards though... there is still hope for RX

 

[Cloudfire777]

still comparing the FE big vega to gaming cards though... there is still hope for RX

Barely GTX 1080 performance for RX Vega if PCPer is right...
This is what AMD decided to use HBM2 on, which even caused further delay?

Rubbish AMD. Absolutely rubbish decision making.

There is still potential that AMD will be able to find some additional performance in the product before the consumer launch, either in drivers and efficiency improvements, higher clocks, or maybe better binning. At most, I could see AMD adding another 10% to the results we see with an odd-case of a significantly under-developed part of the driver stack going beyond that

 

[EXCellR8]

speculative... plus that suggests reference cards. i'll believe it when I see it

don't get me wrong, FE is definitely a bit of a disappointment but RX could salvage Vega if the power and clocks are more reasonable. oh, and price of course.

 

[crashtech]

It's Fury all over again. Not a bad product, just too little, too late. Now the hopeful can begin waiting for a breakthrough with Navi, because Vega isn't it.

 

[Bacon1]

Look at all the TDR errata, and talk about throttling / sleep issues.
10 bit displays problems? WTH, are they doing, again, this is a PRO card!
And what Pro is going to run Unigine Heaven, as if that means anything to them?

Did you actually read the issues though?

• When switching between Pro Mode and Gaming Mode, Timeout Detection and Recovery (TDRs) have been observed with multiple display configurations

It is live reloading drivers without a reboot. That's something that afaik hasn't been done, and NVidia doesn't even let you use different drivers on their pro/gaming cards.

• During sleep and resume, after multiple instances, a blue screen hang has been observed

This one is an actual issue, but it doesn't sound frequent just that they've had it happen at least once and marked it.

• Specific to Linux users, infrequent system hangs have been observed with 2015 Formula One
• During sleep and resume, a user may experience a TDR when instant replay is enabled
• Intermittent stuttering has been observed with Unigine Heaven in extreme cases

Three gaming specific and infrequent issues

Most of the rest are related to sleep / resume situations and don't happen during normal usage.

 

[[DHT]Osiris]

It's Fury all over again. Not a bad product, just too little, too late. Now the hopeful can begin waiting for a breakthrough with Navi, because Vega isn't it.

I'm concerned that they're slipping back faster though. They're already basically a gen back, this should have come prior/around the timeframe of the 1080 release, to be competitive somewhere around the 980ti-1080 range. We're looking at the 20x0 series what, in another quarter or two? We'll be another gen and a half down *after that* before we're looking at Navi, when AMD will finally be competative what, with the 1080ti range?

Something's gotta step up, I feel like AMD's officially spread too thin between CPU and GPU to maintain.

 

[Bacon1]

If the new features aren't working as you claim, then they will be power gated. Which means power consumption will go up if those features are enabled. Such as when a CPU runs AVX code.

We know what Vega is, stop trying to deny it.

Uhhh you do realize that the architectural features missing like tiled rasterization are perf/watt performance increases right? So they will decrease power usage not increase it, or at least increase performance a lot while keeping power usage the similar.

 

[Cloudfire777]

It's Fury all over again. Not a bad product, just too little, too late. Now the hopeful can begin waiting for a breakthrough with Navi, because Vega isn't it.

Fury wasnt late nor a failure. It launched like 2 months after 980Ti and it matched the 980Ti plus had better build quality and cooling. The only downside was overclocking which 980Ti was good at. But for non overclockers it was a great card.

Vega however is a failure on all aspects.

 

[crashtech]

I'm concerned that they're slipping back faster though. They're already basically a gen back, this should have come prior/around the timeframe of the 1080 release, to be competitive somewhere around the 980ti-1080 range. We're looking at the 20x0 series what, in another quarter or two? We'll be another gen and a half down *after that* before we're looking at Navi, when AMD will finally be competative what, with the 1080ti range?

Something's gotta step up, I feel like AMD's officially spread too thin between CPU and GPU to maintain.

I don't think it's in them to do more, really. So hopefully the price is right.

 

[crashtech]

Fury wasnt late nor a failure. It launched like 2 months after 980Ti and it matched the 980Ti plus had better build quality and cooling. The only downside was overclocking which 980Ti was good at. But for non overclockers it was a great card.

Vega however is a failure on all aspects.

It couldn't overclock at all. I mostly go for AMD GPUs, but to me that was a terrible disappointment. Notice I did not say failure, you said that.

 

[[DHT]Osiris]

I don't think it's in them to do more, really. So hopefully the price is right.

With the size of that die, unless they've done something remarkable on the manufacturing side I don't see the yields being so high that they can push this out this series competitively without taking a serious loss. As pointed out earlier, they've got until NV gets the 2060/2070 series out the door at a $250-$400 price point to make back their R&D before whatever RX card this is based on gets murdered in its sleep.

 

[Bacon1]

Sure looks like a great perf/$$ PRO GPU!

Maybe wait for the Gaming drivers and RX before complaining about its gaming performance??