Rumor: Nvidia’s Pascal Architecture Is In Trouble With Asynchronous Compute

[patrickjp93]

4 hours ago, Notional said:

So what happens to Maxwell cards when DX12 games starts using GPU compute for effects and AI?

 

 

 

Yeah Maxwell crashes and burns.

It's a rumor, and I've seen engineering samples. Bitsandchips is wrong 8 ways from Sunday.

[Notional]

26 minutes ago, That Norwegian Guy said:

I called it.

A few of us did. I mean we have heard nothing about Pascal what so ever from NVidia. If they truly are just a few months from launch, we would have heard something! I think NVidia are a bit stressed about this async compute thing. Maybe even performance of TSMC's 16nm FF (not that GloFO's 14nm FF is better in that regard). But who knows.

1 minute ago, patrickjp93 said:

It's a rumor, and I've seen engineering samples. Bitsandchips is wrong 8 ways from Sunday.

And how does that tell you how well the card does in async compute in games? Or hardware scheduling in general? If so please do enlighten us.

[patrickjp93]

16 minutes ago, Notional said:

A few of us did. I mean we have heard nothing about Pascal what so ever from NVidia. If they truly are just a few months from launch, we would have heard something! I think NVidia are a bit stressed about this async compute thing. Maybe even performance of TSMC's 16nm FF (not that GloFO's 14nm FF is better in that regard). But who knows.

And how does that tell you how well the card does in async compute in games? Or hardware scheduling in general? If so please do enlighten us.

Teslas, so no gaming.

[Serin]

Here's a question: Do any 3d design render engines require async? Or any that are planning to take advantage of in some way in the future?
Talking cycles, mental ray etc etc?

 

[D-Dos Dan]

Is there another law similar to Godwin's that states the longer an online chat goes about 2 companies, the higher the chance of a flame war?

[Prysin]

5 hours ago, patrickjp93 said:

Teslas, so no gaming.

Meaning whatever you saw was a nvidia card running pure compute. Be it mixed half precision or double precision.

Either way you have no proof whether they re-added out of order schedulers or not.

[xAcid9]

4 hours ago, Serin said:

Here's a question: Do any 3d design render engines require async? Or any that are planning to take advantage of in some way in the future?
Talking cycles, mental ray etc etc?

 

Probably not because those task use pure compute?

[Citadelen]

30 minutes ago, xAcid9 said:

Probably not because those task use pure compute?

From what In know whenever the GPU is used there will always be bubbles of idleness. I would assume Async can be used within anything that issues tasks to the GPU.

[D-Dos Dan]

2 hours ago, Quibiss said:

You mean the  "defending multimillion dollar companies that do not give a single shit about me but since I define myself and my whole personality with things I own I have to choose a side so I have something to fight for on the internet"-law?

Yeah something along the lines of that

[patrickjp93]

6 hours ago, Prysin said:

Meaning whatever you saw was a nvidia card running pure compute. Be it mixed half precision or double precision.

Either way you have no proof whether they re-added out of order schedulers or not.

Nope. You can use Teslas to render graphics. You just need supplementary video outputs like VGA or you need to send the results to an external machine if you're on a server board.

 

Nvidia could already do asynchronous compute calls while in compute mode and do asynchronous shader calls while in rendering mode. What has been added is the ability to freely do both without heavy context switching.

[Prysin]

1 minute ago, patrickjp93 said:

Nope. You can use Teslas to render graphics. You just need supplementary video outputs like VGA or you need to send the results to an external machine if you're on a server board.

 

Nvidia could already do asynchronous compute calls while in compute mode and do asynchronous shader calls while in rendering mode. What has been added is the ability to freely do both without heavy context switching.

In layman terms, for those who don't know ;

In Fermi and kepler they had hardware schedulers. So existing Teslas could do "pure/true" async. Maxwell doesn't feature hardware schedulers so it must use software to schedule compute workloads.

 

However Patrick do tell me. If you know.

Did you run mixed half/normal or DP?

 

Because nvidia has been touting 10X compute over maxwell. However most of the info is pointing towards 10X mixed precision rather than 32 or 64

[patrickjp93]

1 hour ago, Prysin said:

In layman terms, for those who don't know ;

In Fermi and kepler they had hardware schedulers. So existing Teslas could do "pure/true" async. Maxwell doesn't feature hardware schedulers so it must use software to schedule compute workloads.

 

However Patrick do tell me. If you know.

Did you run mixed half/normal or DP?

 

Because nvidia has been touting 10X compute over maxwell. However most of the info is pointing towards 10X mixed precision rather than 32 or 64

All of the above over the course of a few tests.

 

There may be a peak point where mixed precision has nearly 10X speedup over Maxwell, but the best I found was 8x.

[RagnarokDel]

14 hours ago, AresKrieger said:

I'm going to assume this information is false since nvidia isn't that stupid, plus it's very rare that tech rumors from amd or nvidia are true due to all the stupid fanboyism. Either way dx12 is still not relevant to me, and until a game I actually want to play come out as dx12 exclusive it will remain that way.

Pascal was in the design process before Maxwell was released.

[patrickjp93]

1 hour ago, RagnarokDel said:

Pascal was in the design process before Maxwell was released.

And how long have consoles and Mantle been out, giving Nvidia a perfect roadmap of where AMD's features would be headed? There's been plenty of time to modify the designs since the launch of the 750/TI.

[SurvivorNVL]

6 minutes ago, patrickjp93 said:

And how long have consoles and Mantle been out, giving Nvidia a perfect roadmap of where AMD's features would be headed? There's been plenty of time to modify the designs since the launch of the 750/TI.

Not to mention that nVidia has purposely removed certain feature-sets and optimized the chips they put out at a given time.  Pascal should be marvelous from what you are saying, and with that - it may solidfy nVidia's place or maybe AMD will have a home run.  I remember awhile ago you said the FirePro's of Arctic Island were incredible at DP, but Pascal was better at SP while the new Intel chips which are Knight's Landing or Xeon Phi were about in the middle and more cost-effective watt wise.  Right?  I might be wrong on that.  Going off memory.

[Energycore]

Let us just not hope AMD takes over the market because then we'll see the same effects of monopoly on the other side. But I still don't think a single generation will be enough to have that happen. There's just too much market to catch up with.

[Prysin]

3 hours ago, Energycore said:

Let us just not hope AMD takes over the market because then we'll see the same effects of monopoly on the other side. But I still don't think a single generation will be enough to have that happen. There's just too much market to catch up with.

No we won't. Amd have too much debt to pay down to try any shenanigans that might jeopardise their market share.

 

For AMD to pull any of the shit nvidia have been doing the past few years, they would need to hold a 60%+ market share in the DGPU space and Zen must give them at least 20% market share in the cpu space. 

Most importantly they must keep this market share until 2020 in order to be remotely safe enough to start doing any sort of consumer unfriendly shenanigans 

[Energycore]

May thy words be true.