[bitsandchips]: Pascal to not have improved Async Compute over Maxwell

[Silverforce11]

problem is most of the commands that dx12 incorporates doesnt seems to help at all any maxwell card not in the way they kept bragging about since 2014... http://www.legitreviews.com/nvidia-highlights-directx-12-strengths-amd_138178
fact is since we saw the revised roadmap that included pascal in the early 2015 (pascal didnt existed on any roadmap before that) we already knew they not gonna have anything ready

NV certainly did huff and puff back then, knowing full well nobody will expose their lies for some time, while they sell a ton of fake DX12 hardware in that time period.

Still lol-worthy how all their official stuff claims Fermi will get DX12 support and also get Vulkan support. Basically all they've proven with these recurring episodes, is that NV simply cannot be trusted with paper specs, nor can they be trust to keep their word on any future feature/support.

In case this point flew over some folk's heads, it's not whether NV has Async Compute on their GPUs or not, it's the fact they lied, many times, so blatantly to sell more hardware at consumer's expense.

 

[airfathaaaaa]

NV certainly did huff and puff back then, knowing full well nobody will expose their lies for some time, while they sell a ton of fake DX12 hardware in that time period.

Still lol-worthy how all their official stuff claims Fermi will get DX12 support and also get Vulkan support. Basically all they've proven with these recurring episodes, is that NV simply cannot be trusted with paper specs, nor can they be trust to keep their word on any future feature/support.

In case this point flew over some folk's heads, it's not whether NV has Async Compute on their GPUs or not, it's the fact they lied, many times, so blatantly to sell more hardware at consumer's expense.

you know what is the proof of all of that? on the 780ti and how it almost competes with 980 on aots...(and i think it surpass it while the async is off) gotta love superscalar designs

 

[railven]

Because according to some questionable posters that like to post things that they think they know about are parading the videocard forums that AC = DX12 and that if you somehow if one does not dont support this in the hardware its the end of the particular company? Sorry for my ignorance but this is the vibe I get from VC&G these days.

Yet I see benchmarks and see financials/marketshare, I keep seeing the opposite happening? From my small knowledge in GPUs, its a little more complicated than that and async compute is one part many of features in the DX12 API just like increased drawcalls, decreased overheads etc.

Regardless of how one IHV does it or not in their GPUs, as long as the performance is there its literally meaningless.

Not ignorance. Every other thread I'm reminded how my GTX 980 Ti is obsolete, and has no future in PC gaming. I've been reading that since late 2015, with a laundry list of upcoming DX12 games.

At this point, I'm just enjoying the ride! My card serves my purposes and whoever inherits it from me gets a monster OC'er with a quiet pump that with all the freebies/rebates I got with it cost me less than a Fury X. Can't complain

 

[Genx87]

I will laugh if Pascal has AC support and does it better than AMD ala Tessellation. Then Im sure the argument will shift back to developers over doing AC to hurt AMD just like the tessellation arguments of past.

Anyways AC is but a part of a game engine anyways. Didnt one of these devs a few weeks ago explain AC makes up less than 10% of performance?

 

[Sweepr]

I will laugh if Pascal has AC support and does it better than AMD ala Tessellation. Then Im sure the argument will shift back to developers over doing AC to hurt AMD just like the tessellation arguments of past.

Anyways AC is but a part of a game engine anyways. Didnt one of these devs a few weeks ago explain AC makes up less than 10% of performance?

True, up to ~10% for some AMD GPUs in AotS, according to ComputerBase. 15% best case scenario - Fiji @ 4K in a scene with many draw calls. Is it really a deal breaker if Pascal doesn't (properly) support it?

 

[railven]

True, up to ~10% for some AMD GPUs in AotS, according to ComputerBase. 15% best case scenario - Fiji @ 4K in a scene with many draw calls. Is it really a deal breaker if Pascal doesn't (properly) support it?

Personally for me, nope. I had radeons during the over-tessellated era, and I had no issues turning down tessellation via drivers (though find it a little misleading if it's true that AMD is now doing "optimizations" ie turning down tessellation through force, not user option - ie Fallout 4 "Ultra BBQ" being "faster" on AMD).

If AC takes off an I'm stuck with an AC-less Pascal for whatever reason, I'd hope NV can man up and create a slider like AMD did for tessellation.

 

[Erenhardt]

It seems some of you are very confused about AC and don't have an idea what it is about.

It is not an additional work for a GPU like tessellation is. It is doing the same tasks but scheduled differently, asynchronously.

If you want a slider to compensate for lack of AC support, every game already have that - it is called graphics options. Turn down AA, shadows, reflections, etc and you will match the performance of equivalent cards that support AC.

Early tests show 10% boost from AC. I think it is only a tip of an iceberg. Time will tell.

 

[krumme]

True, up to ~10% for some AMD GPUs in AotS, according to ComputerBase. 15% best case scenario - Fiji @ 4K in a scene with many draw calls. Is it really a deal breaker if Pascal doesn't (properly) support it?

I think the problem is you dont know what impact it will have for yor favorite games. It can eg be 0% or 30%. The problem is it introduces uncertainty.

Ofcource if the games is released you are interested in and the bm shows insignificant difference from your perspective then its of minor importance.

I have tried gaming with Mantle on bf4 and know what difference eg drawcalls can do in heavy fighting scenes in big 64 pers servers. Then you absolutely dont want to miss such a performance feature. Its a fundamental technology.

We dont know for sure how agressive devs will use asynch compute but 50M consoles is the perfect compettitive place and you want to make the most of your market here. Aparently as Zlatan says its quite difficult to do asynch compute as an integrated part of your arch but i still have a hard time understanding we have to accept weak tack on solutions after so many years. Its not a small thing this technology.

We are on the brink of several new games that will use this technology so when its not there its just sad. End of story. Pascal might be fantastic in many other ways and i expect a lot from the new gen but clearly it would be better if it had that technology.

Downplaying asynch compute as drawcalls was downplayed initially is just pathetic and unworthy. Its fundamental technology that should not be missed.

 

[ShintaiDK]

NVidia/Intel just needs to sponsor games with ROV/CR like AMD sponsors Async. And we have the opposite cases.

Even AMD said that no GPU yet supports all DX12 features. And it will be a long time before we see any that does. It will be post Pascal/GCN1.3. MS is already adding new features to DX12 that will be introduced in the end of the year.

And with the look on future GPU performance, we may end up buying new cards due to features rather than performance, from both discrete IHVs.

 

[krumme]

Great technology is great technology. Take eg NV optimus. It took amd several years to do the same. They were sleeping and they lost control of the mobile market even under fermi era.

Downplaying this tech with comments like it didnt work above eg. 555m or there was driver teething problems initially is just idiotic. It made a huge difference on mobile.

So will a lot of the technology introduced in the consoles together with the new api. Like ac. Great. Its just good for us consumers.

 

[Kenmitch]

Even AMD said that no GPU yet supports all DX12 features.

Yet just means on sale today in the end. Not sure what they or NVidia have planned as far as Polaris/Pascal goes and what features they will or won't support.

 

[Samwell]

It seems some of you are very confused about AC and don't have an idea what it is about.

It is not an additional work for a GPU like tessellation is. It is doing the same tasks but scheduled differently, asynchronously.

If you want a slider to compensate for lack of AC support, every game already have that - it is called graphics options. Turn down AA, shadows, reflections, etc and you will match the performance of equivalent cards that support AC.

Early tests show 10% boost from AC. I think it is only a tip of an iceberg. Time will tell.

In some parts it's adding work and you might get even a penalty as the Hitman devs mentioned.

Not only you have to tune it for every architecture, but it seems you even need to tune it for every chip to really gain performance. Tonga and Tahiti both don't gain anything by using Async in Hitman. So while it's a nice feature, it's "just" 1 of the new DX12 features.

 

[railven]

NVidia/Intel just needs to sponsor games with ROV/CR like AMD sponsors Async. And we have the opposite cases.

Even AMD said that no GPU yet supports all DX12 features. And it will be a long time before we see any that does. It will be post Pascal/GCN1.3. MS is already adding new features to DX12 that will be introduced in the end of the year.

And with the look on future GPU performance, we may end up buying new cards due to features rather than performance, from both discrete IHVs.

I just bought a 290X used for $160 about a month ago. How dare you say it's already obsolete!

Hell, my 980 Ti was obsolete the moment I installed it! haha. I'll shed no tears when I replace both GPUs before the end of the year.

Everyone needs to remember the great sage:
"My new computer's got the clocks, it rocks
But it was obsolete before I opened the box"

 

[krumme]

In some parts it's adding work and you might get even a penalty as the Hitman devs mentioned.

Not only you have to tune it for every architecture, but it seems you even need to tune it for every chip to really gain performance. Tonga and Tahiti both don't gain anything by using Async in Hitman. So while it's a nice feature, it's "just" 1 of the new DX12 features.

It is damn difficult but i think the same goes for getting the most of a new thin api on several different pc platforms. Its all over the implementation and we have seen eg memory managent and leaking problems in mantle. Going to a thin api have proved to be very difficult.

It will take several years before the new engines is ready and tuned. It will also be interesting to see what engines will get used most.

 

[mysticjbyrd]

Which seems very possible if Nvidia has to overcompensate in GPU power to make up for lacking async.

Isn't Nvidia using an inferior process? Their dies already have to be 9% bigger.

 

[garagisti]

I'll take a smooth 60 fps over a jittery 70 fps anytime.

This was before 290, and anything new. It was hilarious how the web sites were missing the FCAT equipment when AMD was doing better. No, nothing suspicious about it at all. it doesn't say anything about their credibility at all that they wouldn't publish results when AMD was doing better.

 

[ShintaiDK]

Isn't Nvidia using an inferior process? Their dies already have to be 9% bigger.

TSMC still got the better node electrically.
http://www.eetimes.com/document.asp?doc_id=1329279

I also bet you TSMC got better yields.

 

[Bacon1]

In some parts it's adding work and you might get even a penalty as the Hitman devs mentioned.

Not only you have to tune it for every architecture, but it seems you even need to tune it for every chip to really gain performance. Tonga and Tahiti both don't gain anything by using Async in Hitman. So while it's a nice feature, it's "just" 1 of the new DX12 features.

Of course there are times when it will be slower, because it takes some time to split up work and time to bring that work back together. If its faster to just do it parallel than to take the time to split it up and have multiple people work it, then it will be slower in the end.

For most cases though, it is faster to split up that work, especially for any "hard" jobs. Splitting up easy work isn't worth it.



I'll try to make a super simple example.

Think of it like taking lunch orders for 12 people.

No-Async:

You can have a single person ask each 12 people what they want and then call the delivery place (monitor) with the results. This is simple and straight forward, just ask each person one at a time and then call once you have all 12 orders.

Async:

You have the main order taker tell 4 people to each ask 3 other people what they want to order. When done they give the results of what sandwich each person wanted to the main order taker.

Faster or Slower (Worth it?):

Now if those people are all right next to eachother (easy work) it would have been faster to just ask them in a line.

If those people were in different buildings (hard work) it would be faster to have 1 person go to each building to take the orders, then come back instead of the single person having to walk between each building.

Hopefully that can simplify it for the people that aren't programmers here

 

[garagisti]

I will laugh if Pascal has AC support and does it better than AMD ala Tessellation. Then Im sure the argument will shift back to developers over doing AC to hurt AMD just like the tessellation arguments of past.

Anyways AC is but a part of a game engine anyways. Didnt one of these devs a few weeks ago explain AC makes up less than 10% of performance?

AMD had tesselation support since HD2900 i think. They didn't purposefully help develop a game, which gimped performance by tesselating invisible water under the streets, concrete blocks which didn't need as much tesselation and so on which was discovered in Crysis 2.

Liking a brand and its offerings is fine, but don't reward anti-consumer behaviour, or consumers in the long term will have a problem. Of course, if you're a share holder, and in it to make money, i could understand that approach of supporting dodgy behaviour.

 

[3DVagabond]

GM200 beats Hawaii by ~10% and Hawaii is a total failure. Situation reversed? No big deal.

 

[krumme]

Isn't Nvidia using an inferior process? Their dies already have to be 9% bigger.

As i understand it they trade area for efficiency. And its not small amount of efficiency they get for those few mm2. It means they are far superior at the mobile market. A typical good nv decision that shows they know the market needs.

But we have to take into account that at launch eg 970 was faster than 290 and therefore not big. Today its oposite. But still a very balanced card imo and the market agrees.

But a pascal without a new core that is genuine ac capable is imo a blunder. I for one will not take a chance buying it if true.

 

[Kenmitch]

But a pascal without a new core that is genuine ac capable is imo a blunder. I for one will not take a chance buying it if true

Allegedly they've been working with Microsoft for years on DX12 so why would they leave the feature out of the design? Too me it would be somewhat counterproductive if they chose to skip the feature. Who doesn't like free performance....If used properly.

Guess we'll find out within the next month or so.

 

[krumme]

Allegedly they've been working with Microsoft for years on DX12 so why would they leave the feature out of the design? Too me it would be somewhat counterproductive if they chose to skip the feature. Who doesn't like free performance....If used properly.

Guess we'll find out within the next month or so.

The reason is the 50M consoles, origin (dice) and Mantle that changed the prior game fundamentally. Look at dx12 documentation its a total Mantle ripoff. And even at the coding level there is similarities between consoles and pc. Its unheard until now. It was a gamechanger.

I expected ac for Pascal but aparently its not that easy going by Zlatans remarks? Thats how i understand it and why there is no real ac.

 

[itsmydamnation]

TSMC still got the better node electrically.
http://www.eetimes.com/document.asp?doc_id=1329279

I also bet you TSMC got better yields.

Your link doesn't state what you just said in any way shape or form, other then TMSC has had a better ramp on 16nm then they did on 28nm........

 

[Kenmitch]

The reason is the 50M consoles, origin (dice) and Mantle that changed the prior game fundamentally. Look at dx12 documentation its a total Mantle ripoff. And even at the coding level there is similarities between consoles and pc. Its unheard until now. It was a gamechanger.

I expected ac for Pascal but aparently its not that easy going by Zlatans remarks? Thats how i understand it and why there is no real ac.

In that case with the current info/rumors it doesn't really look good at all for them.

It was more of a dose of hope reply in the end.

NVidia was playing tic,tac,toe while Intel played tic, toc....AMD was playing chess it looks like. Nobody knows how the game will end at this time.