Doom Alpha GPU test

[Ieldra]

I won't begrudge the results - they are what they are and the beta numbers are far more realistic.

I think a lot of the argumentative side of people you are seeing is, prior to seeing the beta, you could sum up a lot of (not all) the responses to "NO! Cuz Nvidia!"

Just a bunch of preprogrammed drivel and people swapping benches that had nothing to do with DOOM to win a pissing contest.

Usually I find when people go in hard and double down on the brand loyalty and emotion, they get 4x the hate back. It's not a way to win an argument (and since when should people try to "win" arguments anyway? The only way you ever truly win is if the other person changes their mind, and that RARELY ever happens).

But I digress....

This is nothing new, the only recent development, in my eyes, is this that it's fashionable to be vehemently anti-nvidia. Is it better/worse than being vehemently anti-amd ? No, it's the same damn thing.
This async shader affair is particularly ugly, because it's a pretty technical detail, and it's become a buzzword in an entrenched PR battle and there's no good way of taking the level of technical detail down a notch to enable a less complex discussion. I've seen the 'async = HTT ' argument many times, most people don't know HTT is a marketing term itself It's never easy

 

[Zion Halcyon]

This is nothing new, the only recent development, in my eyes, is this that it's fashionable to be vehemently anti-nvidia. Is it better/worse than being vehemently anti-amd ? No, it's the same damn thing.
This async shader affair is particularly ugly, because it's a pretty technical detail, and it's become a buzzword in an entrenched PR battle and there's no good way of taking the level of technical detail down a notch to enable a less complex discussion. I've seen the 'async = HTT ' argument many times, most people don't know HTT is a marketing term itself It's never easy

Not better nor worse. But I do see the NVidia side always starting more shit than the AMD side. This of course could also be due to more people using NVidia cards than AMD, so there is a larger pool of "true believers" to draw from.

As far as async goes, there just is not enough out there yet for anyone to just say "async" as a trump card over NVidia, but early returns do look promising. Nvidia needs to be "bloodied" a bit for the consumer's sake, if we ever are to get another round of card wars and better prices.

But ultimately, none of this is going to matter much unless Polaris, which is coming in June, can either:

1. Straight up beat the 980TI head to head, or else
2. Be priced so competitively that it offers a performance in-between the 390x and the 980 TI, but is priced cheaper than the 390x, with the other current offerings other than Fury, Fury X, and Nano similarly dropping radically in price.

In other words, if the card is not a straight up beast, it needs to be so stunning a price-performance ratio that it will sticker-shock Nvidia to have to drop cards prices to compete.

Based on what Raja has already said, it might likely be 2 that they go with, but the question will remain if they are bold enough and can afford to have the new cards set at a cheaper price point than the 390x is currently, which is what I think it would take to really put Nvidia on its heels. Not unless Polaris ends up being a true monster. We have until End of June to find out.

 

[JustReason]

First these new results mirror why I posted this thread to begin with, even and reasonable across the board.

Second async may be the new buzz word, but its importance is no less important. What in recent history, say last 10 years shows this kind of performance boost. And keep in mind this is early usage, likely to get more and extensive use as we go along. So I wouldn't underestimate its benefits simply because for now it hasn't tip the balance heavily.

 

[Ieldra]

Not better nor worse. But I do see the NVidia side always starting more shit than the AMD side. This of course could also be due to more people using NVidia cards than AMD, so there is a larger pool of "true believers" to draw from.

As far as async goes, there just is not enough out there yet for anyone to just say "async" as a trump card over NVidia, but early returns do look promising. Nvidia needs to be "bloodied" a bit for the consumer's sake, if we ever are to get another round of card wars and better prices.

But ultimately, none of this is going to matter much unless Polaris, which is coming in June, can either:

1. Straight up beat the 980TI head to head, or else
2. Be priced so competitively that it offers a performance in-between the 390x and the 980 TI, but is priced cheaper than the 390x, with the other current offerings other than Fury, Fury X, and Nano similarly dropping radically in price.

In other words, if the card is not a straight up beast, it needs to be so stunning a price-performance ratio that it will sticker-shock Nvidia to have to drop cards prices to compete.

Based on what Raja has already said, it might likely be 2 that they go with, but the question will remain if they are bold enough and can afford to have the new cards set at a cheaper price point than the 390x is currently, which is what I think it would take to really put Nvidia on its heels. Not unless Polaris ends up being a true monster. We have until End of June to find out.

I don't expect them to release something Fury level, Fiji is a very shortlived architecture in that case. I hate how AMD name each of their GPUs differently it's confusing as fuck. GCN 1.2 is the architecture, Fiji is the gpu right ?

I think what most people really are underestimating is the increased cost on the new nodes and yields; the fact that TSMC can produce 600mm dies this early in the game is nothing short of a small engineering miracle, both AMD and NVIDIA will only gain by slowing it down a bit; they don't need to dethrone Fiji and GM200, they just need to offer 390/970 390x/980 performance at lower prices and lower TDP

AMD needs mobile marketshare, they seriously have none of that. None.

 

[Ieldra]

First these new results mirror why I posted this thread to begin with, even and reasonable across the board.

Second async may be the new buzz word, but its importance is no less important. What in recent history, say last 10 years shows this kind of performance boost. And keep in mind this is early usage, likely to get more and extensive use as we go along. So I wouldn't underestimate its benefits simply because for now it hasn't tip the balance heavily.

Man I have no interest in downplaying the importance of async, it is great! Even if the performance boost never reaches 10% (and we've seen it does) it's still great in terms of flexibility and making life for developers a little easier.

I posted Doom numbers that directly challenge the claim (backed by Doom alpha bench) that AMD has a distinct advantage in low level APIs, you repeatedly point out that there's a cap at 60 fps, when it's evident I'm looking at 4k

You do this twice.

Then you tell me I have issues with comprehension, claim Maxwell should be benchmarked at reference clocks and that AMD wins at every tier except the top one, accusing me of focusing on the 980ti when I made no mention of it. How do you expect this to be taken seriously ?

If I'm not mistaken you're the guy who really took aim at me for mentioning overclocking as a means to close the gap caused by async, do you understand my point now ?

Hypothetical scenario.

Fury X w/o Async 100fps
Titan X w/o async 100 fps
Fury X W/ Async 110 fps
Titan X w/ 10% oc 110fps

Of course you can always say AMD overclocks too! Sure. By how much in relative terms? It will NEVER match the %oc on maxwell, which is also the basis for me saying that GTX 970 is a lot more competitive vs 390 than people give it credit for nowadays

Fury X 15% oc + async 126.5 fps
TX @ 25% oc NO async 125 fps

This is particularly valid in AotS because it's very compute limited. My benchmark scores was 25% higher than stock with a 25% overclock. PERFECT clock scaling

Addendum: IN THOSE CASES, none so far, in which async provides more than 10% benefit for GCN, i can almost guarantee you that without async enabled they will be outperformed by the relevant nvidia gpu in the same tier

 

[JustReason]

You are reading way to much into my posts. I prefer to keep it simple. Discussing a benchmark needs to be just with the facts at hand. When you add other aspects you are making it far too complicated and difficult than it needs to be.

 

[Ieldra]

You are reading way to much into my posts. I prefer to keep it simple. Discussing a benchmark needs to be just with the facts at hand. When you add other aspects you are making it far too complicated and difficult than it needs to be.

The facts at hand point to async providing no tangible performance gains over competing nvidia gpus

If it's simple, concise explanations you want you should find a new hobby

 

[Quix]

I like how it doesn't support SLI or Crossfire. That's awesome...

A 60fps cap doesn't affect me (my monitor is locked at 60hz anyway) but I can see why it would annoy people who spent a bunch of cash on one of those high refresh rate monitors.

 

[Ieldra]

As far as Doom is concerned, I think any game with multi-player DLC leading to fragmentation is, fittingly, doomed to fail

 

[Deleted member 233574]

Guru3D is a very fucking good website for benchmarks, Hilbert puts a lot effort in testing a big number of games on a lot of hardware. OC3D tests two cards...

More data

That's a low system requirement for 1080p gaming. In Doom3 era, one need the best GeForce to run it with the highest image quality.
In this new Doom, all I need is a mid range board?

 

[Zion Halcyon]

The facts at hand point to async providing no tangible performance gains over competing nvidia gpus

If it's simple, concise explanations you want you should find a new hobby

Er, that's not the facts at hand dude. This here is my issue with stuff like this, because what you are saying isn't reality.


The facts at hand is that in DX 11 games, async makes no difference because it is a DX12 feature, therefore Nvidia holds the same lead they always have.
The facts at hand is that in async-enabled DX 12 games, the AMD cards hold about a 10FPS increase over their direct NVidia competitor.

So before you mouth off to someone wanting to keep it simple, get your facts right.

Or was it worded that way on purpose, with you using the word "tangible" as a fallback?

10 fps might be tangible to some people, you know.

 

[Ieldra]

Er, that's not the facts at hand dude. This here is my issue with stuff like this, because what you are saying isn't reality.


The facts at hand is that in DX 11 games, async makes no difference because it is a DX12 feature, therefore Nvidia holds the same lead they always have.
The facts at hand is that in async-enabled DX 12 games, the AMD cards hold about a 10FPS increase over their direct NVidia competitor.

So before you mouth off to someone wanting to keep it simple, get your facts right.

Or was it worded that way on purpose, with you using the word "tangible" as a fallback?

10 fps might be tangible to some people, you know.

Hey, I'm commuting atm hard to respond, read my post again please in the meantime, I'll respond ASAP, and please stop talking to me about fps gains because I don't give a shit; talk %%%%%%

 

[Zion Halcyon]

Hey, I'm commuting atm hard to respond, read my post again please in the meantime, I'll respond ASAP, and please stop talking to me about fps gains because I don't give a shit; talk %%%%%%

If you don't give a shit, that is your prerogative, but until they start benchmarking by % and not FPS, don't give me that "talk %" bullshit. You got called on this and you know it. Nothing further to say - not interested in how you try to spin this.

 

[Ieldra]

If you don't give a shit, that is your prerogative, but until they start benchmarking by % and not FPS, don't give me that "talk %" bullshit. You got called on this and you know it. Nothing further to say - not interested in how you try to spin this.

you not understanding the benefit of talking % gains vs absolute fps gains is indicative of a fundamentally flawed thought process.


Here is my response to your previous post


First things first, there is only one game in which we can isolate the gain from asynchronous shaders from the gain from Dx12. Ashes.

In ashes of the singularity, a compute bound game, there is absolutely no advantage for Gcn w/Async vs maxwell at fp32 parity. This is not debatable, I have proven this.

Conversely, at fp32 parity, nvidia hw has a performance advantage vs amd hw with no async.

Doom is a vulkan title, and the developers have stated they have gotten significant gains out of async shaders for amd hw.

Despite these big gains they perform equal to their nvidia counterparts, which do not make use of asynchronous shaders.

Ergo async shaders provide no tangible benefits over NVIDIA hardware running without them.

The 10% boost is relative to amd hw running without async, not relative to nvidia hw.

You want to include Hitman in the conversation? That's another discussion, a discussion I'm pretty sure we've had previously

Edit: what do you even mean 'unless they start benchmarking by %', are you aware of the fact that % is a ratio?

How can I know if 10fps is tangible or not if I don't know what you're comparing it to.

10 fps gain at 20fps is 50%

10 fps gain at 100 fps is 10%

 

[defaultluser]

What in recent history, say last 10 years shows this kind of performance boost?

Just a few examples:

FXAA - replaces 2x to 4x MSAA at 1/10the performance cost. Also gets around all the caveats of MSAA, like doesn't work on shader effects. Once developers figured out the sharpen pre-process, the result was beautiful. And this is simple shader software, so it works everywhere! And other variants have improved upon it to the point where nobody really needs the MSAA hit anymore.

Delta Color Compression: allows the 380X to perform nearly the same as the 280X with 1/3 less memory bandwidth. Allows the GTX 960 to perform the same as the GTX 770 with 1/2 the memory bandwidth. That results in a much more massive performance increase than Async brings to the table, for both companies!

When the industry hands you lemons instead of cheaper higher-bandwidth memory, you squeeze the lemons until they take up less space. I mean, we've been on the GDDR5 train for EIGHT YEARS NOW, and HBM is more expensive for now!

This is why Async doesn't really matter yet. The problem is, you have to get your hands particularly dirty in the hardware to get any noticeable improvement out of it (hey, just like a console!). So we'll have to see how much effort designers put into this stuff before they run out of budget/patience.

Just about every other popular improvement is either in hardware (DCC), or is so easy to cut and paste (FXAA) that you'd be stupid not to use it.

 

[razor1]

now I want a Tom Collins lol, guess I know what I'm going to drink after work now.

Well in consoles its much more effective as the life time of a console is 4 to 5 years. But for graphics cards, gens change in 1.5 years.

 

[filo]

After reading this idtech crap 60fps lock.. i had to register and enlighten you guyz. Esreality post.
Direct link 120, 144fps

 

[filo]

i was a lurker for far to long, I always compared review data from hardforum, anandtech and tomsharware. My safty checks.

 

[noko]

Numbers are numbers, it is the gaming experience that counts in the end. Once again Async Compute is being argued - I am sure it is the tech of tomorrow for all vendors with AMD supporting now.

I agree, AMD needs to really target market share (turns into mind share) starting with the mobile platform. Polaris 10 and HBM on laptops? Would be extremely cool. Would make playing Doom rather fun at least

 

[Ieldra]

Numbers are numbers, it is the gaming experience that counts in the end. Once again Async Compute is being argued - I am sure it is the tech of tomorrow for all vendors with AMD supporting now.

I agree, AMD needs to really target market share (turns into mind share) starting with the mobile platform. Polaris 10 and HBM on laptops? Would be extremely cool. Would make playing Doom rather fun at least

We're arguing about async shaders, not async compute, I'll be damned if anybody knows or cares what the difference is though.


Async may well be the 'tech of tomorrow' but the fact of the matter is, it's not a defining feature in terms of performance as we've seen here, and also in the AotS thread.

The reason it's being argued, at least why I'm arguing, is that it's consistently treated like the second coming of graphics christ, and it isn't.

 

[MangoSeed]

Async is a very nice feature but it's only one of many factors that will determine final performance.

I suspect async is more useful for GCN than for nVidia's recent architectures. It seems GCN's relatively poor geometry throughput means the shader array is idle for extended periods of time during geometry heavy passes. Scheduling useful work in parallel to the shader core would be very beneficial in those cases.

 

[noko]

We're arguing about async shaders, not async compute, I'll be damned if anybody knows or cares what the difference is though.


Async may well be the 'tech of tomorrow' but the fact of the matter is, it's not a defining feature in terms of performance as we've seen here, and also in the AotS thread.

The reason it's being argued, at least why I'm arguing, is that it's consistently treated like the second coming of graphics christ, and it isn't.

I think you are arguing with yourself mostly. Yes it is a feature that tomorrow cards (AMD just did it first) that will need to be supported for better efficiency. If Nvidia can't do it now, mislead folks etc. is their problem but I am sure they will be able to do a compute and graphic operation on the the gpu at the same time and out of order in the future. They will have to if they want to compete. Doing everything serial wise will fall short.

 

[Ieldra]

Async is a very nice feature but it's only one of many factors that will determine final performance.

I suspect async is more useful for GCN than for nVidia's recent architectures. It seems GCN's relatively poor geometry throughput means the shader array is idle for extended periods of time during geometry heavy passes. Scheduling useful work in parallel to the shader core would be very beneficial in those cases.

You probably know this unless you've purposefully been ignoring my, several, posts but at compute throughout parity the 980Ti actually pulls ahead by 10%, Async gets amd on par.

I'm trying to replicate these results on a gtx 970 but the friend running the benchmarks is cpu limited ' (

 

[Ieldra]

I think you are arguing with yourself mostly. Yes it is a feature that tomorrow cards (AMD just did it first) that will need to be supported for better efficiency. If Nvidia can't do it now, mislead folks etc. is their problem but I am sure they will be able to do a compute and graphic operation on the the gpu at the same time and out of order in the future. They will have to if they want to compete. Doing everything serial wise will fall short.

Yeah nvidia and their antiquated serial execution need to go the way of the dodo,they'll want to start parallelizing execution before it's too late

 

[noko]

You probably know this unless you've purposefully been ignoring my, several, posts but at compute throughout parity the 980Ti actually pulls ahead by 10%, Async gets amd on par.

I'm trying to replicate these results on a gtx 970 but the friend running the benchmarks is cpu limited ' (

It is well know that Maxwell does Context switches way faster then AMD GCN hardware - AMD design is more for multi-threading. Anyways this thread is for Doom which I am looking forward to, Async stuff should go in a Async thread, you can start one and if I don't show up just continue on.

 

[noko]

Yeah nvidia and their antiquated serial execution need to go the way of the dodo,they'll want to start parallelizing execution before it's too late

Well Nvidia I think spotted it right with DX11 type gaming and their hardware does great. I am sure they will do the same with DX12 in the end.

 

[Ieldra]

Well Nvidia I think spotted it right with DX11 type gaming and their hardware does great. I am sure they will do the same with DX12 in the end.

Oops. Forgot the /s

 

[razor1]

Its not question of serial execution, context switches etc. Its the in ability to interleave instructions between two different queues. PS this is not a requirement of DX12.... So, what developers have been doing with async shaders is outside of DX12 specs. Its a nice to have but its not a necessity to be DX12 compatible. Now what will be interesting is what the profiles will be with Vulkan and Doom 3 and if concurrent execution is happening, if it is. Well then, something is up in DX 12 (outside of the spec async shaders) and maxwell 2

So AMD has async (when developers force nV to use the same path as AMD), nV has Gameworks. same shit different companies.

Now for the people that have been arguing about Pascal, and async, I think its quite abundently clear its going to be good for VR and thus why fine grained preemption has been introduced in Pascal, this will promote changes to the scheduler automatically to help async shaders as well. Will they be useful for Pascal is another matter as it seems Pascal will have better ALU utilization than Maxwell 2..... But this will also allow Pascal to do the same things as we are seeing now with Fiji, and the programmable queues. This definitely has more inclination to VR needs to reduce latency. Still don't know how effective it will be for other types of games, mainly because by doing so you will still introduce latency in the form of burning GPU cycles and stalls from using fences when normally they aren't needed. Syncing for VR is quite different than traditional games.

And this goes for Polaris and Vega too, they seem to be dropping crazy amounts of ALU's and improving through put as Vega's rumors are pointing to the same amount of ALU's as Fiji. Which means Both Vega and top end Polaris are going to very close with theoretical SP throughput along with their counterparts and their comparative opponents in their market segments.