NVidia's Dispute over Nehalem Licensing - SLI not involved

Like I said. I have seen no proof that there is additional latency for using the motherboard GPU to do final output. When you mesh Multi GPU latency into the same category like you did earlier. Then it implies a wider implification. As they are totally irrelevant to each other.

Anyway. Yes there is a performance loss from sending the final data to the IGP. I'll respond more deeply when I have more time. Because I do have some actual data regarding this. ((besides performance percentages)) and I currently use the 780A with a Phenom as my primary platform.

I still don't like it. Not enough available tech detail, I guess. Do we know the bandwith between the GPU & NVIO? I'd much prefer if the display was connected to the dGPU, with low power mode routing the opposite way from the mGPU under low load. This way running 3D apps under load doesn't compromise performance. This may not gel with their notions of headless dGPUs & output only via their own chipset mGPUs. Perhaps this is what Intel won't allow - somebody creating a lock-in on "their" platform....

You know until I heard about it here. The headless GPU thing was something I had never heard about. I really dont see any advantage to do this either. Specially when they claim to be CPU agnostic.

While the SLI market is one thing. Cutting off everyone who doesnt own an Nvidia chipset from using a GPU is pretty obsurd and I cant see Nvidia doing that.
 
Thanks, look forward to the info when you have the time.

You know until I heard about it here. The headless GPU thing was something I had never heard about. I really dont see any advantage to do this either. Specially when they claim to be CPU agnostic.
Only speculation on my part. There's no tech reason why not given they've invested in the tech esp with CUDA solutions not requiring direct display feeds & saves a few bucks on components. Nvidia may be CPU agnostic, but obviously less so with chipsets.

While the SLI market is one thing. Cutting off everyone who doesnt own an Nvidia chipset from using a GPU is pretty obsurd and I cant see Nvidia doing that.
I agree it's unlikely. There are also downsides with my reverse suggestion, mainly with less power reduction.
 
With no nvidia Nehalem chipsets/motherboards.... does this mean that nvidia will enable SLI for Intel motherboards?
 
I still disagree with your assertion Chris, but look forward to seeing any relevant data you can provide. I certainly don't have the tools necessary to measure the latency for sending data from the discrete graphics card's frame buffer to the IGP and out to the display port.
 
How can you disagree with something you have no experience with or documented data on? Are you aware of Broadcast which was built onto the Nforce 200 chipset? The entire purpose of this technology was too get rid of what you are talking about. Allowing the system memory and CPU to write to each GPU with 1 command. Its built into all Nforce 200 chipsets.

((Which the 780A/780I support so I wont be surprised if future intel chipsets have an Nforce 200 or a chipset like it)) Its an extremely efficient way of running all the GPUS at the same time without a need for alot of latency/CPU overhead.


 
How can you disagree with something you have no experience with or documented data on? Are you aware of Broadcast which was built onto the Nforce 200 chipset? The entire purpose of this technology was too get rid of what you are talking about. Allowing the system memory and CPU to write to each GPU with 1 command. Its built into all Nforce 200 chipsets.

((Which the 780A/780I support so I wont be surprised if future intel chipsets have an Nforce 200 or a chipset like it)) Its an extremely efficient way of running all the GPUS at the same time without a need for alot of latency/CPU overhead.



Chris:

disprove the following statement:

data transfers between a Geforce's framebuffer and the IGP do not occur instantaneously

Until you do so, you are wrong.

Your "no experience" and "lack of documented data" comments are red herrings. It doesn't take an EE or a physicist to determine that data transfers always have latency.
 
And thats a straw man's point at best. Your asking for something that can't be disproven or proven. And since your making the claim that it does the burden of proof falls on you. And I have yet to see anything from you other than speculative comments and irrelevant Multi GPU latency inclusions.

The point is NVidia's tech allows for a single command to drive all the data through the MCP Motherboard GPU just like it does through the Discreet GPU at one time without requesting data over and over. The performance loss is so minor its irrelevant. This so called "increased latency" argument is completely ridiculous.

There is TONS of information out there discussing the Broadcast technology as well tons of articles on Hybrid Power. And there is not a single editor who said "Hybrid Power brought forth increased latency which made multi GPU/Single GPU end user experience suffer"

Chris
 
HybridPower's transfer scheme is based on the bizarre assumption that anyone sane would play a game with either vsync off or triple buffering on. I'm sorry, but you're supposed to play your games with vsync on and triple buffering off, and you should optimize your settings so that you remain below a given multiple of the minimum frametime as often as possible. Anything else is heresy, and I am deeply unhappy that technologies made with other approaches in mind are being encouraged *shrugs*.

Mind you, this is still substantially less heretic than the complete joke that is Hybrid SLI/CrossFire (the performance aspects, not the power ones). Long live the 3D industry, where naive reviewers and overly excited users who misunderstand the very basics of 3D Graphics are the only judges! And no, I'm not bitter. No, really. I swear. Maybe. With all due respect.
 
HybridPower's transfer scheme is based on the bizarre assumption that anyone sane would play a game with either vsync off or triple buffering on. I'm sorry, but you're supposed to play your games with vsync on and triple buffering off, and you should optimize your settings so that you remain below a given multiple of the minimum frametime as often as possible. Anything else is heresy,

See Skinner not only are you Bizzare you are now a heretic :D
http://forum.beyond3d.com/showthread.php?t=42961&highlight=holy
 
Well, at least triple buffering isn't quite as grave of an offense as disabling vsync... ;)
 
And thats a straw man's point at best. Your asking for something that can't be disproven or proven. And since your making the claim that it does the burden of proof falls on you.

Well this is going nowhere. I don't expect *you* to agree with me, but I doubt anyone else is going to carry that banner for you.

Let's take a poll:

how many people believe Nvidia has figured out how to transfer data between the discrete graphics card and the IGP instantaneously?

And I have yet to see anything from you other than speculative comments and irrelevant Multi GPU latency inclusions.

I never said there's any relation between the two, I'm simply combining and continuing discussion on another latency issue. Latency + more latency = bad for RT computing. I can't believe you're actually arguing against this...

The point is NVidia's tech allows for a single command to drive all the data through the MCP Motherboard GPU just like it does through the Discreet GPU at one time without requesting data over and over. The performance loss is so minor its irrelevant. This so called "increased latency" argument is completely ridiculous.

The only thing ridiculous here is that you seem to think Nvidia's engineers are capable of engineering time machines.

There is TONS of information out there discussing the Broadcast technology as well tons of articles on Hybrid Power. And there is not a single editor who said "Hybrid Power brought forth increased latency which made multi GPU/Single GPU end user experience suffer"

Chris

Well that's all fine and dandy Chris, but now you're the one setting up straw men. I'm not claiming these two latency issues are related at all. It just so happens that two discussions on latency involving data transfer with discrete graphics have popped up around the same time, and I've mentioned the one in the course of discussing the other. I hope this clears up any confusion on the matter.
 
To satisfy ShaidarHaran, no I don't believe Nvidia can transfer data instantaneously :).

But that isn't necessary for achieving zero additional latency between user moving the mouse and the result showing up on the screen. If the transfer can be fully hidden behind other things that would be there anyway, then there's no additional latency.

The obvious candidate is vsync. If you've got a 60Hz vsync interval, and the discrete GPU can render frames at 80fps, you're still only going to be displaying a new frame at 60fps. With a three-frames-in-flight limit, you'll have about 3*16.67ms = 50ms of input-to-display latency (depending on exactly when during each frame the app samples keyboard/mouse input). Now if the transfer to the integrated GPU doesn't drop the framerate below 60fps, you're still going to have exactly that same latency.

Displaying from the integrated GPU adds no more latency than having a lower frame rate does. If it doesn't change the displayed frame rate, there's no additional latency.
 
Well this is going nowhere. I don't expect *you* to agree with me, but I doubt anyone else is going to carry that banner for you.

Let's take a poll:

how many people believe Nvidia has figured out how to transfer data between the discrete graphics card and the IGP instantaneously?



I never said there's any relation between the two, I'm simply combining and continuing discussion on another latency issue. Latency + more latency = bad for RT computing. I can't believe you're actually arguing against this...



The only thing ridiculous here is that you seem to think Nvidia's engineers are capable of engineering time machines.



Well that's all fine and dandy Chris, but now you're the one setting up straw men. I'm not claiming these two latency issues are related at all. It just so happens that two discussions on latency involving data transfer with discrete graphics have popped up around the same time, and I've mentioned the one in the course of discussing the other. I hope this clears up any confusion on the matter.

Yes it clears it up fine. But what does seem clear to me is you really don't understand the technology your addressing. There simply isnt enough overhead for using the IGP to do the final image scanout. The Nforce 200 chip was designed specifically for this purpose. For all the flak the chip gets it'll probably end up in future boards as well to ensure these issues you seem to think are happening dont happen.

I dont need anyone at beyond3d to "Carry" a banner for me. I'm more than capable of discussing points about 3D tech without anyone here being on my "side".

I've brought up 2 points which you have failed to address.

1) The Nforce 200 single commands all GPUS within the PCIE reducing system latency/overhead virtually completely.

2) The only latency lost /gained is the actual performance lost on FPS. Which is a MS issue on the framerate itself. Since the framerate loss is almost non existence at best. The performance loss because completely irrelevant. I addressed this in my very first post.

Mind you, this is still substantially less heretic than the complete joke that is Hybrid SLI/CrossFire (the performance aspects, not the power ones). Long live the 3D industry, where naive reviewers and overly excited users who misunderstand the very basics of 3D Graphics are the only judges! And no, I'm not bitter. No, really. I swear. Maybe. With all due respect.

You know. I dont remember any reviewer praising Geforce Boost mode. I dont even think anyone really even cares about it. When I covered Hybrid Power I ignored it totally. Hybrid SLI Boost mode is more of a "me too" mode than anything else. Despite it not being very useful to the consumer. Most people focused on Hybrid Power. A mode which can be quite awesome for a general user.

I really don't understand this resentment you seem to be showing. We're not all a bunch of stupid hethens who dont know how to play games or judge performance. Rys asked me a while back while I dont visit beyond3d much anymore. Its because of this kind of attitude thats displayed here by some of the senior members.

The fact is benchmarks are commonly judged by having vertical sync off. This is just the way things are. And there are good reasons for reviewers not enabling vsync. While there may be some value in showing FPS values with vsync on. The AVG users whom all major review sites are aiming at would likely not understand anyone the issue of latency and frame buffering to begin with and maintaining the consistent desire of running half your FPS value.
 
Last edited by a moderator:
To satisfy ShaidarHaran, no I don't believe Nvidia can transfer data instantaneously :).

But that isn't necessary for achieving zero additional latency between user moving the mouse and the result showing up on the screen. If the transfer can be fully hidden behind other things that would be there anyway, then there's no additional latency.

The obvious candidate is vsync. If you've got a 60Hz vsync interval, and the discrete GPU can render frames at 80fps, you're still only going to be displaying a new frame at 60fps. With a three-frames-in-flight limit, you'll have about 3*16.67ms = 50ms of input-to-display latency (depending on exactly when during each frame the app samples keyboard/mouse input). Now if the transfer to the integrated GPU doesn't drop the framerate below 60fps, you're still going to have exactly that same latency.

Displaying from the integrated GPU adds no more latency than having a lower frame rate does. If it doesn't change the displayed frame rate, there's no additional latency.

You raise a good point about vsync. It would be interesting to see just how many PC gamers use vsync, how often, and in which titles. I use it for offline FPSes but disable it for online play, and most other titles I disable it (usually strategy and simulation games have a framerate limiter anyway).
 
Yes it clears it up fine. But what does seem clear to me is you really don't understand the technology your addressing.

Come now Chris, you know better than that.

There simply isnt enough overhead for using the IGP to do the final image scanout. The Nforce 200 chip was designed specifically for this purpose. For all the flak the chip gets it'll probably end up in future boards as well to ensure these issues you seem to think are happening dont happen.

You are correct in that I have misunderstood the precise implementation of data transfer for display output utilized by NV, but this does not invalidate my argument. Data transfers take time. There's no getting around it. It may be "miniscule" from a human perspective, but when you have multiple latency-inducing factors (like AFR in SLI) it all adds up, and every little bit counts towards immersion. Everyone's eyes are different, so who are you to say that this latency is irrelevant?

1) The Nforce 200 single commands all GPUS within the PCIE reducing system latency/overhead virtually completely.

The keyword here being "virtually". It exists. Additional latency in RT computing is bad, no matter how negligible on an individual per-frame basis.

2) The only latency lost /gained is the actual performance lost on FPS. Which is a MS issue on the framerate itself. Since the framerate loss is almost non existence at best. The performance loss because completely irrelevant. I addressed this in my very first post.

And I've already addressed this several times.

No matter how you slice it, this move is NOT beneficial for the end-user. It is clear that were NV to implement such a rendering technique they would be doing so as a proprietary measure to preserve marketshare.
 
I dont understand the premise of your last paragraph. You say that this is bad for the consumer. But Hybrid Power tech is pretty sound right now. Being able to shut off discreet GPUS completely is a huge benefit for idling and desktop activity.

Despite the headaches and problems my Phenoms have given me at times. ((which I'll admit arent as bad as I make them out to be but at times irritate me)) This is a huge advantage to SLI users. The Hybrid Power feature is optional ((Not forced)) and you can always just hook your monitor up to the discreet GPU. Heck even disable the onboard IGP all together if you consider this minor MS increase in frametime a big deal.

I consider the fact that hybrid power eats up 256 megs to be its biggest problem right now. It's not one of those things that is forced on you. Just a perk/advantage of upcoming Intel/AMD chipsets that run on an Nvidia platform.
 
You know. I dont remember any reviewer praising Geforce Boost mode. I dont even think anyone really even cares about it.
OEMs care. Joe Consumer might buy a PC with it preferentially given NVIDIA's counter-intuitive naming scheme.

I really don't understand this resentment you seem to be showing. We're not all a bunch of stupid hethens who dont know how to play games or judge performance. Rys asked me a while back while I dont visit beyond3d much anymore. Its because of this kind of attitude thats displayed here by some of the senior members.
Of course not, and I apologize for being seemingly so extreme (and I agree Beyond3D is more elitist than worst, but meh, it shouldn't be exagerated - there's much much worse on nearly every subject out there); however, as you said yourself, many reviewers don't really bother investigating such things in great detail (by naive, I basically meant 'believe what they're told more often than they should and especially when they are not very interested in the technology as individuals') and so problems that are more perceptual in nature are unlikely to be noticed. Furthermore, even if you think you notice something, you might not dare mentioning it in case you're wrong.

Furthermore, as I said, HybridPower seems philosophically optimized(tm) for scenarios with vsync off as far as I can tell; heck, I even saw someone from NV bragging about how their solution resolved tearing. Ugh. It's not such a big deal, mind you, because I don't fully agree with ShaidarHaran: As he claims, it will definitely increase latency, however *if done right* it shouldn't be a major impact AFAICT (PCIe & memory latency aren't really big in the grand scheme of things). Whether it actually is done right, however, I don't know - so it's still well worth pondering upon...

The fact is benchmarks are commonly judged by having vertical sync off. This is just the way things are. And there are good reasons for reviewers not enabling vsync. While there may be some value in showing FPS values with vsync on. The AVG users whom all major review sites are aiming at would likely not understand anyone the issue of latency and frame buffering to begin with and maintaining the consistent desire of running half your FPS value.
Oh, absolutely, it's normal to benchmark with vsync off. What isn't fine is for a company to optimize a feature's implementation to make it most sensical in scenarios with vsync off, despite that making the most sense for reviewers, not real users.
 
Perhaps after the 18th. When I have more time. And hardware that I can discuss more openly. We can ponder that together. Assuming you might have some kind of test methodology which could expose any possible problems. I'd be interested in hearing and sharing my results with you. If you were actually still on MSN we could talk more freely perhaps.

But as of right now. I have mostly only noticed extremely trivial FPS loss. Usually within the range of 1% as most websites seemed to have observed.

Chris.
 
Bloody hell....
why would that be the case, and it that only when the discreet gpu is shut down or all the time?

No it requires 256 megs of video memory ((from system memory)) all the time. I dont recommend using it with anything less than 4 gigs of memory.

Chris
 
Back
Top