GM200/Titan X announced with slow DP

[RolloTheGreat]

Hey, sorry if I expected such big nVIDIA fans to actually read technical material (White Papers) provided by nVIDIA themselves.

Go on on your crusade Gstanford for an holy grail that does not exist (GM200 cannot go over 1:32). I won't delay you more.

[/quote]

I can't speak for other "Nvidia Fans", but debating and/or searching for clues as to software vs hardware disabled FP functionality is about as enticing as say, "Watching paint dry".

I'm a gamer. If Nvidia figures out how to weld a couple transistors to empty beer cans and make them run the games at the resolution/detail level I desire faster than their competition, that's a sale.

This bizarre tradition of "I have read Company Xs white papers, and the presence of Part Y in amount Z should translate to a 13.84% improvement in shader based AA methods!" leaves me cold.

A couple things are true:

1. I make no money memorizing or analyzing white papers
2. Whether disabled by hardware or software, it's out of my hands so I consider it a moot point. I'm not going to start bridging cut traces or running bathtub drivers in a quest for 5% more fps.

Rage on, Picao, but don't expect me to waste my time with it.

[BjorgenFjords]

Picao. I'm not saying you are right or you're wrong. I am saying you could be either, or there is some kind of middle of the road truth. What you do not understand (maybe you do but continue on) is that you do not have all the data necessary to reach a conclusion about this. Neither does Gstanford. You "think" you do, but you don't. The two charts you have shown of the block diagrams of GK104 and GK110 are interesting. Tell me, what was the DP ratio of GK104?

I know you said you aren't posting anymore, but I hope you do.

And lastly, the "emotional" comment of mine was based ENTIRELY on the way you come off in your posts. You show anger. Aggression. Frustration. <---- emo things.

Anyway, if you can change that and just stick to data without throwing a hissy fit, that would be cool.

[Picao84]

Hey, sorry if I expected such big nVIDIA fans to actually read technical material (White Papers) provided by nVIDIA themselves.

Go on on your crusade Gstanford for an holy grail that does not exist (GM200 cannot go over 1:32). I won't delay you more.

I can't speak for other "Nvidia Fans", but debating and/or searching for clues as to software vs hardware disabled FP functionality is about as enticing as say, "Watching paint dry".

I'm a gamer. If Nvidia figures out how to weld a couple transistors to empty beer cans and make them run the games at the resolution/detail level I desire faster than their competition, that's a sale.

This bizarre tradition of "I have read Company Xs white papers, and the presence of Part Y in amount Z should translate to a 13.84% improvement in shader based AA methods!" leaves me cold.

A couple things are true:

1. I make no money memorizing or analyzing white papers
2. Whether disabled by hardware or software, it's out of my hands so I consider it a moot point. I'm not going to start bridging cut traces or running bathtub drivers in a quest for 5% more fps.

Rage on, Picao, but don't expect me to waste my time with it.
[/quote]

As far as I remember, this conversation was with Gstandford, not with you. Our conversation, which is over for me, was on another thread altogether. If you do not find this conversation enticing, you have a very good solution: do not read/comment on it.

---

Picao. I'm not saying you are right or you're wrong. I am saying you could be either, or there is some kind of middle of the road truth. What you do not understand (maybe you do but continue on) is that you do not have all the data necessary to reach a conclusion about this. Neither does Gstanford. You "think" you do, but you don't. The two charts you have shown of the block diagrams of GK104 and GK110 are interesting. Tell me, what was the DP ratio of GK104?

The DP ratio of GK104 was/is 1:24.

From Anandtech:
"The other change coming from GF114 is the mysterious block #15, the CUDA FP64 block. In order to conserve die space while still offering FP64 capabilities on GF114, NVIDIA only made one of the three CUDA core blocks FP64 capable. In turn that block of CUDA cores could execute FP64 instructions at a rate of ¼ FP32 performance, which gave the SM a total FP64 throughput rate of 1/12th FP32. In GK104 none of the regular CUDA core blocks are FP64 capable; in its place we have what we’re calling the CUDA FP64 block. The CUDA FP64 block contains 8 special CUDA cores that are not part of the general CUDA core count and are not in any of NVIDIA’s diagrams. These CUDA cores can only do and are only used for FP64 math."

I hope that satisfies your [possible] curiosity about the absence of DP units on the GK104 SMX diagram.

Further, like it was said above, each SMX has a block of 8 FP64. Lets do the math, shall we?
8*8 [SMX] = 64.
How many FP32 CUDA cores exist in total on GK104? 1536.
How much is 1536:64?
1536/64 = 24
So this gives it a ratio of 1:24. Coincidence?

On GM200:

GM200 is 601mm2 of graphics, and this is what makes it remarkable. There are no special compute features here that only Tesla and Quadro users will tap into (save perhaps ECC), rather it really is GM204 with 50% more GPU. This means we’re looking at the same SMMs as on GM204, featuring 128 FP32 CUDA cores per SMM, a 512Kbit register file, and just 4 FP64 ALUs per SMM, leading to a puny native FP64 rate of just 1/32.

How did Anandtech conclude that? Lets to Maths again.
4*24 [SMX] = 96
How many FP32 CUDA cores exist in total on GM200? 3072.
How much is 3072:96?
3072/96 = 32
So this gives it a ratio of 1:32. Coincidence?

I would have no problem in someone presenting data to backup their stance on the matter, but so far all I've heard is along the lines of "announced FP64 might be the same as the announced ROP count of GTX970" etc. The two situations are not even comparable. GTX970 is, by definition, a cutdown version of a chip we know how it looks like when fully enabled (GTX980). GM200 was presented by nVIDIA as fully enabled on GTX Titan. Saying they may be lying through their teeth without info to back up that claim is not what I call an educated argument. Much less when that same person tells me that the existence of "1, 4 or 4000 FP64 units on it" is irrelevant. I do not believe in magic or spontaneous generation, sorry.

Concerning the "emotional issues", I will only say that you should look at some other people posts as well - but so far I have not seen you telling that to anyone else. In fact this thread started getting "emotional" right after Gstanford said this: "You are spouting bullshit here, Picao84!" I did not see you quote him and telling him he was being emotional. I'm not the only one with this kind of writing here, neither is it my usual personal style, as you can see in other forums (feel free, my nick is always the same - Picao84). But somehow, I'm the only one targeted as being "emo"? Of course it ends up feeling like double standards and pack behavior. But perhaps its a consequence of myself being relatively new here. You are so used to the way Rollo and Gstanford express themselves that you do not notice anymore. Or maybe some moderation is indeed needed, other than just trying to defend the forum and incentive people out. You might say that moderation is not needed, but then again, why quote me and say I was being emotional? If its a free for all, let it be.

[RolloTheGreat]

Rage!

[Picao84]

Rage!

No Darth Vader, you cannot pull me again to the Dark Side!!!

[BjorgenFjords]

But GK104 doesn't have any DP units. At least according to the block diagram like GK110 does. How is it that it's able to complete DP tasks if there isn't any hardware to do it?

[Picao84]

But GK104 doesn't have any DP units. At least according to the block diagram like GK110 does. How is it that it's able to complete DP tasks if there isn't any hardware to do it?

I already answered that with a quote on what nVIDIA told Anandtech. So unless you have better source on info than nVIDIA themselfs or actually say that they are lying I don't know what's your point.

And now my question to you. If GK104 could be more powerful at DP than it already is under GeForce, why didn't nVIDIA promote it as such in Tesla? After all its a smaller chip, cheaper to manufacture. But they promote it as a Single Precision workhorse instead as Tesla K10.

[BjorgenFjords]

But GK104 doesn't have any DP units. At least according to the block diagram like GK110 does. How is it that it's able to complete DP tasks if there isn't any hardware to do it?

I already answered that with a quote on what nVIDIA told Anandtech. So unless you have better source on info than nVIDIA themselfs or actually say that they are lying I don't know what's your point.

And now my question to you. If GK104 could be more powerful at DP than it already is under GeForce, why didn't nVIDIA promote it as such in Tesla? After all its a smaller chip, cheaper to manufacture. But they promote it as a Single Precision workhorse instead as Tesla K10.

Humor me. How is it that it's able to complete DP tasks if there isn't any hardware to do it?

And to answer your question (not with another question), I don't presume to know why Nvidia does what it does. Sometimes a company may do things that to you and I, does not make any sense, but to them and their plans might be perfect for them. So don't ask me why or why not Nvidia or AMD or Intel does what they do. They have their reasons even if you think they might be stupid reasons.

[Picao84]

But GK104 doesn't have any DP units. At least according to the block diagram like GK110 does. How is it that it's able to complete DP tasks if there isn't any hardware to do it?

I already answered that with a quote on what nVIDIA told Anandtech. So unless you have better source on info than nVIDIA themselfs or actually say that they are lying I don't know what's your point.

And now my question to you. If GK104 could be more powerful at DP than it already is under GeForce, why didn't nVIDIA promote it as such in Tesla? After all its a smaller chip, cheaper to manufacture. But they promote it as a Single Precision workhorse instead as Tesla K10.

Humor me. How is it that it's able to complete DP tasks if there isn't any hardware to do it?

And to answer your question (not with another question), I don't presume to know why Nvidia does what it does. Sometimes a company may do things that to you and I, does not make any sense, but to them and their plans might be perfect for them. So don't ask me why or why not Nvidia or AMD or Intel does what they do. They have their reasons even if you think they might be stupid reasons.

Did you even read the Anandtech quote? It has DP units but they are not shown on the diagram and are in much lower in number than in GK110. GK104 has 8 per SMX, while GK110 has 64. Did I ever imply that GK104 does not have DP units? All my calculations on this thread shown its DP units. Along with the diagrams, nVIDIA provides an explanation. Or are you going to hang on to the diagrams alone to make a point, ignoring the remaining information?

Or are you confusing me with Gstanford? Because he is the one saying that doesn't matter how many DP units the chip has. My whole point is that less DP units = less DP performance. GM200 has only 4 DP units per SMM, must like GM204, therefore it's DP performance is low by design. GK210 succeeded GK110 for DP compute (the chip is not even available as GeForce/Quadro).

[BjorgenFjords]

But GK104 doesn't have any DP units. At least according to the block diagram like GK110 does. How is it that it's able to complete DP tasks if there isn't any hardware to do it?

I already answered that with a quote on what nVIDIA told Anandtech. So unless you have better source on info than nVIDIA themselfs or actually say that they are lying I don't know what's your point.

And now my question to you. If GK104 could be more powerful at DP than it already is under GeForce, why didn't nVIDIA promote it as such in Tesla? After all its a smaller chip, cheaper to manufacture. But they promote it as a Single Precision workhorse instead as Tesla K10.

Humor me. How is it that it's able to complete DP tasks if there isn't any hardware to do it?

And to answer your question (not with another question), I don't presume to know why Nvidia does what it does. Sometimes a company may do things that to you and I, does not make any sense, but to them and their plans might be perfect for them. So don't ask me why or why not Nvidia or AMD or Intel does what they do. They have their reasons even if you think they might be stupid reasons.

Did you even read the Anandtech quote? It has DP units but they are not shown on the diagram and are in much lower in number than in GK110. GK104 has 8 per SMX, while GK110 has 64. Did I ever imply that GK104 does not have DP units? All my calculations on this thread shown its DP units. Along with the diagrams, nVIDIA provides an explanation. Or are you going to hang on to the diagrams alone to make a point, ignoring the remaining information?

Or are you confusing me with Gstanford? Because he is the one saying that doesn't matter how many DP units the chip has. My whole point is that less DP units = less DP performance. GM200 has only 4 DP units per SMM, must like GM204, therefore it's DP performance is low by design. GK210 succeeded GK110 for DP compute (the chip is not even available as GeForce/Quadro).

There are no DP units on GK104 according to the block graphic you linked.

[Picao84]

Feel free to read and compare:

GK104:
http://www.anandtech.com/show/5699/nvidi...0-review/2

GK110:
http://www.anandtech.com/show/6760/nvidi...tan-part-1

You are deliberately ignoring information. With that attitude we will not get anywhere.

[BjorgenFjords]

Picao, I've asked you a question twice already and you refuse to directly answer it. You're afraid of the point I have to make otherwise you wouldn't be so "careful" answering questions with questions. So, you'll never admit to possibly being wrong about GM200. I can see that. It takes intelligence to be open about things, rather than a closed, my way or the highway mentality you've exibited thus far. Not another peep from me until you answer my question DIRECTLY. No beating around the bush BS, no questions to answer my question.
Cheers.

[RolloTheGreat]

Picao, I've asked you a question twice already and you refuse to directly answer it. You're afraid of the point I have to make otherwise you wouldn't be so "careful" answering questions with questions. So, you'll never admit to possibly being wrong about GM200. I can see that. It takes intelligence to be open about things, rather than a closed, my way or the highway mentality you've exibited thus far. Not another peep from me until you answer my question DIRECTLY. No beating around the bush BS, no questions to answer my question.
Cheers.

Picao is probably a bargain basement ambulance chaser, looking for some cheesy class action suit.

"Your honor in this interview with Ryan Smith, the defendants represented to the buying public that the Titan X is a FULLY ENABLED Titan level GPU, and yet their employee Keysplayr admits the FP64 performance has been disabled on AlienBabelTech! My clients take FP64 VERY seriously and are entitled to damages!"

[ocre]

The DP ratio of GK104 was/is 1:24.

From Anandtech:
"The other change coming from GF114 is the mysterious block #15, the CUDA FP64 block. In order to conserve die space while still offering FP64 capabilities on GF114, NVIDIA only made one of the three CUDA core blocks FP64 capable. In turn that block of CUDA cores could execute FP64 instructions at a rate of ¼ FP32 performance, which gave the SM a total FP64 throughput rate of 1/12th FP32. In GK104 none of the regular CUDA core blocks are FP64 capable; in its place we have what we’re calling the CUDA FP64 block. The CUDA FP64 block contains 8 special CUDA cores that are not part of the general CUDA core count and are not in any of NVIDIA’s diagrams. These CUDA cores can only do and are only used for FP64 math."

I hope that satisfies your [possible] curiosity about the absence of DP units on the GK104 SMX diagram.

Further, like it was said above, each SMX has a block of 8 FP64. Lets do the math, shall we?
8*8 [SMX] = 64.
How many FP32 CUDA cores exist in total on GK104? 1536.
How much is 1536:64?
1536/64 = 24
So this gives it a ratio of 1:24. Coincidence?

On GM200:


How did Anandtech conclude that? Lets to Maths again.
4*24 [SMX] = 96
How many FP32 CUDA cores exist in total on GM200? 3072.
How much is 3072:96?
3072/96 = 32
So this gives it a ratio of 1:32. Coincidence?

I would have no problem in someone presenting data to backup their stance on the matter, but so far all I've heard is along the lines of "announced FP64 might be the same as the announced ROP count of GTX970" etc. The two situations are not even comparable. GTX970 is, by definition, a cutdown version of a chip we know how it looks like when fully enabled (GTX980). GM200 was presented by nVIDIA as fully enabled on GTX Titan. Saying they may be lying through their teeth without info to back up that claim is not what I call an educated argument. Much less when that same person tells me that the existence of  "1, 4 or 4000 FP64 units on it" is irrelevant. I do not believe in magic or spontaneous generation, sorry.

Concerning the "emotional issues", I will only say that you should look at some other people posts as well -  but so far I have not seen you telling that to anyone else. In fact this thread started getting "emotional" right after Gstanford said this: "You are spouting bullshit here, Picao84!" I did not see you quote him and telling him he was being emotional. I'm not the only one with this kind of writing here, neither is it my usual personal style, as you can see in other forums (feel free, my nick is always the same - Picao84). But somehow, I'm the only one targeted as being "emo"? Of course it ends up feeling like double standards and pack behavior. But perhaps its a consequence of myself being relatively new here. You are so used to the way Rollo and Gstanford express themselves that you do not notice anymore. Or maybe some moderation is indeed needed, other than just trying to defend the forum and incentive people out. You might say that moderation is not needed, but then again, why quote me and say I was being emotional? If its a free for all, let it be.

I completely believe that the GM200 will be restricted to 1/32 DP. I dont think we will see cards with a higher rate. But i know that is my guess. I have been known to be wrong. But for the record, I am with you on the GM200 being limited to 1/32. I feel like it is.

But, i dont understand how the gk110 can switch between a 1/23 DP rate to a 1/3 DP rate just by clicking a check in the the control panel. How the heck is this happening? I really really dont know.

If there is special HW that allows the clock/timing/ or some multiplier - that would be really great to know. Or maybe it is schedulers? I have no idea. But if we can find whatever magic that allows the gk110 to switch between states, that would be awesome.

[ocre]

When you can't compete, obstruct.

Amd tool A,
"People who buy nvidia are so dumb. They just giving away them their money for something they want, a powerful GPU today. They are so brainwashed, spending their money now when we all know that one day cheaper cards will come out."

AMD tool B
"Yeah, they are mindless sheeple. Buying nvidia is just dumb and irresponsible. The value will plummet overnight just like it did on the original Titan and Titan black."

AMD tool A,
"Yep. That doesn't happen when you buy AMD gpus. They are the most awesome of awesome. The holy of holy. AMD gpus never loose value, ever."

76% of the market
"Huh?"

http://www.fool.com/investing/general/20...an-bl.aspx

Seriously,