Codemasters builds DX12 support into its EGO Engine 4.0

[Mauller]

I should also add that DX12 on its own is limited to 4 threads when it comes to drawcalls.

Where did you read this? As far as i am away the low level API's are only limited to the number of threads you code your renderer for.

And if this is based on the API Overhead test, the test runs into the limit of how many Drawcalls a GPU can process before being limited by the CPU.

 

[Mauller]

There's no real reason not to either, There's nothing you can't change to make it more private.

I loved Win 7 and hated Win 8.1 when I moved to it but having now moved to Win 10 it's like being on 7 again. It's a stable OS that with a few settings tweaks does nothing I don't want it too.

All you need is a third party firewall, you then block the reporting and feedback framework when it asks for internet access.

 

[humbug]

Where did you read this? As far as i am away the low level API's are only limited to the number of threads you code your renderer for.

And if this is based on the API Overhead test, the test runs into the limit of how many Drawcalls a GPU can process before being limited by the CPU.

Have a look at the thread. In DX12 Matt 5960 scores 18m. In mantle it's 33m.

DX12 stopped Scaling at a certain number of threads. Probably more than 4 but it's still very limited compared with mantle

I need to run it on my i5 to see if the 8 thread i7 scales any better

 

[Mauller]

Have a look at the thread. In DX12 Matt 5960 scores 18m. In mantle it's 33m.

DX12 stopped Scaling at a certain number of threads. Probably more than 4 but it's still very limited compared with mantle

Could be a few reasons why the score was like that. one of them being the MGPU code not being used in DX12 version etc. Also i think if he did that a while ago then the DX12 drivers were not as good at the time.

 

[humbug]

Could be a few reasons why the score was like that. one of them being the MGPU code not being used in DX12 version etc. Also i think if he did that a while ago then the DX12 drivers were not as good at the time.

What does the mgpu code have to do with it?
Matt was also one of the last to run it in DX12 and Mantle. Recently.

They scale the same until about 4 to 8 threads, I need to run my i5 on it to see which

 

[Mauller]

What does the mgpu code have to do with it?
Matt was also one of the last to run it in DX12 and Mantle. Recently.

They scale the same until about 4 to 8 threads, I need to run my i5 on it to see which

Because the API test scales with the number of cards in your system. But if it was recent then i am not sure what is up with that. Could be something with how the mantle code is run. I remember the oxide guys mentioning something about Small batch optimisations.

But i doubt it is thread limited to four threads, there must be something more to it in the way the code is written.

 

[humbug]

Because the API test scales with the number of cards in your system. But if it was recent then i am not sure what is up with that. Could be something with how the mantle code is run. I remember the oxide guys mentioning something about Small batch optimisations.

But i doubt it is thread limited to four threads, there must be something more to it in the way the code is written.

It doesn't. It's completely CPU bound, that's the whole point. Don't believe me or matt ask Kaap to do an independent assessment. He will agree with me. He was the one who first spotted it.

 

[Mauller]

It doesn't. It's completely CPU bound, that's the whole point. Don't believe me or matt ask Kaap to do an independent assessment. He will agree with me. He was the one who first spotted it.

You need to compare DX12 and Mantle in a single card setup. there must be something wrong with the multigpu in DX12 for the API test. that or it doesnt have support for it yet.

If you notice in the DX12 numbers, the multi gpu people are getting around the same numbers as what single card users are getting if you give some leeway to overclocking etc. it is more than likely the implementation of MGPU rather than a problem with the API and cores it can use. in regards to the API overhead test.

It is only when you compare multicard users on mantle to single cards that you see a dramatic uplift in Batches since it needs to submit more to feed each gpu. And if you look at Savagesams results you see that DX12 even beats his mantle scores on his fury x.

 

[humbug]

There is no difference between mgpu and other users just as there is no difference between higher and lower performance GPU's. There is also no difference if you run it at 4K or 720P.

Why. Because it's measuring drawcalls per second not 3D rendering. It's Purley a CPU test, the GPU is only relevant in what API it can run, it's performance and how many does not come into it.
It's an API OVERHEAD TEST

 

[Gregster]

There is no difference between mgpu and other users just as there is no difference between higher and lower performance GPU's. There is also no difference if you run it at 4K or 720P.

Why. Because it's measuring drawcalls per second not 3D rendering. It's Purley a CPU test, the GPU is only relevant in what API it can run, it's performance and how many does not come into it.

Are you talking about 3DMark's API test here?

 

[humbug]

Yes

 

[Mauller]

There is no difference between mgpu and other users just as there is no difference between higher and lower performance GPU's. There is also no difference if you run it at 4K or 720P.

Why. Because it's measuring drawcalls per second not 3D rendering. It's Purley a CPU test, the GPU is only relevant in what API it can run, it's performance and how many does not come into it.
It's an API OVERHEAD TEST

It is to do with the number of GPU's as well, you need more draw calls when you drive multiple GPU's. And if the MGPU code is not working or implemented then the DX12 mode is only testing a single card.

Look at Soldierboys scores, he has dual Furyx, he gets 18mil calls in DX12 and 27mil in Mantle.

then compare it to savage sam with a single fury X and stock cpu, he gets 15mil in dx12 and 14mil in mantle.

And then you look at MAtt, his DX12 score is 18mil for 4x fury x, around the same as soldierboys 2xfuryx then trailed by savagesam at 15mil on his stock processor. Although Matt's Mantle score is 33 mil compared to soldier boys 27mil.

The API test is just as CPU dependant as it is Dependant on the number of GPU's in the system if MGPU is implemented.

 

[Gregster]

Yes

So then it is also using the GPU (no idea if SLI works and I can't test as I no longer run SLI but when I ran it, my GPU got hotter and if you look here, there is a difference in scores between the 290X and 285



And Mantle scored a little over but nowhere near what you was saying.



It would make no sense for DX12 to be limited to 4C and should/would be a simple task of enabling more like Mantle did.

 

[Mauller]

You posted the same chart twice there greg

plus that doesn't show draw calls scaling with number of gpus like i am talking about. Although it gives a good representation of the scaling with cores etc.

 

[humbug]

He gets more in mantle because it has better thread scalling than DX12, the same reason matt gets the same Scaling in mantle vs DX12, The same reason Kaap again gets the same scalling.
It's like your trying to make the evidence fit your theory while the peg you hold is square. With that it's like trying to explain internal combustion to a caveman when really it's all so simple.
Drawcalls have nothing to do with the GPU, IT IS A CPU ONLY THING.

 

[Gregster]

You posted the same chart twice there greg

plus that doesn't show draw calls scaling with number of gpus like i am talking about. Although it gives a good representation of the scaling with cores etc.

Yer ooops. I can't find SLI results but not looked too hard.

 

[Mauller]

He gets more in mantle because it has better thread scalling than DX12, the same reason matt gets the same Scaling in mantle vs DX12, The same reason Kaap again gets the same scalling.
It's like your trying to make the evidence fit your theory while the peg you hold is square. With that it's like trying to explain internal combustion to a caveman when really it's all so simple.
Drawcalls have nothing to do with the GPU, IT IS A CPU ONLY THING.

It is just as much to do with the GPU as it is the CPU. plus greg just showed good evidence that scaling in DX12 is just as good as mantle with a single card.

A gpu can only process so many draw calls to begin with and the cpu will need to generate more draw calls when addressing multiple cards. hence why Matt has such a high number of draw calls. i am also sure that in an old thread he did the api test with different numbers of cards that showed multi card scaling in the api test.

 

[Gregster]

I should also add that DX12 on its own is limited to 4 threads when it comes to drawcalls. That's 4x better than DX11 but still needs clever innovation to help when it comes to serious work.
Mantle BTW is 12 at least.

He gets more in mantle because it has better thread scalling than DX12, the same reason matt gets the same Scaling in mantle vs DX12, The same reason Kaap again gets the same scalling.
It's like your trying to make the evidence fit your theory while the peg you hold is square. With that it's like trying to explain internal combustion to a caveman when really it's all so simple.
Drawcalls have nothing to do with the GPU, IT IS A CPU ONLY THING.

Where did Matt get 33m in this test? And no it isn't limited to 4 threads (and 4 threads would be 2C4T?) And it still uses the GPU regradless of how you put it in capitals to drive your point across.

 

[humbug]

It's in the thread greg you only need look

It is just as much to do with the GPU as it is the CPU. plus greg just showed good evidence that scaling in DX12 is just as good as mantle with a single card.

A gpu can only process so many draw calls to begin with and the cpu will need to generate more draw calls when addressing multiple cards. hence why Matt has such a high number of draw calls. i am also sure that in an old thread he did the api test with different numbers of cards that showed multi card scaling in the api test.

What Gregs chart shows is that a 290x doesn't scale as well in drawcalls threading as a Fury. Wich is very interesting in its self but has absolutely nothing to do with the performance of it.
Probably a difference in sync threading. It is interesting.

 

[Mauller]

here is matts score from 3dmark itself

http://www.3dmark.com/3dm/8108517