[DX12] Fable Legends Beta Benchmarks

[railven]

Railven zlatan has some access to people in the industry judging from him getting to see nano before release as well.

But that Said, wait for more results/info

I get that he works in the industry, which is an issue in itself. It reminds me of the shill hunt (no I am not implying he is a shill), I get various posters here work for various parts of the industry, but you don't see them drawing targets on their backs. More so you don't see them hand feeding some of our residents "odd, my results are this" posts which contradict everything we've seen then disappear into the night.

Did you not see the conspiracy theory literally spring up on the same page of his post "DID NV GET THEM TO TURN OFF AS!!!!"

We're back to an insider saying something, no verification and people are running with it. Didn't we just have a "Maxwell2 can't do Asynch Compute" scare?

 

[RussianSensation]

Never had any issues with balanced settings. My GPU and CPU ramp up as they should and there is 0 performance difference for me in games or benchmarks.

That doesn't mean there are no cases/games where there is no penalty. I've documented this myself because I run various programs that peg my laptop's CPU to 99-100% and it does matter. Since you can't possibly test every single game/benchmark, the only way to guarantee absolute maximum performance in call cases is to use the High Performance mode.

ExtremeTech on Balanced vs. Performance power management in Windows:

"The next thing to check was power management — and this is where we found our smoking gun. We tested Windows 10 in its “Balanced” power configuration, which is our standard method of testing all hardware. While we sometimes increase to “High Performance” in corner cases or to measure its impact on power consumption, Windows can generally be counted on to handle power settings, and there’s normally no performance penalty for using this mode....but we’ve retested both GPUs in High Performance mode. These new results are significantly different from our previous tests. 4K performance is unchanged, and the two GPUs still tie, but 1080p performance improves by roughly 8% on the GTX 980 Ti and 6% on the Fury X." ~ ExtremeTech

Unless you can test every real world CPU/GPU application known to man and document it to prove that there is no difference, it's already been documented by others that it does have a performance hit. Since it also can vary from system to system, you cannot guarantee that there is penalty for others even if your system shows no penalty in the applications you run. Can you guarantee that 200 million PCs worldwide will not have a performance penalty under Balanced mode? No. For that reason, to ensure uniform testing methodologies aimed to show maximum benchmarking performance, all CPU/GPU benchmarks done by professionals must be done in High Performance mode or there is a risk of outliers as ExtremeTech found out.

It doesn't take a lot of time googling to find out there are gamers who have real world experiences of Balanced mode hurting performance.

"In Balanced Mode, even a simple game like Dota 2 experiences occasional FPS drop (as low as 15 fps), as opposed to when I use High Performance - of which it's a constant 30 fps Vsync-ed." ~ Link

"I ran World of Warcraft and Dirt 3 games. I have Intel i5 2500K CPU. When Windows runs a game in Balance mode, I get either 1605 Mhz or between 3300 and 3600. When I use High Performance, it goes up to 3700." ~ Link

There is absolutely 0 reason to leave a desktop in the Balanced mode when trying to benchmark or the goal is maximum performance. If you really care to save pennies on power usage, you can run Balanced mode and switch to High Performance when you launch a game. Since there are already distributed computing apps I ran over the years that show a degradation in performance for me over 24-48 hours of testing, I don't even need a 3rd party validation as my own system shows the performance penalty exists when running Distributed Computing + having 100+ tabs open across various browsers.

I actually did even more digging to prove this point since it seems in 2015 some people still think Balance = HP and there is no performance penalty.

Professional Source #1

"Our 8-cpu database server has a ton of traffic, but extremely low CPU utilization (just due to the nature of our SQL queries -- lots of them, but really simple queries). It's usually sitting at 10% or less. So I expect it was downclocking even more than the above screenshot. Anyway, when I turned power management to "high performance" I saw my simple SQL query benchmark improve by about 20%, and become very consistent from run to run.

As of Windows Server 2008, the default setting of the OS was switched from High Performance to Balanced. Energy efficiency was becoming a larger factor in the real world, and our ability to balance between the oft-opposing poles of Power and Perf was improving. That being said, there will always be environments where our recommendation is that the power policy should be switched back to High Performance. Anything super latency sensitive will clearly fall into that bucket, such as banking, stock markets, etc.

OEMs have the flexibility to add custom tunings onto their factory settings if they want to put in the additional effort to find a balance that works better for their specific customers. System administrators also have that flexibility. But tuning the power/perf knobs in the OS is a very tricky business, not for the faint of heart." - Scott Hanselman is a former professor, former Chief Architect in finance, now speaker, consultant, father, diabetic, and Microsoft employee.
http://www.hanselman.com/blog/TheNe...ManagementAndPerformanceOnWindowsServers.aspx

Professional Source #2
> This shows gigantic performance hit of Balanced vs. HP. Not surprising because in environments where you performance is tied to low latency, a platform running on Balanced will never be fast enough to engage into full performance on a millisecond (nanosecond) basis.

Even AnandTech already proved this point 5 years ago:

"The Xeon X5670 using the “high performance power plan” is able to boosts its clockspeed regularly, and this results in a 11% higher throughput over the same CPU in “the balanced power plan”. Or you could say that disabling turboboost (by using the power plan “balanced&#8221 results in an 10% throughput disadvantage. The way the queues are handled, this 10% advantage in throughput results in a 31% higher response time. The really interesting thing is the comparison between the L5640 and the Xeon X5670 in “balanced mode”."

The Balanced mode is set to default because > 50% of Windows 7->10 users are likely laptop users and HP would destroy battery life for them, as well as a lot of Windows users do not require maximum performance at the cost of power usage (even in the data center/server environment).

It's been widely documented and proven that HP > Balanced when seeking maximum performance. Since it's not possible to isolate every CPU/GPU intensive application and every PC build/laptop in the world, this is why all benchmarks must be run in HP mode if the goal is to maximize performance not battery life.

 

[3DVagabond]

I get that he works in the industry, which is an issue in itself. It reminds me of the shill hunt (no I am not implying he is a shill), I get various posters here work for various parts of the industry, but you don't see them drawing targets on their backs. More so you don't see them hand feeding some of our residents "odd, my results are this" posts which contradict everything we've seen then disappear into the night.

Did you not see the conspiracy theory literally spring up on the same page of his post "DID NV GET THEM TO TURN OFF AS!!!!"

We're back to an insider saying something, no verification and people are running with it. Didn't we just have a "Maxwell2 can't do Asynch Compute" scare?

Sorry, but I missed that. Do we now have Maxwell2 working with Async Compute?

 

[Mondozei]

IIRC Maxwell GM200 can't run Async Compute natively but only via software emulation, so I don't see why railven accept that as an issue.

Most people, including Maxwell owners like myself, don't really count software emulation as genuine support of a feature.

 

[3DVagabond]

IIRC Maxwell GM200 can't run Async Compute natively but only via software emulation, so I don't see why railven accept that as an issue.

Most people, including Maxwell owners like myself, don't really count software emulation as genuine support of a feature.

That's what I understood, but I thought I might have missed something. Thanks.

So, have they accomplished that yet?

 

[Silverforce11]

It only matters if games use a lot of compute and make it compatible for async mode. Otherwise its of no consequence.

I mean Lionhead has said, the benchmark has 5% ac. At best, there's a 5% uplift for GPU that support it over GPUs that don't.

It doesn't explain the normal UE4 25-50% gap for AMD though. Possibly linked to AMD's worse DX11 driver overhead no longer being an issue in DX12??

 

[.vodka]

http://www.anandtech.com/show/9667/amd-releases-catalyst-159-beta-drivers

Today AMD has released AMD Catalyst 15.9 Beta as their latest driver update, with display driver version 15.201.1151. This driver provides optimizations for the upcoming Star Wars: Battlefront Beta that we will be seeing next week and for the Fable Legends DX12 benchmark that we saw last week.

When the benchmark becomes available, there's the driver that was mentioned in the article.

------------------------

Yeah, I think that's the point here, DX12 frees AMD's cards from the bottlenecks in their DX11 driver, allowing GCN to perform like Maxwell on DX11. Async compute, if available, is the icing on the cake. We still need more data points to draw conclusions, yet by now it's quite clear that DX12 is just what AMD needed.

Not that their DX11 driver wasn't getting better lately, 980's lead over the 290x/390x shrunk a lot since its launch up to now which is commendable, with all these optimizations also benefiting the rest of the lineup... even with all of that, maxwell is clearly superior in DX11 games, especially if gameworks is involved. That lead, at least according to this benchmark, has disappeared.

 

[antihelten]

It only matters if games use a lot of compute and make it compatible for async mode. Otherwise its of no consequence.

I mean Lionhead has said, the benchmark has 5% ac. At best, there's a 5% uplift for GPU that support it over GPUs that don't.

It doesn't explain the normal UE4 25-50% gap for AMD though. Possibly linked to AMD's worse DX11 driver overhead no longer being an issue in DX12??

Slight correction here. Lionhead hasn't actually said anything about compute load directly to my knowledge.

The 5% claim originates from Extremetech if I'm not mistaken, but it's worth noting that Joel Hruska doesn't mention what the 5% refers to (5% of all commands are compute commands or 5% of the frame render time is taken up by compute), which makes quite a big difference for the potential of async compute. It is also worth noting that Joel Hruska specifically mentions that async is turned on. (source)

For a bit more indepth information on this, we have this post on B3D. Here it is mentioned (from someone who has seen GPUView results of both AMD and Nvidia runs), the 5% is frame render time. However interestingly enough, it is also mentioned that whilst async compute is technically turned on, it is only copy commands that are capable of running asynchronously (with either graphics or compute commands), whilst graphics and compute commands are not running asynchronously together (which is kinda the whole point of async compute, at least as far as the large performance gains that have been discussed are concerned).

Of course all of this still doesn't really explain why Zlatan saw such a large uplift in performance with async compute, but I suppose he might have a newer build, where more tasks have been implemented as compute.

 

[littleg]

Slight correction here. Lionhead hasn't actually said anything about compute load directly to my knowledge.

The 5% claim originates from Extremetech if I'm not mistaken, but it's worth noting that Joel Hruska doesn't mention what the 5% refers to (5% of all commands are compute commands or 5% of the frame render time is taken up by compute), which makes quite a big difference for the potential of async compute. It is also worth noting that Joel Hruska specifically mentions that async is turned on. (source)

In contrast to this we have this post on B3D. Here it is mentioned (from someone who has seen GPUView results of both AMD and Nvidia runs), the 5% is frame render time. However interestingly enough, it is also mentioned that whilst async compute is technically turned on, it is only copy commands that are capable of running asynchronously (with either graphics or compute commands), whilst graphics and compute commands are not running asynchronously together (which is kinda the whole point of async compute, at least as far as the large performance gains that have been discussed are concerned).

Of course all of this still doesn't really explain why Zlatan saw such a large uplift in performance with async compute, but I suppose he might have a newer build, where more tasks have been implemented as compute.

He's a PS4 dev, maybe he has a build optimised for that platform.

 

[antihelten]

He's a PS4 dev, maybe he has a build optimised for that platform.

Seeing as Fable Legends isn't coming out for PS4, I would say that that is extremely unlikely to put it mildly.

 

[Silverforce11]

The 5% claim originates from Extremetech if I'm not mistaken, but it's worth noting that Joel Hruska doesn't mention what the 5% refers to (5% of all commands are compute commands or 5% of the frame render time is taken up by compute), which makes quite a big difference for the potential of async compute. It is also worth noting that Joel Hruska specifically mentions that async is turned on. (source)

Joe Hruska said Lionhead briefed the press on the benchmark, that's where he got the 5% figure from.

Either way its such a minuscule figure that its pointless in this bench. But the bench itself, lets not forget, is a fly through of the environment. It's actually devoid of all gameplay and far from being representative of DX12 games. It's more a game engine showcase.

Also, most devs are cross-platform, its quite rare to have PS4 only devs or xbone only. Most studios definitely are all multi-platform.

 

[AtenRa]

Seeing as Fable Legends isn't coming out for PS4, I would say that that is extremely unlikely to put it mildly.

Fable Legends will come to PS4 one year after the XBone and PC.

 

[railven]

Sorry, but I missed that. Do we now have Maxwell2 working with Async Compute?

IIRC Maxwell GM200 can't run Async Compute natively but only via software emulation, so I don't see why railven accept that as an issue.

Most people, including Maxwell owners like myself, don't really count software emulation as genuine support of a feature.

That's what I understood, but I thought I might have missed something. Thanks.

So, have they accomplished that yet?

See how that works.

My "issue" is that people ran with a false notion. Maxwell2 doesn't have GCN AS capabilities, that wasn't even my claim. More so, GCN > Maxwell2 @ AS != Maxwell2 can not run AS. Read Silver's post. Clearly he gets it.

Fable Legends will come to PS4 one year after the XBone and PC.

That's interesting, where did you read that?

 

[Azix]

Zlatans version might have async enabled more extensively. If he wasn't derping. If they have ported the feature over to PC then my guess is nvidia managed to get it turned off completely. They failed to do that with ashes and only got it turned off for themselves (if I understand what happened with oxide and nvidia correctly).

 

[railven]

Zlatans version might have async enabled more extensively. If he wasn't derping. If they have ported the feature over to PC then my guess is nvidia managed to get it turned off completely. They failed to do that with ashes and only got it turned off for themselves (if I understand what happened with oxide and nvidia correctly).

And now we're back to the theory:

Did NV get AS turned off for AMD on the Fable Bench? Without any verification of his results, why are we even barking up that tree?

If someone else started a new account and posted something along the lines of "I got Maxwell2 card, he's right Asynch Compute is turned off for press version, here is 980 Ti results with it on" and it's faster than Fiji.

Oh the can of worms.

 

[AtenRa]

That's interesting, where did you read that?

Heh cant find the article now, ill keep looking.

 

[Stuka87]

And now we're back to the theory:

Did NV get AS turned off for AMD on the Fable Bench? Without any verification of his results, why are we even barking up that tree?

If someone else started a new account and posted something along the lines of "I got Maxwell2 card, he's right Asynch Compute is turned off for press version, here is 980 Ti results with it on" and it's faster than Fiji.

Oh the can of worms.

This has been my question since the beginning of all of this. Does the Fable benchmark even have AS turned on? Or did NV have them disabled it.

 

[AtenRa]

I believe when people say Maxwell 2 doesnt have Async Compute they mean Hardware based AC.

 

[Stuka87]

I believe when people say Maxwell 2 doesnt have Async Compute they mean Hardware based AC.

Is there really a difference? The cards have no physical support for it. Emulating the support has a performance hit, and obviously is not going to be as fast as a hardware implementation.

 

[poofyhairguy]

Is there really a difference? The cards have no physical support for it. Emulating the support has a performance hit, and obviously is not going to be as fast as a hardware implementation.

All it means is there will be a higher CPU performance hit for the Nvidia driver, right? Much like AMD cards were boosted by more powerful CPUs in the Directx 11 age.

 

[thesmokingman]

This has been my question since the beginning of all of this. Does the Fable benchmark even have AS turned on? Or did NV have them disabled it.

If they followed Nv Guide to DX12, then it would be off.

 

[Headfoot]

Is there really a difference? The cards have no physical support for it. Emulating the support has a performance hit, and obviously is not going to be as fast as a hardware implementation.

The answer depends on whether we're talking as intellectual honesty participants or partisan cheerleaders.

You can either say yes or no to this question: "Is acceptably fast performing software emulation the same as hardware for purposes of calling a feature 'Supported'?"

If you say yes this means:
* Maxwell2 supports Async Compute
* AMD could theoretically support DX FL 12_1 if they can write fast enough software paths for those features

If no:
* Maxwell2 does not support Async Compute (Based upon current information)
* AMD can not support DX FL 12_1

The problem is that cheerleaders want to have it both ways for their side, which is intellectually dishonest and illogical. This division also appears between Haswell iGPU based h.265 decode vs gm206 decode

My opinion is that if you can write a fast enough software path (using a mix of CPU, GPGPU, Fixed hardware where appropriate) then you support it because ultimately the consumer gets to experience the feature. Hardware accelerated is always going to be faster and use less energy, and software based is always going to be more flexible and can be updated after its been shipped.

If the question is: "Is there fixed function hardware built into the GPU which supports X feature" then the answers are (based upon current information):
* Maxwell2 does not support fixed-function-hardware accelerated async compute
* AMD does not support fixed-function-hardware accelerated DX12 FL12_1 features

If anyone has verifiable information which proves otherwise I am of course completely willing to change my final assessments on these points, but based on current information that is how it stands.

 

[tential]

I really only care if the performance hit hurts me as a card owner given my cpu. Most people purchased gtx 970s which also means they have i5s or i7s. I really doubt the software emulation hit is that bad. So far I haven't seen any test that makes it look like the nvidia cards are a horrible buy.

 

[AtenRa]

Is there really a difference? The cards have no physical support for it. Emulating the support has a performance hit, and obviously is not going to be as fast as a hardware implementation.

I believe you answered your own question

 

[KingFatty]

The answer depends on whether we're talking as intellectual honesty participants or partisan cheerleaders.

You can either say yes or no to this question: "Is acceptably fast performing software emulation the same as hardware for purposes of calling a feature 'Supported'?"

Would this just shift the argument to whether something is "acceptably fast"?

I think even if a card can achieve a feature even with slow performance, my brain likes to check the imaginary box of that feature being supported.

But it just seems that the term "hardware support" is like a term of art, and means something special. I'd still like to retain both characteristics - 1) whether the feature is supported, and 2) whether it's supported via hardware support.

The "acceptably fast" part would just get absorbed under your typical benchmarking stuff, so it just doesn't seem to be needed as a criteria to evaluate whether a feature is supported in hardware.