AMD CPU speculation... and expert conjecture

[jdwii]

The only reason AMD have a shot with Mantle is because of the market share the consoles will give them imo, some seem to forget that there will be alot more x86/gcn apus around the world soon.
edit I read nvidia 'snubbed' the consoles because margins were too low, it makes sense now, AMD completely undercut them.

No nvidia was incapable of providing what they needed at the time. X86+GPU from 1 source

 

[de5_Roy]

But my question is why is everyone assuming it's AMD trying to skirt balanced system benchmarks when it's perhaps EA and DICE trying to hide the fact that they did a lazy job of implementing Mantle in BF4.

both can be true. starswarm bench is a demonstration while bf4 is a first-attempt implementation in a game. tech demo's gonna seem more optimized.

the links got me interested enough to google starswarm and search for a bit more info on oxidegames, ars technica and elementalgames forums. turns out oxide games have 3 games in development, based on nitrous engine. the only title i found was star control and in the ars article, 2015 is mentioned as a time "when they will have a good idea about progress"... so that's when i think star swarm level visual improvement is coming. i just skimmed over the info, coulda missed something vital. however, the closest-to-release new game with mantle support seems to be the new thief game. it'll also have tress fx support. on the oxidegames site (juanrga posted the link already) mentions that gpu-bottlenecked systems won't be seeing big gains... in a rather roundabout way. at least that's how i read it. if the baseline was already high enough, then it won't matter.

But I've given you evidence that Mantle can offer lots of improvement when using a "balanced" system. The problem with BF4 Mantle benchmarks isn't Mantle, but it's BF4. If Mantle was a disappointment in general, we'd see the same style results out of Star Swarm benchmark, with balanced systems and Intel quads not gaining anything.

"can offer" is the same as "has a lot of potential" i've been saying. i don't disagree with you. right now i'm looking forward to how the new thief game handles mantle (sans tressfx, reviewers can help it).

edit: i also read that frostbite 3 engine will be used in 15+ games. so, one more year at least, games are on the way, now it's up to amd to keep up the debugging and getting the updates out timely. but right now....no.

 

[wh3resmycar]

The issue here is AMD made an API, but did so without input from the rest of the players in the market. Traditionally, API's like that don't get adopted, and some OTHER API comes along later with industry wide support that takes over. That's what I expect we'll see at the end of the day: Intel and NVIDIA forcing MSFT to improve DX.

from: http://www.guru3d.com/articles_pages/amd_mantle_preview,3.html

"Our intention is for Mantle, or something that looks very much like it, to eventually become an industry standard applicable to multiple graphics architectures and platforms."

dx HAS to improve.

 

[palladin9479]

When the first accelerator cards were out very few games used them and many considered them a novelty at best. HW assisted 3D rendering was for CAD/CAM workstations only. That eventually proved false when Voodoo released the first widely successful card and paved the way for eventual standardization.

But its worth noting: While initially niche, the rendering process itself scales naturally. So once you got the API standardized, you had explosive growth and mass acceptance.

The issue here is one of standardization. If Intel and NVIDIA don't jump on board, or if MSFT improves DX [even if it isn't quite as much of a gain], Mantel dies a quiet death.

The issue here is AMD made an API, but did so without input from the rest of the players in the market. Traditionally, API's like that don't get adopted, and some OTHER API comes along later with industry wide support that takes over. That's what I expect we'll see at the end of the day: Intel and NVIDIA forcing MSFT to improve DX.

Hence I said

I expect the current iteration of Mantle to be like MMX / 3DNOW / Glide. Useful in some situations, especially the low cost ones, but not widely adopted. In the next two years we'll see several updates to the API and further refinements as developers start working with it and providing more feedback to what they want and what they need. Eventually it'll be replaced by another version of itself or by another API that does the same thing but better.

What we're seeing now is a technology demo. Some programs will use it, some won't, and it'll be hit or miss for awhile. There will be a version 2 and 3 later that will improve many of the problems and introduce new features. Other venders will get on board and it'll spread in popularity. Or another product will be introduced that does the same thing but significantly better. The original HW rasterizers didn't have a single standard API, each had it's own as each company was jockeying to be the leader. OpenGL has been designed around Unix CAD / CAM usage in industrial applications and originally didn't port well to a MS-DOS gaming. Eventually DX was created and OpenGL was updated to start supporting those needed features and the environment settled around Win32 and NT32.

That is the way technology evolves.

 

[8350rocks]

BF4, and all games before it, get around the draw call bottleneck by sending a lot of data at a time, rather then submitting a call every time you want to render a texture. Factor in the PCI-E latency, and you realize this is the CORRECT approach to take anyways, regardless of the overhead involved in draw calls.

As for DX, it wouldn't be that hard to introduce a new set of extensions that remove a glut of the abstraction layer, at the cost of universal code support.

Ask yourself why they didn't do that already and maybe you will understand why we need something else. There is a reason why Steam is being created, people are sick of Microsoft

Because that abstraction allows DirectX to work with EVERY GPU. Remove the abstraction layer, and you have to start using different code for different GPU's. That's the downside. OpenGL is no different in this regard.

FYI, MSFT already did remove a bit of the glut in the latest version of DX11; see BF4 on Win 7 versus Win 8 to see the effects of this.

^ not to mention there was word that their will be no Directx 12 which means Nvidia and Amd will have to take over

Which MSFT shot down. That was AMD farther trying to push their API by spreading FUD about the future of DX.

M$FT still has not put out an entirely 64 bit OS, and you expect that they will somehow suddenly rush to push a new version of DX that essentially rewrites the API entirely to further streamline their issues?

Not without getting closer to the metal.

Additionally, I have not heard of anything about DX12 or anything even remotely close...DX 11.2 still isn't fully adopted in the PC Gaming world. The cutting edge engines have support for it, though you have to do a lot of the work yourself still in some respects. With 11.2 being so new, I doubt we see DX12 any time in the near future. Additionally, I know of multiple major vendors with big name engines working out details to include Mantle support in their engines. The engine we are licensing for use in our MMO is one such engine, and once the tech becomes available we will be adding the Mantle backend into our rendering paths. We want our MMO to function natively on Win/OSX/Linux and we feel that Mantle will actually take that a tremendous step in the right direction by being able to abandon DX entirely.

That is a much more palatable system than what we are currently working with which is putting a wrapper on DX draw calls and having to middle man between DX and OGL. The performance losses are at an acceptable percentage in terms of performance (when there is a performance loss), though why not have the penalty erased entirely?

It is a brilliant idea, and if NVidia ever manages to pull their head out, they will get on board and back the open source project too. Though, that remains to be seen.

 

[palladin9479]

M$FT still has not put out an entirely 64 bit OS, and you expect that they will somehow suddenly rush to push a new version of DX that essentially rewrites the API entirely to further streamline their issues?

Huh? Starting with Windows XP Pro x64 there has been a native 64-bit OS by Microsoft for x86. Prior to that it was Windows NT Server 4.0 for DEC Alpha.

There is no such thing as "entirely 64 bit OS". An OS is either 64-bit or it's not, if it is then the Kernel will run 64-bit long and any privileged (Ring 0/1) software will have to be 64-bit also. If it's not then it'll run 32-bit mode and any privileged software will also be 32-bit.

 

[Cazalan]

This benchmark of Kaveri ES was used in this thread to predict that A10 would be at SB i5 level

http://browser.primatelabs.com/geekbench3/compare/226070?baseline=223722

Benchmarks of final silicon confirm the prediction

http://browser.primatelabs.com/geekbench3/compare/226070?baseline=375716

Another prediction confirmed :-D

That benchmark is practically useless. Scores are all over the place.

http://browser.primatelabs.com/geekbench3/compare/375716?baseline=390643

 

[Cazalan]

But my question is why is everyone assuming it's AMD trying to skirt balanced system benchmarks when it's perhaps EA and DICE trying to hide the fact that they did a lazy job of implementing Mantle in BF4.

But I've given you evidence that Mantle can offer lots of improvement when using a "balanced" system. The problem with BF4 Mantle benchmarks isn't Mantle, but it's BF4. If Mantle was a disappointment in general, we'd see the same style results out of Star Swarm benchmark, with balanced systems and Intel quads not gaining anything.

Those games couldn't be any more different though. BF4 is actually a playable game. StarSwarm is a pure demo which doesn't look anything like what the final game will be. The ship detail is like a game 10 years old. It's purely to demonstrate draw call maximization. There's little point in trying to compare the two.

 

[de5_Roy]

TSMC and GlobalFoundries Led Foundry Market in 2013 – IC Insights.
Top 13 Foundries Account for 91% of Total Foundry Sales in 2013
http://www.xbitlabs.com/news/other/display/20140129235543_TSMC_and_GlobalFoundries_Led_Foundry_Market_in_2013_IC_Insights.html

First look: AMD Mantle CPU performance in Battlefield 4
A new frontier for PC gaming?
http://techreport.com/review/25995/first-look-amd-mantle-cpu-performance-in-battlefield-4

 

[gamerk316]

Techreport noted something worth noting:

http://techreport.com/review/25995/first-look-amd-mantle-cpu-performance-in-battlefield-4/2

One thing we didn't expect to see was Nvidia's Direct3D driver performing so much better than AMD's. We don't often test different GPU brands in CPU-constrained scenarios, but perhaps we should. Looks like Nvidia has done quite a bit of work polishing its D3D driver for low CPU overhead.

Which shows up most in this chart:

Note the results at both the high and low end:

At the low end, the 780 TI VASTLY outperforms the R9 290x with a A10-7850k (93 vs 69 FPS). Enable Mantel though, and you start getting 110 FPS.

At the high end, the 780 TI outperforms the R9 290x with an i7 4770k (118 vs 145 FPS). Enable Mantel, and NVIDIA remains ahead, as AMD still only gets 130 FPS.

So here's another line of investigation that probably needs to be done: Is there a SIGNIFICANT different in driver overhead for both AMD and NVIDIA, which is actually making the Mantel results overinflated due to a buggy D3D driver? Because looking at the A10 results, I don't think anyone is going to argue that a 780 TI is 33% faster then the R9 290x. Something else is going on here...

 

[-Fran-]

Techreport noted something worth noting:

http://techreport.com/review/25995/first-look-amd-mantle-cpu-performance-in-battlefield-4/2

One thing we didn't expect to see was Nvidia's Direct3D driver performing so much better than AMD's. We don't often test different GPU brands in CPU-constrained scenarios, but perhaps we should. Looks like Nvidia has done quite a bit of work polishing its D3D driver for low CPU overhead.

Which shows up most in this chart:

Note the results at both the high and low end:

At the low end, the 780 TI VASTLY outperforms the R9 290x with a A10-7850k (93 vs 69 FPS). Enable Mantel though, and you start getting 110 FPS.

At the high end, the 780 TI outperforms the R9 290x with an i7 4770k (118 vs 145 FPS). Enable Mantel, and NVIDIA remains ahead, as AMD still only gets 130 FPS.

So here's another line of investigation that probably needs to be done: Is there a SIGNIFICANT different in driver overhead for both AMD and NVIDIA, which is actually making the Mantel results overinflated due to a buggy D3D driver? Because looking at the A10 results, I don't think anyone is going to argue that a 780 TI is 33% faster then the R9 290x. Something else is going on here...

From what I read somewhere, it seems that nVidia has implemented (for DX11) mutithreaded support nicely and AMD hasn't polished it. There's no way for me to investigate that, but seems to be a sound reason for the way MANTLE scales in poor CPU scenarios against upper tier CPUs compared to nVidia. How can that be explored?

Cheers!

 

[gamerk316]

You'd have to compare, in a CPU bound situation, a NVIDIA GPU against a similar AMD GPU. It's quite clear, based on this result at least, NVIDIA has the higher performing driver. As such, one would reason any gains they would get using a lower level API would be much lower then what AMD is getting.

Now, I'm not going to suggest AMD intentionally wrote a crappy DX driver in order to play up Mantel, but on the other hand, if the current situation was reversed, how many people would be claiming NVIDIA did, in fact, just that?

 

[juggernautxtr]

to me that chart looks pretty good for the 7850k to even come to 100fps with added dgpu.
amd needs to step up on drivers, for a low end desk top parts it actually quite impressive. even if it isn't on par with the high end part.

wonder if that is stock or overclocked bench though.

 

[gamerk316]

One other point I want to make: Has anyone bothered to do a graphics fidelity comparison? It IS a totally different API after all...

 

[Cazalan]

So here's another line of investigation that probably needs to be done: Is there a SIGNIFICANT different in driver overhead for both AMD and NVIDIA, which is actually making the Mantel results overinflated due to a buggy D3D driver? Because looking at the A10 results, I don't think anyone is going to argue that a 780 TI is 33% faster then the R9 290x. Something else is going on here...

Looks like that will keep reviewers busy for quite some time. Having to maintain 2 sets of drivers is an extra burden for AMD.

 

[Cazalan]

One other point I want to make: Has anyone bothered to do a graphics fidelity comparison? It IS a totally different API after all...

Like this?

http://wccftech.com/amd-mantle-washed-graphics-compared-direct-battlefield-4/

DICE is taking the blame for it though.

 

[de5_Roy]

Like this?

http://wccftech.com/amd-mantle-washed-graphics-compared-direct-battlefield-4/

DICE is taking the blame for it though.

it makes the promo slide at the top of the article (edit: mildly) ironic.

 

[juanrga]

This benchmark of Kaveri ES was used in this thread to predict that A10 would be at SB i5 level

http://browser.primatelabs.com/geekbench3/compare/226070?baseline=223722

Benchmarks of final silicon confirm the prediction

http://browser.primatelabs.com/geekbench3/compare/226070?baseline=375716

Another prediction confirmed :-D

That benchmark is practically useless. Scores are all over the place.

http://browser.primatelabs.com/geekbench3/compare/375716?baseline=390643

It was previously mentioned this benchmark is very sensitive to memory speed

http://browser.primatelabs.com/geekbench3/compare/226070?baseline=390643

Techreport noted something worth noting:

http://techreport.com/review/25995/first-look-amd-mantle-cpu-performance-in-battlefield-4/2

One thing we didn't expect to see was Nvidia's Direct3D driver performing so much better than AMD's. We don't often test different GPU brands in CPU-constrained scenarios, but perhaps we should. Looks like Nvidia has done quite a bit of work polishing its D3D driver for low CPU overhead.

Which shows up most in this chart:

Hum, I like more this part from techreport:

Of course, Nvidia has known for months, like the rest of us, that a Mantle-enabled version of BF4 was on the way. You can imagine that this game became a pretty important target of optimization for them during that span. Looks like their work has paid off handsomely.

Therefore Nvidia has really made a benchmarketing optimization. It works for this game but not for others.

 

[gamerk316]

Therefore Nvidia has really made a benchmarketing optimization. It works for this game but not for others.

Possibly, but we'd more Mantel enhanced titles to prove that.

 

[Cazalan]

Don't most major games have custom profiles for them now anyway? That's not unique to NVidia.

 

[juanrga]

AMD's Feldman interview

http://venturebeat.com/2014/02/04/amd-micro-server-boss-welcomes-the-age-of-arm-based-servers-interview/view-all/

He mentions again that AMD will be no more competing against Intel servers with powerhungry CPUs:

AMD plans to disrupt the huge x86 (Intel-based) server chip market by using smaller and more power efficient ARM cores in AMD-designed server processors. [...] AMD isn’t going toe to toe with Intel in designing super-fast server chips anymore. And that may be its path to riches, according to Feldman. Its chips won’t need giant air conditioning systems.

I interpret this as another way to say that there is no plans for Steamroller/Excavator/... Opterons (which implies no Steamroller/Excavator/... FX). If someone disagree he or she can sign this petition

https://www.change.org/petitions/advanced-micro-devices-amd-amd-release-high-end-fx-processor-with-steamroller-microarchitecture#share

 

[Rum]

You'd have to compare, in a CPU bound situation, a NVIDIA GPU against a similar AMD GPU. It's quite clear, based on this result at least, NVIDIA has the higher performing driver. As such, one would reason any gains they would get using a lower level API would be much lower then what AMD is getting.

Now, I'm not going to suggest AMD intentionally wrote a crappy DX driver in order to play up Mantel, but on the other hand, if the current situation was reversed, how many people would be claiming NVIDIA did, in fact, just that?

If the performance disparity was greater I would say yes to them playing up Mantle, but 12 FPS? Ya figure they would go for a bigger gain if they wanted to make it look super awesome! Insert (Nvidia) I feel the same way there is no reason to tank your performance numbers in an effort to sell an api on quite possibly one of the worst show case games. BF4 is a buggy turd.
AMD will refine their drivers to perform better in D3D and continue to refine Mantle. Kudos to Nvidia for killing it with their D3D implementation.

 

[jdwii]

Found out my 6950hd alone has more TFLOPS then the PS4. Man this gen is so disappointing i have no understanding why anyone would even want a console anymore. I really could build a 600$ PC that could do more.

I was thinking Nintendo would be a genus if they built a steambox for their next gen think about it microsoft and sony would still never see a Nintendo game on their system and this could even help Nintendo with 3rd party as well as online play.

 

[esrever]

Found out my 6950hd alone has more TFLOPS then the PS4. Man this gen is so disappointing i have no understanding why anyone would even want a console anymore. I really could build a 600$ PC that could do more.

I was thinking Nintendo would be a genus if they built a steambox for their next gen think about it microsoft and sony would still never see a Nintendo game on their system and this could even help Nintendo with 3rd party as well as online play.

What good are TFLOPS? You can't play TFLOPS.

 

[palladin9479]

Found out my 6950hd alone has more TFLOPS then the PS4. Man this gen is so disappointing i have no understanding why anyone would even want a console anymore. I really could build a 600$ PC that could do more.

I was thinking Nintendo would be a genus if they built a steambox for their next gen think about it microsoft and sony would still never see a Nintendo game on their system and this could even help Nintendo with 3rd party as well as online play.

What good are TFLOPS? You can't play TFLOPS.

Their like megawidgets and whizzlepops. Imaginary numbers used to indicate the theoretical vector performance of a particular processor. It has no bearing on real life applications because outside of science and data modeling, real applications use a mixture of scalar and vector with 90~99% of the code being scalar (Integer). All logical operations are integer operations, so any compares or jumps (ASM for If / Then type statements) are handled by the scalar processor.