[DX12] Fable Legends Beta Benchmarks

[Headfoot]

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.

Agreed, I think hardware support is one check box, and "support" generally is another check box. And I agree the debate would move to "Acceptably fast" which is the actual conclusion that the engineers are considering when they choose which design strategy to follow when they originally create the hardware. (Acceptably fast meaning whether that feature then bottlenecks total performance or if another bottleneck kicks in before it/at the same time as it). I completely agree with you.

I'm basically saying we should debate based on reality, not on fantastical definitions of "support" which change and don't apply equally depending on whether its their favorite company or not.

"acceptably fast" analysis already occurs in most benchmark reviews indirectly when they evaluate whether a card is for example ROP bound at a particular resolution, or whether a certain VRAM capacity is acceptable for a certain resolution. Whether a software emulated feature is fast enough could easily slot in with these other factors competent reviewers are already evaluating. This is basically what is being benchmarked in the AOTS benchmark right now in all of these "Does Maxwell2 support Async" threads

 

[railven]

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.

What makes you think it was turned off?

EDIT:

Heh cant find the article now, ill keep looking.

If I had to guess, you misread an article. This is a MSFT first party studio and a first party published game. It will not see the light of day outside of Xbox/PC unless Lionhead studios gets bought out by another company.

Which brings me back to the question I asked: what kind of NDA does a PS4 dev have on a MSFT first party game?

 

[96Firebird]

Heh cant find the article now, ill keep looking.

I think you're confusing Fable Legends with Rise of the Tomb Raider.

 

[AtenRa]

I think you're confusing Fable Legends with Rise of the Tomb Raider.

Yea figures why I couldnt find the article, thanks

 

[3DVagabond]

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.



That's interesting, where did you read that?

Maybe there was a communication breakdown. Let me ask simply. Is Async compute now active either in drivers or hardware with any nVidia DX12 cards? Fermi, Kepler, Maxwell 1 or 2? If it is, which is it, drivers or hardware?

Thanks.

 

[antihelten]

So a bit more info has come to light. Someone over at B3D shared a snapshot of the AMD press deck for Fable, which contained some of AMD's performance numbers for async on/off.

Async apparently applies to GI and Simulation and Culling. With async off, GI takes 1.53ms and Simulation and Culling takes 1.63ms. With async on, GI takes 0.766ms and Simulation and Culling takes 0.51ms.

PCPer's reported numbers for these stages correspond to AMD's async on numbers (anandtech's numbers are much higher, but they tested at 4K and not 1080P like AMD and PCPer). As such it seems relatively safe to assume that async is running, at least for AMD.

We can then calculate the total amount of time taken up by Compute, which is roughly 20% with async turned off and 10% with async turned (technically the various compute commands should take roughly the same time whether async is on or off, but with async on, some of that time will overlap with graphics commands, thus lowering the effective time taken). Furthermore the numbers reported by AMD would entail a performance boost of roughly 15%.

A 15% boost is relatively close to what Zlatan reported (he got 20%), although the average FPS numbers are still much lower than his (most sites report 75-80 FPS versus his 105.8 FPS). I suppose this might be a result of Zlatan having a more optimised build.

 

[N7SpectreElite]

Link to the snapshot please

 

[Mondozei]

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.

Silver just said that when you have sub-5% of AC, it doesn't matter much. He never said software emulation means Maxwell supports AC.

And you keep refusing to admit that it's an issue, including in the section I'm quoting now.

You're saying that just because GCN is better at AC, that doens't mean GM200 can't run it.

And we're back to square one: software emulation means you can't run it because it isn't done natively. So you've achieved nothing in your takedown, other than to display your own ignorance in what the issue was and remains(and you still don't understand it, judging by your post) and even managed to misattribute silver's quote while you were at it. Well done.

Everyone deserves a chance to understand a basic concept. When they show themselves unable to do that repeatedly, they no longer deserve other people's time and attention. So I'm done with you.

 

[tential]

Silver just said that when you have sub-5% of AC, it doesn't matter much. He never said software emulation means Maxwell supports AC.

And you keep refusing to admit that it's an issue, including in the section I'm quoting now.

You're saying that just because GCN is better at AC, that doens't mean GM200 can't run it.

And we're back to square one: software emulation means you can't run it because it isn't done natively. So you've achieved nothing in your takedown, other than to display your own ignorance in what the issue was and remains(and you still don't understand it, judging by your post) and even managed to misattribute silver's quote while you were at it. Well done.

Everyone deserves a chance to understand a basic concept. When they show themselves unable to do that repeatedly, they no longer deserve other people's time and attention. So I'm done with you.

You're caught up on software/hardware.
I couldn't care less personally. If I'm a maxwell user and I still get a boost from ac, even software, then I don't care....

This whole thing is being talked about by people to say gcn>maxwell.

If I'm a 970/980ti owner (the only 2 cards that matter if you purchase another maxwell card lol at you but another matter entirely), I doubt you care this much. It's being overblown like no other.

And before someone calls me a maxwell/nvidia lover, I have the hd7950, and just bought an r9 290.

People are making any argument to say how finally gcn is better, or whatever. It's boring.
If you're a gcn owner great you're happy your card will benefit from Async when it's actually used. If you're a maxwell owner, you got amazing performance from maxwell since maxwell release. Games that we unplayable on gcn work on maxwell and every game is focused on optimizing for a maxwell owner first.

Pros and cons of being in each boat....

 

[maddie]

You're caught up on software/hardware.
I couldn't care less personally. If I'm a maxwell user and I still get a boost from ac, even software, then I don't care....

This whole thing is being talked about by people to say gcn>maxwell.

If I'm a 970/980ti owner (the only 2 cards that matter if you purchase another maxwell card lol at you but another matter entirely), I doubt you care this much. It's being overblown like no other.

And before someone calls me a maxwell/nvidia lover, I have the hd7950, and just bought an r9 290.

People are making any argument to say how finally gcn is better, or whatever. It's boring.
If you're a gcn owner great you're happy your card will benefit from Async when it's actually used. If you're a maxwell owner, you got amazing performance from maxwell since maxwell release. Games that we unplayable on gcn work on maxwell and every game is focused on optimizing for a maxwell owner first.

Pros and cons of being in each boat....

For the present owners of maxwell & GCN cards this is a fine viewpoint, but what if I'm in the market for a new card?

 

[railven]

Silver just said that when you have sub-5% of AC, it doesn't matter much. He never said software emulation means Maxwell supports AC.

And you keep refusing to admit that it's an issue, including in the section I'm quoting now.

You're saying that just because GCN is better at AC, that doens't mean GM200 can't run it.

And we're back to square one: software emulation means you can't run it because it isn't done natively. So you've achieved nothing in your takedown, other than to display your own ignorance in what the issue was and remains(and you still don't understand it, judging by your post) and even managed to misattribute silver's quote while you were at it. Well done.

Everyone deserves a chance to understand a basic concept. When they show themselves unable to do that repeatedly, they no longer deserve other people's time and attention. So I'm done with you.

Where have we seen that GM200 can not process Asynch Compute? How many examples do we have? Whether it is done "native" or via "emulation" doesn't change if something can not process something.

There was an even better post after mine that went further into this.

Can GM200 run Asynch Code? Yes, the two examples we've seen show it running Asynch Code. If the recent info on Fable is true, it has Asynch Compute code. Now, can GM200 run Asynch Code as fast as GCN, most likely not even close.

Can Xbox One play Xbox 360 games? Sure, it does it through emulation. "NOOOO, IT CANT IT DOESNT HAVE THE HARDWARE. ITS NOT NATIVE." [Get out of here] with that nonsense.

Maybe there was a communication breakdown. Let me ask simply. Is Async compute now active either in drivers or hardware with any nVidia DX12 cards? Fermi, Kepler, Maxwell 1 or 2? If it is, which is it, drivers or hardware?

Thanks.

You're asking the wrong person. I'm not sure what Nvidia is doing with their hardware. So far, I've yet to see any proof that GM200 can't do AS. Even AOTS shows that it can do it, just so ridiculously poor you're better off turning it off.

Profanity isn't allowed in the technical forums, acronym or otherwise.
-- stahlhart

 

[tential]

For the present owners of maxwell & GCN cards this is a fine viewpoint, but what if I'm in the market for a new card?

Then you have no idea since no games are fully released, no drivers are finalized or anything.
Read the post earlier with a link about citizen and dx12. We don't even have engines that fully utilize dx12 yet.

Neither vendors current gpus are useful in dx12 in the long run. Just like every time before, we won't see the best gpus for the api until after its widely used. Think 2017 gpus.

Also just like last gen it's all about implementation. Some implementation of Async will favor how nvidia handled it in maxwell (gameworks games) others will favor gcn, and others still will be useful. Despite Async. The 980ti is still the top dog.

If you're in the marker for a new gpu then like everyone who intelligent has said, you need a stopgap gpu to hold you over until you know what the best choice is. That's why I have an r9 290. I'll upgrade to whatever gpu handles new games best afterwards.

Look at the performance charts at the end of the day it's not like it drastically changes anything. An r9 390x and gtx 980 are still the same performance class after dx12.

A good stopgap card, 290, 970 is what I recommend. If you got the cash a 980ti. Then sell the card for whatever card you need once you know the best dx12 performers. You'll lose at most 50 dollars.... Its the cost of a game and to be honest this is pc gaming. To those who think it's cheap your being utterly ignorant.

 

[tential]

Also let's be even more real, we bought gcn cards that couldn't be fully utilized in dx11 and I don't we the massive outrage from anyone about this. Yet a huge deal is made about maxwell now. This is being made into a big deal because there are fanboys on both sides that look for one thing their favorite vendor does then blows it out of proportion. At the end of the day no matter what choice you make now you're making a compromise. Neither vendor has straight up 100% winner at the moment.

 

[railven]

Also let's be even more real, we bought gcn cards that couldn't be fully utilized in dx11 and I don't we the massive outrage from anyone about this. Yet a huge deal is made about maxwell now. This is being made into a big deal because there are fanboys on both sides that look for one thing their favorite vendor does then blows it out of proportion. At the end of the day no matter what choice you make now you're making a compromise. Neither vendor has straight up 100% winner at the moment.

At this point I don't even care if Maxwell2 handles AS poorly (which all accounts so far show that it does, and if AS is gonna be a huge factor in future titles - I'd advise against Maxwell2 if you wish to keep your card for a while), but people are running to the hills that "AS doesn't work with Maxwell2" If they said "it isn't run natively", sure so far that's been shown to be true.

But these people are hinging their arguments on something not being done native (or on hardware) as the sole definition of "support" or "processing" is just ridiculous.

 

[railven]

And you keep refusing to admit that it's an issue, including in the section I'm quoting now.

Forgot to answer this. It doesn't affect me. It isn't an issue to me, right now. None of the games I play are negatively affected by Maxwell's current inferior version of Asynch Compute. I also have zero interest in Fable and AOTS (I've yet to play a single Fable game, and AOTS is not a genre I follow).

I only represent myself on these forums, but because it isn't an issue to me currently doesn't mean it shouldn't be criticized. If NV has inferior AS support, well that's on them to fix. If it cost them sales, well that's on them to explain.

When new cards come out if AMD has the better product, that's where my money goes. If NV pulls a rabbit out of it's hat, that's where my money goes.

 

[EarthwormJim]

Where have we seen that GM200 can not process Asynch Compute? How many examples do we have? Whether it is done "native" or via "emulation" doesn't change if something can not process something.

There was an even better post after mine that went further into this.

Can GM200 run Asynch Code? Yes, the two examples we've seen show it running Asynch Code. If the recent info on Fable is true, it has Asynch Compute code. Now, can GM200 run Asynch Code as fast as GCN, most likely not even close.

Can Xbox One play Xbox 360 games? Sure, it does it through emulation. "NOOOO, IT CANT IT DOESNT HAVE THE HARDWARE. ITS NOT NATIVE." [Get out of here] with that nonsense.

GM200 cannot run Asynch code though, you're talking about a physical chip. Nothing on it physically can do mixed mode compute and graphics in parallel. Nvidia's drivers can accept Asynch commands though and pass them onto the host processor. The hardware cannot handle it, but the system can.

Your Xbox analogy doesn't hold up when describing GM200, but it works fine if you talk about the hardware and software together. GM200 cannot do Asynch, but GM200 plus Nvidia's software can

 

[railven]

GM200 cannot run Asynch code though, you're talking about a physical chip. Nothing on it physically can do mixed mode compute and graphics in parallel. Nvidia's drivers can accept Asynch commands though and pass them onto the host processor. The hardware cannot handle it, but the system can.

Your Xbox analogy doesn't hold up when describing GM200, but it works fine if you talk about the hardware and software together. GM200 cannot do Asynch, but GM200 plus Nvidia's software can

Holy crap. Where back at this. Please, show me where Asynch Compute is defined as being hardware driven.

I'm pretty sure this horse was beaten to death in the AOTS thread. Even MSFT doesn't define it as "hardware." You can have hardware accelerated AS, such as GCN, or software driven AS, such as Maxwell2.

End of the day both are AS, however, one implementation has been shown to be incredibly more efficient.

Again, Maxwell2 (or I guess since I said the GM200, but I was referring to the family) doesn't have AS hardware, but it can certainly handle AS code. [HINT: It cheats by offloading some of the scheduling to the CPU.]

We done with this or is someone gonna bark "it isn't being done native" again?

 

[96Firebird]

GM200 cannot run Asynch code though, you're talking about a physical chip. Nothing on it physically can do mixed mode compute and graphics in parallel. Nvidia's drivers can accept Asynch commands though and pass them onto the host processor. The hardware cannot handle it, but the system can.

A GPU cannot process anything without drivers.

 

[EarthwormJim]

Holy crap. Where back at this. Please, show me where Asynch Compute is defined as being hardware driven.

I'm pretty sure this horse was beaten to death in the AOTS thread. Even MSFT doesn't define it as "hardware." You can have hardware accelerated AS, such as GCN, or software driven AS, such as Maxwell2.

End of the day both are AS, however, one implementation has been shown to be incredibly more efficient.

Again, Maxwell2 (or I guess since I said the GM200, but I was referring to the family) doesn't have AS hardware, but it can certainly handle AS code. [HINT: It cheats by offloading some of the scheduling to the CPU.]

We done with this or is someone gonna bark "it isn't being done native" again?

The whole point is to make use of idling hardware on the GPU. Doing it in software with a lengthy switch is self defeating.

Let's get down to why we even use GPUs; so we can run code directly on hardware. If all people care about is features, let's go back to software rendering, certainly can tick off a bunch more features that way.

Does Nvidia's Asynch support let marketing tick off a check box? Yes.

Does it serve any real use? Probably not, and it was actually detrimental in Ashes according to the developer.

 

[railven]

The whole point is to make use of idling hardware on the GPU. Doing it in software with a lengthy switch is self defeating.

And the recent discoveries are showing that was mostly AMD's problem, not Nvidia. Why Nvidia doesn't seem to gain as much using AMD's version of AS. And so far, we don't have anything from NV to represent an optimized version of their AS. Heck, even Fable isn't showing much use of AS (at least from what I've read) and we don't even know conclusive if it was on or off from the benchmarks that were run.

Let's get down to why we even use GPUs; so we can run code directly on hardware. If all people care about is features, let's go back to software rendering, certainly can tick off a bunch more features that way.

/facepalm

This is the equivalent of putting the cart before the horse. Kudos, AMD had ACE/AC since debut of Tahiti 2011. OMG you're finally now getting to see it used in a 2015 benchmark for a game slated for 2016 release! Of course we want hardware native stuff, but right now - why? When more games use it if NV still is being all obsolete with their inferior implementation, they will pay for it. Right now - only people really caring are forum posters trying to score points.

Does Nvidia's Asynch support it let marketing tick off a check box? Yes.

Does it serve any real use? Probably not, and it was actually detrimental in Ashes according to the developer.

Marketing. It's all about marketing. But I tell you what, I didn't buy my Maxwell2 card based on DX12 performance. By the time DX12 performance actually matters to me Maxwell2 will be a distant memory.

 

[EarthwormJim]

And the recent discoveries are showing that was mostly AMD's problem, not Nvidia. Why Nvidia doesn't seem to gain as much using AMD's version of AS. And so far, we don't have anything from NV to represent an optimized version of their AS. Heck, even Fable isn't showing much use of AS (at least from what I've read) and we don't even know conclusive if it was on or off from the benchmarks that were run.



/facepalm

This is the equivalent of putting the cart before the horse. Kudos, AMD had ACE/AC since debut of Tahiti 2011. OMG you're finally now getting to see it used in a benchmark of a 2015 game slated for 2016 release! Of course we want hardware native stuff, but right now - why? When more games use it if NV still is being all obsolete with their inferior implementation, they will pay for it. Right now - only people really caring are forum posters trying to score points.



Marketing. It's all about marketing. But I tell you what, I didn't buy my Maxwell2 card based on DX12 performance. By the time DX12 performance actually matters to me Maxwell2 will be a distant memory.

This whole debate is fueled by Nvidia marketing. It's eerily similar to the 4GB of ram on the GTX 970. Sure it's "there" but is all of it useful?

It wouldn't be a big deal if they hadn't advertised Async as a feature for Maxwell cards.

 

[railven]

This whole debate is fueled by Nvidia marketing. It's eerily similar to the 4GB of ram on the GTX 970. Sure it's "there" but is all of it useful?

It wouldn't be a big deal if they hadn't advertised Async as a feature for Maxwell cards.

First bold: oddly, if you ask me this is all stemming from AMD's marketing side. And I'm not even mad that AMD is finally playing dirty. They are using their hardware advantage to murky up the water. They sponsored a game that puts their version of AS (ie hardware driven) in an extreme positive light. Kudos AMD, bring the gauntlet!

Second bold: WTF!? Let's talk about concepts. I guess Bulldozer is not really two cores per module, because...well I won't even get into it, since we'd just get all kinds of messy.

The Red: Aye dios mio. It's probably as useful as the second 4GBs of VRAM on 390/390x. IE, you'll probably run out of GPU processing before you can use it all up. "But what about mGPU!!!" But 4GBs is good enough for Fiji for 4K!!!!!

 

[EarthwormJim]

First bold: oddly, if you ask me this is all stemming from AMD's marketing side. And I'm not even mad that AMD is finally playing dirty. They are using their hardware advantage to murky up the water. They sponsored a game that puts their version of AS (ie hardware driven) in an extreme positive light. Kudos AMD, bring the gauntlet!

Second bold: WTF!? Let's talk about concepts. I guess Bulldozer is not really two cores per module, because...well I won't even get into it, since we'd just get all kinds of messy.

The Red: Aye dios mio. It's probably as useful as the second 4GBs of VRAM on 390/390x. IE, you'll probably run out of GPU processing before you can use it all up. "But what about mGPU!!!" But 4GBs is good enough for Fiji for 4K!!!!!

Bulldozer is two integer cores, 1 floating point core. AMDs slides and material make this clear. Great comparison on how to properly document your hardware.

 

[3DVagabond]

You're asking the wrong person. I'm not sure what Nvidia is doing with their hardware. So far, I've yet to see any proof that GM200 can't do AS. Even AOTS shows that it can do it, just so ridiculously poor you're better off turning it off.

I missed it running in AOTS. There's a lot of back and forth about Maxwell and Async compute but not much of substance. All I read is the "don't worry, be happy", it doesn't matter responses. So, where is AOTS running ASync compute on Maxwell?

 

[maddie]

I always find it so strange that some here act as if their situation is LAW. Reading some responses one might think that ALL of the present GCN and Maxwell2 users will dump their cards and get the next gen as soon as they go on sale.