Titan X is Single Precision

BTW, another reason for NV leaving DP computing might be the coming Intel Knights chip and cards. With 3 Teraflop DP performance and a version of the chip able operate as a standalone computer it might be a game changer in this area which NV can not match. If the chip would be available for making ultra high-performance workstations there would be total shakeup at the high-end (imagine Xeon on overdose of steroids) though I tend to think Intel will offer it only to exotic markets by limiting sales and/or by setting exorbitant unit price.

The new Knights Chip can indeed be a game changer, and is hardware wise. It will all depend on how much Intel wants to push these rather than any competition IMO.
 
I say yes to your logic. (keeping what I have for a few years) I still can't keep myself from perusing the forums.

Still rocking my GTX 780.

Kid
 
http://www.pcworld.com/article/2898...-1000-titan-x-the-most-advanced-gpu-ever.html

"Huang, however, also said the card would hit roughly 200 GFLOPs in double-precision floating point performance. That puts it on par with the company’s discontinued—and two generations old GeForce GTX 780 Ti—in double-precision performance. That’s a break from previous Titans, which have all ruled the roost in double-precision performance. The original GeForce Titan, for example, could hit 1,500 GFLOPS.

"It is a gaming card," says analyst Jon Peddie of Jon Peddie Research. "Floating point performance is limited."

Huang seemed unconcerned with the double precision performance, however, and showed off numbers that the Titan X could perform. Used for neural net analysis, Huang said a 16-core Intel Xeon would take 43 days to process the data. The original GeForce Titan would take six days and its replacement, last year’s GeForce Titan Black, would take five days."

That's pretty sad, considering my GTX480 could hit 168 GFLOPS DP, and that GPU is from 2010.
I really think that NVIDIA's BS is starting to show in ways that even their PR department can no longer put a positive spin on it.

I might be going with AMD next time around; inferior product or not, at least I know what I am getting.
NVIDIA is kind of turning into Apple circa 2006. :(
 
That's pretty sad, considering my GTX480 could hit 168 GFLOPS DP, and that GPU is from 2010.
I really think that NVIDIA's BS is starting to show in ways that even their PR department can no longer put a positive spin on it.

I might be going with AMD next time around; inferior product or not, at least I know what I am getting.
NVIDIA is kind of turning into Apple circa 2006. :(


? DP performance? Are you a programmer, mathematician or scientist that needs DP?

The new Knights Chip can indeed be a game changer, and is hardware wise. It will all depend on how much Intel wants to push these rather than any competition IMO.


Back to the topic Knight's landing is going up agaist Pascal, so its probably going to be over matched by alot.
 
That's pretty sad, considering my GTX480 could hit 168 GFLOPS DP, and that GPU is from 2010.
I really think that NVIDIA's BS is starting to show in ways that even their PR department can no longer put a positive spin on it.

I might be going with AMD next time around; inferior product or not, at least I know what I am getting.
NVIDIA is kind of turning into Apple circa 2006. :(

I'm still not sure why anyone is surprised or even disappointed with the new Titan X having lacking Double Precision performance. It's a gaming card, nothing more. NVidia is compartmentalizing their product lines: high Double Precision performance is available with the K-series Quadro and all Tesla products. They realized their screw-up with the Kepler-based Titans having outstanding DP performance and robbing themselves of even higher profit Quadro and Tesla sales.
 
I'm still not sure why anyone is surprised or even disappointed with the new Titan X having lacking Double Precision performance. It's a gaming card, nothing more. NVidia is compartmentalizing their product lines: high Double Precision performance is available with the K-series Quadro and all Tesla products. They realized their screw-up with the Kepler-based Titans having outstanding DP performance and robbing themselves of even higher profit Quadro and Tesla sales.

Sadly, this is true, and you are a right, it isn't a surprise at all.
But this is what also made the original Titan stand out amongst the crowd.
 
I'm still not sure why anyone is surprised or even disappointed with the new Titan X having lacking Double Precision performance. It's a gaming card, nothing more. NVidia is compartmentalizing their product lines: high Double Precision performance is available with the K-series Quadro and all Tesla products. They realized their screw-up with the Kepler-based Titans having outstanding DP performance and robbing themselves of even higher profit Quadro and Tesla sales.

I was perplexed by this as well. I just assumed it was because people were jealous.

The only negative is it may hurt resale value but I don't think most of the people who were bitching had that in mind.
 
I never used the DP on my og Titan anyway. Wasted die space for gaming it seems.
 
The reason they took it out of Maxwell is simple:

When you have the same old-ass process node you've been using for years, and it's up to the chip architects to find both (1) spare TDP for more performance and (2) more efficient layout to free up more TDP, and to cram more processing units into the same area.

Maxwell managed to improve performance by 50% for the same Power by:

(!) adding larger L2 cache and better losssless compression, allowing them increase performance by 50% using the exact same memory interface, saving the die space and power of adding more external I/O.

(2) some removal of lesser-used parts of the shading blocks, meaning nearly the same performance with smaller die and less power.

(3) specifically for big Maxwell: they dropped the massive amount of dedicated double-precision FP units. IF they had kept these, the die space would have been too large to be manufacturerable, so they just didn't bother.
 
Last edited:
That's pretty sad, considering my GTX480 could hit 168 GFLOPS DP, and that GPU is from 2010. I really think that NVIDIA's BS is starting to show in ways that even their PR department can no longer put a positive spin on it.

Nonsense. Dropping DP is brilliant step NV since there are new huge markets opening which need exclusively SP or even limited SP and lots of memory. Brainy people are talking about future in which e.g. driving will be prohibited due to self-driving cars being safer or machines seeing things better than people. GPU architecture fits fine to these apps. Apps requiring DP will be microsize comparing to this type of SP apps

Back to the topic Knight's landing is going up agaist Pascal, so its probably going to be over matched by alot.

This is not very likely. NV might be dropping DP altogether due to the niche market size. Intel is more interested since Knights Landing can also work as stand alone processor, this will be then step into general architecture with very many cores. SP is interested in specialized architecture.

I'm still not sure why anyone is surprised or even disappointed with the new Titan X having lacking Double Precision performance. It's a gaming card, nothing more. NVidia is compartmentalizing their product lines: high Double Precision performance is available with the K-series Quadro and all Tesla products. They realized their screw-up with the Kepler-based Titans having outstanding DP performance and robbing themselves of even higher profit Quadro and Tesla sales.

No,no, not like this. Look at the new Quadro M, it is professional Titan X. Titan X is absolutely not only a gaming card. Its SP and huge VRAM makes it ideal for the new machine learning applications. NV is completely deemphasizing DP, the boss said clearly that Pascal emphasis is on reduced SP. This is because the new applications require it and volumes will be huge.

The reason they took it out of Maxwell is simple:
When you have the same old-ass process node you've been using for years, and it's up to the chip architects to find both (1) spare TDP for more performance and (2) more efficient layout to free up more TDP, and to cram more processing units into the same area. Maxwell managed to improve performance by 50% for the same Power by:
(!) adding larger L2 cache and better losssless compression, allowing them increase performance by 50% using the exact same memory interface, saving the die space and power of adding more external I/O.
(2) some removal of lesser-used parts of the shading blocks, meaning nearly the same performance with smaller die and less power.
(3) specifically for big Maxwell: they dropped the massive amount of dedicated double-precision FP units. IF they had kept these, the die space would have been too large to be manufacturerable, so they just didn't bother.

First and foremost, if there would be prospective growing market for DP they would keep it. But there is suddenly emerging potential huge new market for GPU architecture with exclusive S , and gaming does not need SP too. This is why NV is concentrating exclusively on SP with Maxwell and with Pascal the headlines will be on reduced SP (e.g. 16-bit SP) and not DP. This is due to the breakthroughs which make possible machines better than people in dealing with visual tasks like driving, DP is not needed there, massive reduced SP is best.
 
If you require DP Titan X is evidently not for you. NV said it very clearly, if you need highest performance DP you buy Titan Z. Why NV did this you can see from their presentation at GPU Tech conference, they noticed there are tons of new applications requiring only SP (vision, machine learning, self-driving cars, robots). Even more, some of these applications require floating point less precise than SP and this will be supported in the Pascal architecture. NV thus relegated DP to a niche since market is very small comparing to the SP market.

Nonsense. Dropping DP is brilliant step NV since there are new huge markets opening which need exclusively SP or even limited SP and lots of memory. Brainy people are talking about future in which e.g. driving will be prohibited due to self-driving cars being safer or machines seeing things better than people. GPU architecture fits fine to these apps. Apps requiring DP will be microsize comparing to this type of SP apps.

You do realize that the DP capabilities in GeForce/Titan (Z/X) are only disabled/gimped in hardware, not completely removed, right?
NVIDIA did this so that they wouldn't cannibalize their Tesla and Quadro GPU markets, which have DP fully enabled.

This was a business move, not because "learning applications" can take advantage of SP; from a technical standpoint, that doesn't even make sense.
If anything, one would want DP to ensure further accuracy, especially for everything you listed.

You can save your statements for NVIDIA's ad campaign and PR discussions. ;)
 
Last edited:
You do realize that the DP capabilities in GeForce/Titan (Z/X) are only disabled/gimped in hardware, not completely removed, right?
NVIDIA did this so that they wouldn't cannibalize their Tesla and Quadro GPU markets, which have DP fully enabled.
This was a business move, not because "learning applications" can take advantage of SP; from a technical standpoint, that doesn't even make sense.If anything, one would want DP to ensure further accuracy, especially for everything you listed.
You can save your statements for NVIDIA's ad campaign and PR discussions. ;)

Do you have credible information DP is only disabled in Maxwell? According to sources Maxwell is designed from the bottom for SP only.

About the new machine learning applications you are clearly ignorant. They stem from new theories, keywords: Deep Neural Networks, Convolutional Neural Networks. DP makes no sense there, what makes even bigger sense there than SP is reduced SP. Reduced SP is now key for Pascal. Pascal will not put any emphasis on DP, DP is small niche market comparing to these new applications. NV boss was showing NVidia board for Tesla cards. Imagine high-end NVidia chips in each card, this the new market size and many more.
 
Rumor says that Pascal Mixed precision is supposed to use the compute units in either DoublePrecision, SinglePrecision or half precision mode. Just like you can do 16bit fp with 32bit units and effectively double the performance.

So it would only have half the performance in DP instead of 1/32th. Without additional 64bit units.
 
You do realize that the DP capabilities in GeForce/Titan (Z/X) are only disabled/gimped in hardware, not completely removed, right?
NVIDIA did this so that they wouldn't cannibalize their Tesla and Quadro GPUs, which have DP fully enabled.

This was a business move, not because "learning applications" can take advantage of SP; from a technical standpoint, that doesn't even make sense.
If anything, one would want DP to ensure further accuracy, especially for everything you listed.

You can save your statements for NVIDIA's ad campaign and PR discussions. ;)

No, the Titan X, unlike the Kepler Titan, has no fast dedicated DP compute path:

http://wccftech.com/nvidia-quadro-m6000-official/

Unlike previous Quadro cards such as the K6000 itself, the NVIDIA Quadro M6000 features only 1/32 the DP (Double Precision) performance due to lack of FP64 hardware available on the Maxwell core architecture that was speculated a while ago. Hence, the M6000 features just 0.2 TFlops of FP64 compute performance like the GeForce GTX Titan X which FP32 performance is rated at 7.0 TFlops.

The Kepler Quadro was the largest they could make the die and still have high-performance DP units. They added too many improvements (including double the ROPs) to the Big Maxwell die to afford the space for fast dedicated DP hardware.

Let me make this easier for you to understand: making die sizes over 500mm2 is very difficult and expensive, because the number of dies is less, and the defect percentage per-die is higher. The larger it gets., the harder things become.

http://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/2

GM200 (Big Maxwell): 601mm2, 1/32 DP rate
GK110 (Big Kepler): 551mm2, 1/3 DP rate
GF110 (Big Fermi): 520mm2, 1/2 DP rate

They introduced fast DP support with Fermi: 512 shaders and 256 DP units, clocked at 2x core speed.

The move from 40nm Fermi to 28nm Kepler allowed them to keep those beefy DP units because of the die shrink, improvements to the layout of the shaders, and the fact that the two chips used roughly the same bulky or game/-oriented resources (i.e 48 ROPS, 384-bit memory interface, and approximately the same number of texture units).

But they still had to cut the ratio to 1/3, since the number of shader units skyrocketed from 512 to 2880 units!

And now they had to make the same magic performance improvement happen without a die shrink (or suffer the pain of being beaten badly by Fiji), so the first thing to go was the DP. If people need the DP, they can still buy GK110. And this being Nvidia, they've already found different markets to try and sell the Quadro version of the card to. This includes being packaged with IRAY, which only uses 32-bit DP, and of course the machine learning thingy :D
 
Last edited:
No,no, not like this. Look at the new Quadro M, it is professional Titan X. Titan X is absolutely not only a gaming card. Its SP and huge VRAM makes it ideal for the new machine learning applications. NV is completely deemphasizing DP, the boss said clearly that Pascal emphasis is on reduced SP. This is because the new applications require it and volumes will be huge.

Double. Precision.

...is what I said. Higher DP performance is available in the K-series Quadro and Tesla products. Not in the M-series. If an enterprise segment customer wants only SP, then just about any GPU from any segment would fit the bill as long as whatever they choose meets their SP needs. If they want massive SP capability and a fuckton of VRAM, then they would do just fine with any upper-end Quadro, Tesla, or even a GeForce GTX Titan X (the fastest GAMING video card currently available as of today on planet Earth).
 
Do you have credible information DP is only disabled in Maxwell? According to sources Maxwell is designed from the bottom for SP only.

I stand corrected. :)
I was working off of Kepler info and did not realize that Maxwell had changed the GPU design and structure so much.

If you do find any interesting info, please share it as I really would like to read through it.
 
Last edited:
You do realize that the DP capabilities in GeForce/Titan (Z/X) are only disabled/gimped in hardware, not completely removed, right?
NVIDIA did this so that they wouldn't cannibalize their Tesla and Quadro GPU markets, which have DP fully enabled.

This was a business move, not because "learning applications" can take advantage of SP; from a technical standpoint, that doesn't even make sense.
If anything, one would want DP to ensure further accuracy, especially for everything you listed.

You can save your statements for NVIDIA's ad campaign and PR discussions. ;)

That used to be the case for Kepler cards but not maxwell. While Geforce Kepler cards had most of its 64bit cuda cores disabled. Titan X only has 1/32 of DP cuda cores and fully enabled btw.
 
The funny thing is that however they decide to brand the cards; workstation or gaming, they have their work cut out for them since AMD's R295 x2 beat every single nVidia card in almost every single category of performance. In some cases if not most nVidia was being run in SLI and still fell short and I am a disappointed nVidia fan boy.
 
Take all of those $ values and subtract $100-$200 after they get #rekt by AMD in June.
I'm wetting my whistle for a sub-$400 GTX 980. Let's make it happen.

Last time AMD was months late to the market their cards were underwhelming. Although they did lower the price of the GTX 260s when they finally arrived ~ 8 months later. :p
 
That used to be the case for Kepler cards but not maxwell. While Geforce Kepler cards had most of its 64bit cuda cores disabled. Titan X only has 1/32 of DP cuda cores and fully enabled btw.

Nice, thanks for the info, I stand corrected. :)

I totally missed your post.
Thank you for the documentation and the examples, I am reading through them now!
 
Last edited:
The easiest way show this is that the Quadro based on Titan X also lacks DP.
 
The funny thing is that however they decide to brand the cards; workstation or gaming, they have their work cut out for them since AMD's R295 x2 beat every single nVidia card in almost every single category of performance. In some cases if not most nVidia was being run in SLI and still fell short and I am a disappointed nVidia fan boy.

R295x2 is not beating Titan X in any significant way, besides Titan X oveclocked is tough to beat.

The easiest way show this is that the Quadro based on Titan X also lacks DP.

Indeed, this is good indicator. It indicates NV is totally decreasing stake in DP since they bet on new markets related to vision, machine learning, silicon brains, silicon humanoids, where there might be potential for billions of chips. These areas rely on massively parallel architectures a la GPU and they do not need much precision, even full SP is not necessary For the market potential think about NV chips in the vision system of each car. Comparing to this the DP is niche market they can leave for others.
 
Indeed, this is good indicator. It indicates NV is totally decreasing stake in DP since they bet on new markets related to vision, machine learning, silicon brains, silicon humanoids, where there might be potential for billions of chips. These areas rely on massively parallel architectures a la GPU and they do not need much precision, even full SP is not necessary For the market potential think about NV chips in the vision system of each car. Comparing to this the DP is niche market they can leave for others.

Pascal will have DP.
Tegra in vehicles, while having a higher potential market, is already a highly competitive market, ie lower margins, and the vision/self-drive aspect is still quite a way down the road, ie years of R&D before you get any significant revenue.
DP HPC they already have at least a guaranteed half of the market from CUDA and massive guaranteed contracts and/or high margins.

You don't actively remove yourself from a market that is part of the backbone to your success just because you see potential in other markets. That would be like Intel removing itself from the Enterprise/Server marketplace in the early 2000's because they saw the market potential of the mobile/ultra-mobile markets.
 
Pascal will have DP.
Tegra in vehicles, while having a higher potential market, is already a highly competitive market, ie lower margins, and the vision/self-drive aspect is still quite a way down the road, ie years of R&D before you get any significant revenue.
DP HPC they already have at least a guaranteed half of the market from CUDA and massive guaranteed contracts and/or high margins.

You don't actively remove yourself from a market that is part of the backbone to your success just because you see potential in other markets. That would be like Intel removing itself from the Enterprise/Server marketplace in the early 2000's because they saw the market potential of the mobile/ultra-mobile markets.

Exactly. Lord knows what he's talking about. Nvidia will continue to produce DP-heavy products when they have a fresh die shrink to fit more units. Their next core should be on 14nm, and should be DP-heavy. I imagine they will follow-up that product with an SP-heavy refresh version on the same 14nm process.

This is just Nvidia recognizing clearly what the limits of fab processes will be in the future: you can expect 3-4 years between major fabrication improvements. Thus, if you want a product refresh in the between times (i.e. Maxwell), you'll have to do it by improving efficiency, or temporarily removing die-hungry features that don't affect gaming like DP.

Nvidia just knows how to craft a product and sell it no matter what they are burdened with, which is why they continue to increase their profits in a shit market.
 
Last edited:
Indeed, this is good indicator. It indicates NV is totally decreasing stake in DP since they bet on new markets related to vision, machine learning, silicon brains, silicon humanoids, where there might be potential for billions of chips. These areas rely on massively parallel architectures a la GPU and they do not need much precision, even full SP is not necessary For the market potential think about NV chips in the vision system of each car. Comparing to this the DP is niche market they can leave for others.

Knowing what I know now about Maxwell, I can guarantee you there will be a DP-heavy variant for their Tesla GPUs and potentially a Quadro variant.
Again, NVIDIA doesn't want to cannibalize their two separate markets, which makes perfect sense.

But to say they are completely getting away from DP all together is just silly.
 
Back
Top