[Updated] Oxide responds to AotS Conspiracies, Maxwell Has No Native Support For DX12 Asynchronous Compute

[HKZeroFive]

EDIT: Just found out that it was already in OP. Silly me

[Briggsy]

Again, these are all assumptions. You don't know if Pascal will beat out Arctic Islands.

So far it looks like AMD is winning DX12.

 

Considering Maxwell and Pascal were announced in at least 2010 and 2013, Its a safe bet that Nvidia will not Support Native DX12 Async Compute (out-of-sync to cpu draw calls) for at least 2-3 years unless Nvidia pump billions of reserve cash into a complete redesign that implements the proper hardware. They can do it, but will they? 

[zMeul]

snip

designing Pascal without full DX12 hardware support can be quite catastrophic for nVidia; but .. I don't believe they'll do such thing

it was shown time and time again that the vast majority of the people buying computer hardware, that they don't have a fucking clue what they're buying - so, for them to do that, would be suicidal; that market share will turn on it's head so fast that it wouldn't know what hit them

so no, I don't believe nVidia to be stupid

[patrickjp93]

They want people to buy their cards and really just force them to - AMD put everything into GCN 1.2 - Tessellation, A-sync compute, and more whereas Nvidia likely anticipated the outcome and planned so that their cards get a shorter lifespam and the enthusiasts are forced to upgrade.

Yup. Nvidia gets you the best for the here and now and charges a premium for it. AMD goes cheap and has its tech age well without taking the performance crown which makes them look both cheap and uncompetitive. AMD needs to act and think like a business, not a charity.

[patrickjp93]

I generally agree with your post, but these two points stick out:

If people get the notion that AMD can handle DX12 better, AMD should gain market share both in the near future and long term, as people are looking forward to upcoming titles, utilizing DX12. So I very much doubt AMD will suffer from this. On the contrary.

NVidia hardware is not faster. Not in flops and not in DX12. NVidia's performance gains are due to multithreaded low overhead DX11 drivers. Those are rendered irrelevant in DX12, which is one of the biggest reasons we are seeing this shift.

Nvidia wins in flops, just not DP in consumer cards. Until the Fury X, Nvidia has had the flops lead since Kepler if not Fermi. And games don't use DP.

[Zah]

http://www.overclock.net/t/1569897/various-ashes-of-the-singularity-dx12-benchmarks/2130#post_24379702

 

New post by the dev.

[Prysin]

Nvidia wins in flops, just not DP in consumer cards. Until the Fury X, Nvidia has had the flops lead since Kepler if not Fermi. And games don't use DP.

 

Nvidia hasnt bean in lead in FLOPS since pre fermi. Ever since GCN came out, Nvidia has LOST in flops, in every fucking card to card comparison. Stop spreading bullshit.

 

R9 380 

Pix rate: 31 GPixels/s

Tex Rate: 109 GTexel/s

FLOPS: 3,476 GFLOPS

 

GTX 960

Pix rate: 36.1 GPixels/s

Tex Rate: 72.1 GTexel/s

FLOPS: 2,308 GFLOPS

-------------------------------------------------

 

R9 390

Pix rate: 64 GPixels/s

Tex Rate: 160 GTexel/s

FLOPS: 5,120 GFLOPS

 

 

970

Pix rate: 58.8 GPixels/s

Tex Rate: 109 GTexel/s

FLOPS: 3,494 GFLOPS

-------------------------------------------------

 

R9 390X

Pix rate: 67.2 GPixels/s

Tex Rate: 185 GTexel/s

FLOPS: 5,914 GFLOPS

 

 

GTX 980

Pix rate: 71.2 GPixels/s

Tex Rate: 144 GTexel/s

FLOPS: 4,616 GFLOPS

-------------------------------------------

 

R9 Fury

Pix rate: 64 GPixels/s

Tex Rate: 224 GTexel/s

FLOPS: 7,168 GFLOPS

-------------------------------------------

 

R9 Fury X

Pix rate: 67.2 GPixels/s

Tex Rate: 269 GTexel/s

FLOPS: 8,602 GFLOPS

 

 

980Ti

Pix rate: 96 GPixels/s

Tex Rate: 176 GTexel/s

FLOPS: 5,632 GFLOPS

 

Titan X

Pix rate: 96 GPixels/s

Tex Rate: 192 GTexel/s

FLOPS: 6,144 GFLOPS

------------------------------------------

 

 

Now for the past

GTX 680

Pix rate: 32.2 GPixels/s

Tex Rate: 129 GTexel/s

FLOPS: 3,090 GFLOPS

 

HD 7970GHz Edt.

Pix rate: 32 GPixels/s

Tex Rate: 128 GTexel/s

FLOPS: 4,096 GFLOPS

 

 

Ever since AMD dropped VLIW in favor of GCN, NIVIDIA HAS NEVER. ON ANY ACCOUNT. EVER. CAUGHT UP TO THEM IN FLOPS

[Humbug]

http://www.overclock.net/t/1569897/various-ashes-of-the-singularity-dx12-benchmarks/2130#post_24379702

New post by the dev.

So NVIDIA is working on a driver update for async compute. That explains why the Ark Survival DX12 patch was delayed at the last moment? They said it was a driver thing.

[Prysin]

So NVIDIA is working on a driver update for async compute. That explains why the Ark Survival DX12 patch was delayed at the last moment? They said it was a driver thing.

well drivers are only going to do so much.... if the hardware scheduler is not up to the task. No matter how great of a driver, you will lose to the hardware solution with the better scheduler...

 

Unfortunatly, some things, you just cannot get over by simple driver magic.

[don_svetlio]

So NVIDIA is working on a driver update for async compute. That explains why the Ark Survival DX12 patch was delayed at the last moment? They said it was a driver thing.

Not really - drivers cannot make A-sync compute magically appear. Think of it as hyperthreading.

"It allows jobs to be cycled into the GPU during dormant phases. In can vaguely be thought of as the GPU equivalent of hyper threading. Like hyper threading, it really depends on the workload and GPU architecture for as to how important this is. In this case, it is used for performance. I can’t divulge too many details, but GCN can cycle in work from an ACE incredibly efficiently. Maxwell’s schedular has no analog just as a non hyper-threaded CPU has no analog feature to a hyper threaded one."

If a driver could fix it - then I could simply flash my CPU's software to that of a 4790 and have an i7 at half the price

[Icecube26]

So NVIDIA is working on a driver update for async compute. That explains why the Ark Survival DX12 patch was delayed at the last moment? They said it was a driver thing.

https://docs.unrealengine.com/latest/INT/Programming/Rendering/ShaderDevelopment/AsyncCompute/index.html .... UE4 will use assync compute... so looks like .... they wont be able to fix... cuz its a architecture issue... AoTS dev that works with DX 12 said that and i believe him...he even tried to run on a 980ti... and the 980ti got rekt pretty much...

[Mahigan]

Not really - drivers cannot make A-sync compute magically appear. Think of it as hyperthreading.

"It allows jobs to be cycled into the GPU during dormant phases. In can vaguely be thought of as the GPU equivalent of hyper threading. Like hyper threading, it really depends on the workload and GPU architecture for as to how important this is. In this case, it is used for performance. I can’t divulge too many details, but GCN can cycle in work from an ACE incredibly efficiently. Maxwell’s schedular has no analog just as a non hyper-threaded CPU has no analog feature to a hyper threaded one."

If a driver could fix it - then I could simply flash my CPU's software to that of a 4790 and have an i7 at half the price

nVIDIA do support Asynchronous Compute on a hardware level but a large part of their scheduler is implemented in software. Therefore what feeds their Asynchronous Warp Schedulers is a series of Software solutions.

 

nVIDIAs solution is also limited to 32 Compute Queues (or 1 Graphic + 31 Compute). Therefore under any sort of high Asynchronous Compute workloads, they run the risk of hitting a bottleneck. This is further compounded by the Latency introduced under Preemption due to their slow context switching.

 

Face it... nVIDIA was not as prepared as AMD for DX12. That's to be expected given that AMD has had a lot of experience with Mantle as well as the console API's.

 

We may need to wait until Pascal in order to see the level of engineering talent over at nVIDIA as it pertains to DX12.

[Briggsy]

https://docs.unrealengine.com/latest/INT/Programming/Rendering/ShaderDevelopment/AsyncCompute/index.html .... UE4 will use assync compute... so looks like .... they wont be able to fix... cuz its a architecture issue... AoTS dev that works with DX 12 said that and i believe him...he even tried to run on a 980ti... and the 980ti got rekt pretty much...

 

They will definitely have to use the ACEs in the Xbone to maximize its performance for ARK.

[patrickjp93]

Nvidia hasnt bean in lead in FLOPS since pre fermi. Ever since GCN came out, Nvidia has LOST in flops, in every fucking card to card comparison. Stop spreading bullshit.

R9 380

Pix rate: 31 GPixels/s

Tex Rate: 109 GTexel/s

FLOPS: 3,476 GFLOPS

GTX 960

Pix rate: 36.1 GPixels/s

Tex Rate: 72.1 GTexel/s

FLOPS: 2,308 GFLOPS

-------------------------------------------------

R9 390

Pix rate: 64 GPixels/s

Tex Rate: 160 GTexel/s

FLOPS: 5,120 GFLOPS

970

Pix rate: 58.8 GPixels/s

Tex Rate: 109 GTexel/s

FLOPS: 3,494 GFLOPS

-------------------------------------------------

R9 390X

Pix rate: 67.2 GPixels/s

Tex Rate: 185 GTexel/s

FLOPS: 5,914 GFLOPS

GTX 980

Pix rate: 71.2 GPixels/s

Tex Rate: 144 GTexel/s

FLOPS: 4,616 GFLOPS

-------------------------------------------

R9 Fury

Pix rate: 64 GPixels/s

Tex Rate: 224 GTexel/s

FLOPS: 7,168 GFLOPS

-------------------------------------------

R9 Fury X

Pix rate: 67.2 GPixels/s

Tex Rate: 269 GTexel/s

FLOPS: 8,602 GFLOPS

980Ti

Pix rate: 96 GPixels/s

Tex Rate: 176 GTexel/s

FLOPS: 5,632 GFLOPS

Titan X

Pix rate: 96 GPixels/s

Tex Rate: 192 GTexel/s

FLOPS: 6,144 GFLOPS

------------------------------------------

Now for the past

GTX 680

Pix rate: 32.2 GPixels/s

Tex Rate: 129 GTexel/s

FLOPS: 3,090 GFLOPS

HD 7970GHz Edt.

Pix rate: 32 GPixels/s

Tex Rate: 128 GTexel/s

FLOPS: 4,096 GFLOPS

Ever since AMD dropped VLIW in favor of GCN, NIVIDIA HAS NEVER. ON ANY ACCOUNT. EVER. CAUGHT UP TO THEM IN FLOPS

780TI says hello. Just not in DP.

[Prysin]

780TI says hello. Just not in DP.

780Ti was matched against the 290X (at the time). So was the Titan Black...

 

So lets see how these add up, shall we?

 

R9 290X

Pix rate: 64 GPixels/s

Tex Rate: 176 GTexel/s

FLOPS: 5,632 GFLOPS

 

GTX 780Ti

Pix rate: 52.5 GPixels/s

Tex Rate: 210 GTexel/s

FLOPS: 5,040 GFLOPS

 

GTX Titan Black

Pix rate: 53.3 GPixels/s

Tex Rate: 213 GTexel/s

FLOPS: 5,125 GFLOPS

 

 

You were saying?

Time for an apology to all the people you tried to bullshit i guess....

[DuckDodgers]

780Ti was matched against the 290X (at the time). So was the Titan Black...

 

So lets see how these add up, shall we?

 

R9 290X

Pix rate: 64 GPixels/s

Tex Rate: 176 GTexel/s

FLOPS: 5,632 GFLOPS

 

GTX 780Ti

Pix rate: 37.1 GPixels/s

Tex Rate: 210 GTexel/s

FLOPS: 5,345 GFLOPS

 

GTX Titan Black

Pix rate: 39.2 GPixels/s

Tex Rate: 213 GTexel/s

FLOPS: 5,644 GFLOPS

 

 

You were saying?

Time for an apology to all the people you tried to bullshit i guess....

 

Fixed!

[Prysin]

Fixed!

From where are your numbers?

[Notional]

Fixed!

 

Source?

 

Because Wiki says otherwise:

https://en.wikipedia.org/wiki/GeForce_700_series

[othertomperson]

When you see a 290x almost reaching a 980 ti of course peeps will be butthurt

 

The benchmark I saw was 290X and 980, not Ti.

[Prysin]

Source?

 

Because Wiki says otherwise:

https://en.wikipedia.org/wiki/GeForce_700_series

yeah, i cannot find those numbers anywere. May be when OCd to kingdom come. but hey, 290X can OC too... so that would be a fun comparison, wouldnt it....

[Prysin]

The benchmark I saw was 290X and 980, not Ti.

there was a comparison between a 980Ti and FuryX....

 

you know, when you note down the FPS numbers for both the 980Ti and 290X in the same game with the same settings, its not hard to compare them really......

 

Also, not hard to imagine 290X being close, it has exactly the same amount of FLOPs as the 980Ti... So computational power is equal. 980Ti has higher pixel rate by far (nearly double)

[Briggsy]

To confirm what @Prysin said:

 

https://www.techpowerup.com/gpudb/2512/geforce-gtx-780-ti.html

https://www.techpowerup.com/gpudb/2549/geforce-gtx-titan-black.html

 

5040 GFLOPS (780TI) and

5121 GFLOPS (Titan Black)

 

for reference models at least.

 

https://www.techpowerup.com/gpudb/2460/radeon-r9-290x.html

 

5632 GFLOPS (290x)

[Prysin]

https://www.techpowerup.com/gpudb/2512/geforce-gtx-780-ti.html

https://www.techpowerup.com/gpudb/2549/geforce-gtx-titan-black.html

 

5040 GFLOPS (780TI) and

5121 GFLOPS (Titan Black)

 

for reference models at least.

i already said that. We are just calling bullshit on @DuckDodgers more then suspicious numbers.

 

But hey, if he is talking custom models, why not just look at 290X MSI Lighting w/OC.... just to keep things fair i mean....

[Kinda Bottlenecked]

The benchmark I saw was 290X and 980, not Ti.

 

Oh? I thought many have seen this? http://arstechnica.com/gaming/2015/08/directx-12-tested-an-early-win-for-amd-and-disappointment-for-nvidia/

 

 

[Briggsy]

i already said that. We are just calling bullshit on @DuckDodgers more then suspicious numbers.

 

But hey, if he is talking custom models, why not just look at 290X MSI Lighting w/OC.... just to keep things fair i mean....

 

I know, I was just linking to techpowerup. definitely more reliable source than wikipedia for sure. fixed my post to reflect that.