"Inevitable Bleak Outcome for nVidia's Cuda + Physx Strategy"

[Keysplayr]

Hey guys, there has been so many areas discussed in this thread, I wonder if we can make a cliff notes thread and cover all the points discussed here. Save for the personal attacks here and there, (shame) this has been a pretty good thread.
Game? A lot of misconceptions about CUDA, PhysX, OpenCL, DirectX Compute have been addressed.

 

[dadach]

its about processor speed rather than a number of cores once its past dual core...unless the game is specifically written for more cores

 

[Keysplayr]

Originally posted by: dadach
its about processor speed rather than a number of cores once its past dual core...unless the game is specifically written for more cores

Ok, so you mean a 200MHz difference on the uncore, can make up for the performance of an entire 4th core? I don't think so if CPU utilization was 100% across the board on the tri-core and quad-core CPU's. Apparently, it was not. I can see the Dual core being utilized at 100%, and a very noticeable difference going to tri (if Intel had one).

 

[Scali]

Originally posted by: dadach
its about processor speed rather than a number of cores once its past dual core...unless the game is specifically written for more cores

Wait, this was Far Cry 2, which was hailed as one of the games specifically getting relatively good performance from quadcores.

The answer is:
It's still about processor speed (and cache/bandwidth) more than the number of cores past dualcore, EVEN if the game is specifically written for more cores.

I think people are underestimating and even insulting game developers.
Multi-core/multi-CPU isn't exactly new. John Carmack experimented with dual-CPU systems back in the Quake 3 days.
Dualcores have been commonplace for years now, and developers could easily have gotten systems with two or more dualcore CPUs to get systems with more than two cores to optimize their software on. Even quadcores have been mainstream for a while now.
Saying that developers haven't yet started to optimize for multicore is not only short-sighted, it's downright insulting.
Developers have been trying for years (plenty of information from interviews with people like John Carmack, Tim Sweeney and Gabe Newell on sites like Anandtech, discussing how they are working on the multicore problem).
The problem is just that multicore isn't a good solution for most software. The most parallel tasks in a games are graphics and physics. We already have special parallel processors for those, leaving the CPU with the hard-to-parallelize-and-synchronize problems, which just don't scale well by definition.

In other words: I think by now we've seen most of the improvements we're going to get from dualcore/quadcore systems. I don't expect future engines to be significantly better than what current engines alraedy do. They've already managed to squeeze 30-40% extra performance out of multicores, that's probably about as good as it gets.

 

[BenSkywalker]

I said "benefits of accelerated Physx", not benefits of accelerated physics. If the PS3 is capable of accelerating every other physics API, then it's not really an advantage of Physx, is it?

PhysX is turnkey, massively threaded and free. Spend some time working with Cell, you'll soon discover that those traits are not to be marginalized

PhysX started off as a C++ library, and x86 was already supported by PhysX before Nvidia got involved. It's not like any other platform had to adopt Nvidia's Cuda technology to use PhysX.

All the other platforms had low level functionality as publicly available information, none of the current GPUs do. It isn't like we know what ATi is using for an instruction set, nVidia does document how CUDA is mapped compared to normal function calls, ATi could keep all of their design secrets secret if they ported CUDA over while having full functionality. There really isn't another viable approach unless they support OpenCL at some point in the future, but we don't know when/if that will happen for given parts(I assume that their current parts will, but to what level we don't know).

PhysX is not nearly the dominant PC gaming platform like DirecX is.

Given the state of PC gaming, PhysX runs on more gaming devoted hardware then DirectX does currently.

From an even different perspective, if Nvidia was a major player in the x86 market, and created a compiler optimized for their HW, would anyone call Intel stupid for refusing to adopt NV's compiler and instead pushing their own?

Of course not, Intel would absolutely be expected to push their own standard and push it very hard. You have a very good point there, now if you could just link me ATi's alternative

I've highlighted the relevant portion for you. OpenCL from Nvidia is basically a PR thing, as it's not publicly available for consumption.

I had a smart @ss comment typed out for this, but decided instead to be a bit more diplomatic. Can you please show me where to get any OpenCL application right now. Any at all. Just go ahead and link them up for us all. The fact is nVidia is getting the ball rolling on OpenCL- they have shown pretty much all the initiative with the standard. They are the ones helping developers get the applications going. Where is ATi? They are MIA in the actual pushing of OpenCL outside of their PR department talking.

This is from Nigel Dessau - SVP at AMD. I think he pretty much answered your question.

You do realize he is a senior VP of marketing? Why don't you try and get some OpenCL drivers from AMD, tell me how much luck you have. One company is talking a big game, another company is actually doing it- from the initial development, chairing the group and pushing development. One company talks, the other does. The sad thing is, the company doing all the talking is the one saying the other company is against the standard. Since, you know, they only built it, chair it, and support it. How can that hope to compete with a marketing guy talking?

 

[SunnyD]

Originally posted by: BenSkywalker

I've highlighted the relevant portion for you. OpenCL from Nvidia is basically a PR thing, as it's not publicly available for consumption.

I had a smart @ss comment typed out for this, but decided instead to be a bit more diplomatic. Can you please show me where to get any OpenCL application right now. Any at all. Just go ahead and link them up for us all. The fact is nVidia is getting the ball rolling on OpenCL- they have shown pretty much all the initiative with the standard. They are the ones helping developers get the applications going. Where is ATi? They are MIA in the actual pushing of OpenCL outside of their PR department talking.

And your point? Like I said, AMD is taking a typical approach to "beta testing". Go talk to them about getting an OpenCL driver stack, not me.

This is from Nigel Dessau - SVP at AMD. I think he pretty much answered your question.

You do realize he is a senior VP of marketing? Why don't you try and get some OpenCL drivers from AMD, tell me how much luck you have. One company is talking a big game, another company is actually doing it- from the initial development, chairing the group and pushing development. One company talks, the other does. The sad thing is, the company doing all the talking is the one saying the other company is against the standard. Since, you know, they only built it, chair it, and support it. How can that hope to compete with a marketing guy talking?

[/quote]

Like I said, I can talk the PR game too, just like Nvidia. I do know Nigel is a marketing veep. You think AMD isn't heavily involved in the OpenCL panel? You think Nvidia is the "chair" of the group? Funny. Unless you actually go talk to AMD's relations team, you have no grounds to comment further. There's a lot more than "talk" going on on the AMD side.

 

[Wreckage]

Originally posted by: SunnyD
There's a lot more than "talk" going on on the AMD side.

Well let us know when their is some action going on at AMD.

 

[SirPauly]

Originally posted by: SunnyD

Originally posted by: BenSkywalker

I've highlighted the relevant portion for you. OpenCL from Nvidia is basically a PR thing, as it's not publicly available for consumption.

I had a smart @ss comment typed out for this, but decided instead to be a bit more diplomatic. Can you please show me where to get any OpenCL application right now. Any at all. Just go ahead and link them up for us all. The fact is nVidia is getting the ball rolling on OpenCL- they have shown pretty much all the initiative with the standard. They are the ones helping developers get the applications going. Where is ATi? They are MIA in the actual pushing of OpenCL outside of their PR department talking.

And your point? Like I said, AMD is taking a typical approach to "beta testing". Go talk to them about getting an OpenCL driver stack, not me.

This is from Nigel Dessau - SVP at AMD. I think he pretty much answered your question.

You do realize he is a senior VP of marketing? Why don't you try and get some OpenCL drivers from AMD, tell me how much luck you have. One company is talking a big game, another company is actually doing it- from the initial development, chairing the group and pushing development. One company talks, the other does. The sad thing is, the company doing all the talking is the one saying the other company is against the standard. Since, you know, they only built it, chair it, and support it. How can that hope to compete with a marketing guy talking?

Like I said, I can talk the PR game too, just like Nvidia. I do know Nigel is a marketing veep. You think AMD isn't heavily involved in the OpenCL panel? You think Nvidia is the "chair" of the group? Funny. Unless you actually go talk to AMD's relations team, you have no grounds to comment further. There's a lot more than "talk" going on on the AMD side.

I am not going to disagree that there isn't more than talk going on with AMD but the key to me is nVidia is walking and talking and yet this is just a PR game to you. PR game to me is defined as wind, talk, virtually no examples and substance to back claims of talk -- this isn't nvidia, and by your own wording is for AMD at this time to me.

This opposition to nVidia is quite odd and needs to be studied to me. Even though I am pro-GPGPU and behind hardware accelerated Physics, still under the illusion it is marginal based on the need for more compelling examples to showcase for the consumer.

 

[SunnyD]

Originally posted by: SirPauly
I am not going to disagree that there isn't more than talk going on with AMD but the key to me is nVidia is walking and talking and yet this is just a PR game to you. PR game to me is defined as wind, talk, virtually no examples and substance to back claims of talk -- this isn't nvidia, and by your own wording is for AMD at this time to me.

Well, without being any further blunt... I am not able discuss any specifics of what AMD is doing. Fair enough?

I know everyone else here is going to jump all over that and try to call it out. But anyone in the industry knows why I have to leave it at that, and can only rely on already released information to pass along.

 

[Scali]

I don't think anyone needs to question that ATi, nVidia and Intel are all working on OpenCL.
But as far as I can see, nVidia is closest to getting OpenCL to the end-user.
nVidia also had a time when they offered their driver and SDK only to 'select developers'. But a few days ago, they made it available to all registered developers. The next step is probably a public beta for everyone, and finally the official release.

AMD is currently in the 'select developers' stage. They'll probably offer their drivers and SDK to a larger audience at some point, but it seems that their product isn't as mature as nVidia's yet at this point. They will get there eventually, it may just take a few months more than it takes nVidia.

The thing is just that nVidia's quick adoption of the OpenCL standard is in stark contrast with many things that people have said in this thread and on the blog.
nVidia doesn't see OpenCL as a competing solution to C for Cuda, but rather as a companion, as yet another language to promote their Cuda architecture and bring it to a wider audience.

 

[Wreckage]

Originally posted by: SunnyD

Well, without being any further blunt... I am not able discuss any specifics of what AMD is doing. Fair enough?

Do you work for an AMD focus group or something?



You have a week off due to your recent baiting, which you have been warned numerous times about in the past, as well as the other trolling/misleading statements you've made in the past day.

AmberClad
Video Moderator

 

[BenSkywalker]

You think AMD isn't heavily involved in the OpenCL panel? You think Nvidia is the "chair" of the group?

Yes I do think that nVidia holds the chair position for OpenCL. I have this funny habit of thinking about reality-

"The opportunity to effectively unlock the capabilities of new generations of programmable computer and graphics processors drove the unprecedented level of cooperation to refine the initial proposal from Apple into the ratified OpenCL 1.0 specification," said Neil Trevett, chairman of the OpenCL working group, president of the Khronos Group, and a vice president at Nvidia.

Link.

The fact that Trevett has held the chair for OpenCL since the position was created is common knowledge and was never anything resembling an insider secret in any way. You who champion OpenCL don't know this? The president of Khronos, the chair of OpenCL- is a nVidia executive. This is a bit different then AMD's approach, he is the VP of content creation, not marketing, so perhaps that is where the confusion lies?

Unless you actually go talk to AMD's relations team, you have no grounds to comment further. There's a lot more than "talk" going on on the AMD side.

So far along in development that they are in a closed beta that is even under NDA for the fact that it exists? They really must be pushing the standard forward, heh.

 

[habbakuk87]

Originally posted by: BenSkywalker

PhysX started off as a C++ library, and x86 was already supported by PhysX before Nvidia got involved. It's not like any other platform had to adopt Nvidia's Cuda technology to use PhysX.

All the other platforms had low level functionality as publicly available information, none of the current GPUs do. It isn't like we know what ATi is using for an instruction set, nVidia does document how CUDA is mapped compared to normal function calls, ATi could keep all of their design secrets secret if they ported CUDA over while having full functionality. There really isn't another viable approach unless they support OpenCL at some point in the future, but we don't know when/if that will happen for given parts(I assume that their current parts will, but to what level we don't know).

AFAIK Ati has already released the documentation to instruction set of R700 series GPU in addition to R600.

 

[BenSkywalker]

AFAIK Ati has already released the documentation to instruction set of R700 series GPU in addition to R600.

This? Not the level of information you would need to get something like CUDA running properly. I guess saying we don't know what they are using for an instruction set wasn't accurate in an absolute sense, a better way to state it is we don't know exactly how that relates to the hardware available. We don't know enough about low level functionality, and in no way am I knocking them for this btw, they have released as much information as nV has, just in the GPU market you don't let out all of your secrets

 

[Scali]

Originally posted by: habbakuk87
AFAIK Ati has already released the documentation to instruction set of R700 series GPU in addition to R600.

Yea, the Stream SDK is freely available and contains all that sort of info.
But that would imply that nVidia would have to use ATi's proprietary GPGPU solution to build a PhysX implementation that only ATi benefits from.
It should be pretty obvious that nVidia is not interested in investing resources into helping their competitor.

OpenCL is a slightly different story, because it isn't specifically for ATi.
The other platforms already had support before nVidia even acquired PhysX.
In fact, these platforms had support before Ageia acquired NovodeX and rebranded it to PhysX

I have personally played with the SDK when it was still NovodeX, and didn't have any acceleration on PC.
Then when Ageia came out with the PhysX PPU, I was amazed and excited at the potential. I've supported accelerated physics from that moment on, long before nVidia even made any attempts at delivering a GPU solution for physics.

 

[SSChevy2001]

Originally posted by: Scali

Originally posted by: SSChevy2001
Are you serious? Why do people come to sites like anandtech?

It seems you didn't understand my question.

What's not to understand? I clearly gave you a link showing a title that favors a quad core by more than 5%.

Here's another for you and keys, but this one shows X3 720 getting is ass kick by X4 920. It's clearly shows a title that scales well even from 3 -> 4 cores.
http://www.firingsquad.com/har..._performance/page9.asp

Originally posted by: Scali
Okay, so if we assume the dualcore had 100% CPU usage, that would translate to 50% CPU usage for the quadcore.
You now see 30% increase in performance.
That would roughly mean you now have 65% usage. In other words, the extra 2 cores have 15% load (so 7.5% each).
How is that acceptable, when just now you were ranting that quadcores weren't used properly, with your line of argument being that you don't see anywhere near 100% usage in Task Manager?

Where did I ever say 100% usage on quadcore? Read my post again.

Originally posted by: SSChevy2001
Yes I know what taskmgr is. While I'm not a developer, but I can understand that you just can't break up a game to run 100% on a quad core. What I expect to see is more titles favoring quad core cpus by more than just by 5%.

Originally posted by: Scali

Originally posted by: SSChevy2001
Read the links I posted above or look for more yourself. I was shocked the first time I played it on my 4870 that everything worked just fine until that 1st glass section.

Well maybe with a quadcore. I have a dualcore as I said, and for me the whole game crawls without PhysX acceleration. Not a problem for me, since I have a GeForce to fix that. Imagine if I had to buy a new CPU for better physics effect... and still not get it playable at all times

I have no problem with GPU PhysX, which is why I still have my 8800GTS. Forcing an all or nothing deal on the effects that can run on current cpus is another story.

 

[Scali]

Originally posted by: SSChevy2001
What's not to understand? I clearly gave you a link showing a title that favors a quad core by more than 5%.

Well, you started by complaining that your CPU 'only' showed 50% usage, from which you somehow drew the conclusion that the CPU isn't used effectively.
I tried to explain why the CPU usage itself doesn't tell you much.

I never claimed that quadcores can't boost a game by more than 5%. So there's no need to prove that to me.

What I was trying to explain is that even though a quadcore may only show 50% usage or so, that may actually be as good as it gets (depends on the game/situation), because the other 50% of the time the cores may be waiting for hardware or synchronizing between different parts of the engine etc (eg sound effects can only be played AFTER you know if something happened that creates sound... eg the user has fired a gun, or a car has crashed into a wall or something. So one has to wait for the other).

Originally posted by: SSChevy2001
I have no problem with GPU PhysX, which is why I still have my 8800GTS. Forcing an all or nothing deal on the effects that can run on current cpus is another story.

I think this is more a case of PC ports of console games in general.
Mirror's Edge certainly isn't alone in this. If we disregard the extra PhysX effects for a moment, Mirror's Edge on the PC is exactly the same as on XBox or PS3. This means that the game is lighter than it could be on PC. In this case nVidia has added special PhysX effects to the PC version, that's your 'all'. Otherwise it would be a direct port of the console version, and you would indeed get 'nothing', because the original game on the consoles never had 'all' to begin with. It's nVidia who added it, and naturally they only added it for THEIR hardware. PhysX is THEIR thing.

However, there are many other examples of games that are simply ported from consoles to PCs, and don't really challenge the latest PC hardware.
A game like Left 4 Dead isn't exactly heavy on a quadcore and high-end videocard either. My brother plays it on a Pentium 4 2.8 HT with a Radeon X1900XTX videocard, and it runs pretty good with nearly everything maxed out.

There aren't many games out there like Crysis, which really put the smack down on your system. That's just because consoles are very popular these days, and many developers give priority to the console versions, and don't really bother to put in extra effects for PCs with more powerful hardware.

 

[SunnyD]

Originally posted by: BenSkywalker
Yes I do think that nVidia holds the chair position for OpenCL. I have this funny habit of thinking about reality-

Strange, your link is the only link I can find that says Neil is the chair of the OpenCL group. Everything else I see points to Aaftab Munshi from Apple being the chair (Since the position was created incidentally).

Unless you actually go talk to AMD's relations team, you have no grounds to comment further. There's a lot more than "talk" going on on the AMD side.

So far along in development that they are in a closed beta that is even under NDA for the fact that it exists? They really must be pushing the standard forward, heh.

[/quote]

And what exactly are you getting at here? Other than a typical fanboy remark (which I consider odd coming from you), it makes no sense. So what if they do or don't require NDAs? So what if they're in a closed beta? I don't see how that is at all relevant to whether or not they are furthering OpenCL, since it obviously means they ARE committed to it in some fashion. Whether or not you like their practices is irrelevant to the point of whether they may or may not be supporting OpenCL.

 

[Scali]

Originally posted by: SunnyD
Strange, your link is the only link I can find that says Neil is the chair of the OpenCL group. Everything else I see points to Aaftab Munshi from Apple being the chair (Since the position was created incidentally).

What about this?
http://www.khronos.org/develop...ore%20Expo%20Mar09.pdf

 

[AmberClad]

The throwing around of insults and veiled jabs like "fanboy" have no place in this discussion. That goes for everyone.

AmberClad
Video Moderator

 

[Keysplayr]

Scali, look. These guys are getting all bent, because they've been touting OpenCL for the past few weeks/months. It's not acceptable that it is now known, that Nvidia chairs the OpenCL group. Not acceptable that OpenCL is very similar to C for CUDA. Not acceptable that PhysX is the only GPU PhysX game in town, for a long, long while. Cut them some slack, they have no where else to go. It's completely apparent in their postings and their personal attacks on you and others supporting Nvidia and CUDA and PhysX. It's all they got. Leave em be.

 

[BenSkywalker]

Strange, your link is the only link I can find that says Neil is the chair of the OpenCL group. Everything else I see points to Aaftab Munshi from Apple being the chair (Since the position was created incidentally).

I'm not sure where you get your information from, direct from Khronos. That wasn't exactly big detective work to get that link, I went to Khronos group and typed in his name

So what if they do or don't require NDAs? So what if they're in a closed beta? I don't see how that is at all relevant to whether or not they are furthering OpenCL, since it obviously means they ARE committed to it in some fashion. Whether or not you like their practices is irrelevant to the point of whether they may or may not be supporting OpenCL.

nVidia is the one that is pushing the standard, they are the ones chairing the board- if my comments may sound a bit biased to you it may be because I'm rather tired of the loyalists for ATi acting like they are chosing a path to battle nVidia, by picking an open standard against awful proprietary PhysX. Reality is, PhysX is no more or less proprietary then Havok, and OpenCL support as it stands right now is further along for the green team then it is for red. People could make the exceptionally strong point that right now OpenCL makes significantly less difference then PhysX or DX10.1 so why do they need to have full support, and that would be a very valid point. What is not valid is in the insinuation that OpenCL is AMD's way of battling nV. nV is ahead of them in every visible way, and given that what OpenCL needs to make it it support, the fact that one company is pushing support out in a timely fashion without secrecy surrounding their every move is quite telling to me. Why hide the fact that you are working on drivers for a platform that you have been pushing? It makes almost no sense.

 

[Atechie]

This is stupid....the same people whining over "PhysX locks me to NVDIA, boohoo"...
Those are the same peeps how where whining "Why do I need to buy an AGEIA PPU, couldn't my GPU do it, boohoo?"

Must be hard being so scared of progress...

 

[ronnn]

Originally posted by: Keysplayr
Not acceptable that PhysX is the only GPU PhysX game in town, for a long, long while.

Sure there must be more than one game supported. Which of course was the point of that article - maybe next gen will be able to support physics and be realistically useful at the same time. Won't come too soon for nvidia.

 

[Melted Rabbit]

http://www.khronos.org/files/o...ick-reference-card.pdf

Page 4
OpenCL is a trademark of Apple Inc. and is used under license by Khronos.


http://www.khronos.org/develop...ew/opencl_overview.pdf

Page 6
? Apple initially proposed and is very active in the working group
- Serving as specification editor


http://www.khronos.org/develop...view%20GDC%20Mar09.pdf
This document may be a bit misleading as this is Mr. Trevett's bio from here:
http://www.khronos.org/develop...ary/bios/trevett_neil/

Trevett, Neil - Vice President Embedded Content, NVIDIA

Neil Trevett joined NVIDIA in May of 2005. As Vice President of Embedded Content, he is responsible for enabling and encouraging compelling applications on non-PC platforms, including cell phones and automobiles. Prior to joining NVIDIA, Mr. Trevett has spent over twenty years in the 3D graphics industry including Vice President positions at 3Dlabs. Mr. Trevett is currently the elected President of the Khronos Group where Mr. Trevett created and chairs the OpenGL ES working group that has defined the industry standard for 3D graphics on embedded devices. Mr. Trevett holds a Bachelor of Science degree in Computer Science and Electronic Engineering from the University of Birmingham in the UK. (Permalink)

Mr. Trevett may chair the OpenCL group, but his main focus within nVidia and the Khronos Group has been embedded platforms and not PC platforms. OpenCL has applications in the embedded field as well as PC GPU applications. Then there is Apple, which holds the trademark for OpenCL and initially proposed and did the final editing for the current OpenCL standard.