CUDA vs Clock Speed vs VRAM?

[omega_rugal]

cluthz

Don`t try too hard, anything you post to back up your argument NV will dismiss it as being "simply wrong" because for some reason only the benchmarks/reviews he posts are valid.

[NaturalViolence]

But anandtech agrees with me. In fact they were one of the sources I looked up when I was googling it. They even included an image to visually demonstrate it:



See. 8 chips with the same density. 2 chips per memory controller for 2 of the memory controllers and 4 wired to the last controller. Just like I described. They also have PCB shots that back this up.

And they flat out said that they don't use mixed IC densities like tomshardware claims:

anandtech Wrote:For the GTX 660 Ti in 2012 NVIDIA is once again going to use their asymmetrical memory technique in order to outfit the GTX 660 Ti with 2GB of memory on a 192bit bus, but they’re going to be implementing it slightly differently. Whereas the GTX 550 Ti mixed memory chip density in order to get 1GB out of 6 chips, the GTX 660 Ti will mix up the number of chips attached to each controller in order to get 2GB out of 8 chips. Specifically, there will be 4 chips instead of 2 attached to one of the memory controllers, while the other controllers will continue to have 2 chips. By doing it in this manner, this allows NVIDIA to use the same Hynix 2Gb chips they already use in the rest of the GTX 600 series, with the only high-level difference being the width of the bus connecting them.

And here they talk about the interleaving:

anandtech Wrote:Our best guess remains that NVIDIA is interleaving the lower 1.5GB of address while pushing the last 512MB of address space into the larger memory bank, but we don’t have any hard data to back it up.

They admit that they're just guessing. If they are correct then this statement:

anandtech Wrote:The best case scenario is always going to be that the entire 192bit bus is in use by interleaving a memory operation across all 3 controllers, giving the card 144GB/sec of memory bandwidth (192bit * 6GHz / 8). But that can only be done at up to 1.5GB of memory; the final 512MB of memory is attached to a single memory controller. This invokes the worst case scenario, where only 1 64-bit memory controller is in use and thereby reducing memory bandwidth to a much more modest 48GB/sec.

is valid. However tomshardware's description is still wrong.

omega_rugal Wrote:Don`t try too hard, anything you post to back up your argument NV will dismiss it as being "simply wrong" because for some reason only the benchmarks/reviews he posts are valid.

I'm surprised that you would say that considering you're one of the few people here to convince me that I was wrong about something using benchmarking data (the argument we had about willamette a few months back).

I respect this mans arguments because he is in fact backing them up with data. Which most don't bother to do. I was confused about the wording on one of his posts so he elaborated. When I asked how he determined this he posted his source. I then read through it and noticed some flaws. So I posted them. What's wrong with doing that?

[omega_rugal]

Well i just say that is more friendly to say "whoa whoa where did you get that?" instead of "man you are so wrong show me some proof" depends on how you say it.

[NaturalViolence]

I actually didn't even ask for proof this time since I wasn't entirely sure what he was saying. He provided it anyways. Which is a nice change of pace. All I did was ask him what he meant. As far as I can tell I wasn't impolite.

As you know these things usually start with someone making a ridiculous claim then multiple back and forth posts of me pestering them to provide the source. 90% of the time they don't have a source since they're just regurgitating whatever nonsense they heard on an internet forum. So they get pissed, claim that I'm wrong, call me an asshole or something along those lines, ask me to leave them alone, and then stop posting, still without a source. I am very grateful that I didn't have to go through any of that this time. Kudos to cluthz for being polite, elaborating, and providing a source right from the beginning. I wish more people were like you.

[omega_rugal]

I love when we can discuss as gentlemen

http://www.youtube.com/watch?v=UIgwY1pRzTw

[NaturalViolence]

We're pretty far off topic now but considering the OP had his question answered a long time ago and doesn't appear to be coming back I guess it's ok.

omega_rugal Wrote:I love when we can discuss as gentlemen

Yes. It doesn't happen often but when it does it's pretty glorious. I really do wish Internet forums were generally a bit more academic in their discussions. I know this is ultimately just a place to talk about dolphin and for users to ask questions related to it. But there is a lot of potential for people bettering their personal knowledge of underlying technology if they would be open to criticizing each other the same way we do in academia (though those debates certainly get heated sometimes as well).

Rant: (Show Spoiler)

Instead everyone either tries to avoid calling each other out due to fear of reprisal or lashes out at anyone who dares to criticize them with personal insults. And if you act like me then you become known as the local know-it-all or a dick just for asking people to provide a god damn source when asked. Or claiming that people should at least TRY to convince other people who disagree with them instead of being silent, which benefits no one since only one side gets a voice. Unless of course it's something purely subjective like appearance.

[cluthz]

Thanks for the clarification NV!

So, my initial assumption, that "the last" 512MB is addressed thru only 1 64 bit controller, thus with lower bandwidth remains valid then?

For an end user I guess how many IC and mixed density would not matter at all, but how much bandwidth that is available to the user. Although more experienced users uses sites like toms hardware to gain more knowledge, which this time around clearly is faulty in some respects.

I still cannot see how my first post was wrong, where i claim the first 1.5gb does what the full 192 bit and the last 0.5 only 64??

@omega
I'm not trying to pick a fight with NV or try to falsify what he is saying. I'm simply trying to actually find out how the bandwidth on the 660 series work, and NV is clearly a lot more knowledgeable then me (or pretty much everyone on most forums), but probably a bit more stubborn too

@nv
My initial point was that the 660s does have lesser bandwidth on the last 0.5gb. The mixed IC part came in after I cited toms hardware, simply because I did not want to cut too much of the paragraph. Your explanation, and anandtech does indeed make tom's explanation void, except for the lesser bandwidth.

Thanks for taking the time to explain this very detailed in parts. I'm no native english speaker, so if I'm sounding rude or dismissive I can ensure you that is not my intention.

[rpglord]

Yeah you are correct clutz because memory have to be interleaved in order to reech speed declared in gpu specifications.
And interleaving doesn't support asymmetrical design :
-for 192 its either 1.5 GB, 3 GB or 6 GB and so on,
-for 256-bit it's 1,2,4,8 GB
-for 320 its 1.25, 2.5, 5GB
-for 384 same as 192
-for 448 it's 1.75, 3.5, 7 GB
-for 512 same as 256
and so on.

Just look at the Titan and Geforce 780 for example.
They are 384-bit and have 6GB and 3 GB, which is what is symmetrical with 384-bit. Coincidence ? I don't think so...

So what happens if you don't follow this symmetrical design ? It seems nobody is 100% sure, but it seems quite probable that some part of memory won't run at full bandwith, such as last 512mb on geforce 660/660 ti

[NaturalViolence]

cluthz Wrote:So, my initial assumption, that "the last" 512MB is addressed thru only 1 64 bit controller, thus with lower bandwidth remains valid then?

Yes. Though how the interleaving is managed isn't entirely known.

cluthz Wrote:For an end user I guess how many IC and mixed density would not matter at all, but how much bandwidth that is available to the user. Although more experienced users uses sites like toms hardware to gain more knowledge, which this time around clearly is faulty in some respects.

Exactly. It doesn't matter to the end user but it's still wrong.

The performance hit is going to depend on how the application and drivers manage it. Most games of course can't use more than 1.5GB of video memory and for the ones that can the drivers will likely place the data that is accessed the least often in the last 512MB of memory address space. Which will likely reduce the performance hit to almost nothing in most scenarios.

cluthz Wrote:I still cannot see how my first post was wrong, where i claim the first 1.5gb does what the full 192 bit and the last 0.5 only 64??

Depends on whether you're talking about the physical or logical layer. At the physical layer that's wrong. At the logical layer that's right.

I didn't fully understand your original post so I had to make assumptions about what you meant. I assumed you were talking about the physical layer.

cluthz Wrote:Thanks for taking the time to explain this very detailed in parts. I'm no native english speaker, so if I'm sounding rude or dismissive I can ensure you that is not my intention.

You weren't. You were far more polite than most of the native speakers and non-native speakers on this forum.