GM200/Titan X announced with slow DP

[Picao84]

Like I speculated in the old forum, Big Maxwell, unlike Big Kepler/Fermi, is not strong in DP: only 0,2 TFlops.
One more for the record, following my speculation that Maxwell would be most if not all done in 28nm, several months before the release of GM204. As usual, I was faced at the time with some fierce opposition to this idea, never mind how many arguments and facts I thrown at it.

On a side note, the keynote is very very interesting. Lots of information about Pascal and performance estimates: 10x faster than Maxwell!!

http://www.ustream.tv/channel/gpu-techno...rence-2015

[SickBeast]

There's no way that Pascal will be 10x faster outside of a few custom made niche apps that support some type of new functionality. That's pure hype IMO.

The Titan X looks really good though. nVidia just keeps pumping out new technology while AMD has been stagnant for some time now. I hope that Captain Jack can bring prices down.

[BjorgenFjords]

Was it 10X faster, per watt? Was there a qualifier like that? Just got home from work and now I can watch it in it's entirety.

[Picao84]

Was it 10X faster, per watt? Was there a qualifier like that? Just got home from work and now I can watch it in it's entirety.

It was, as Jen put it, "CEO Math". He just took some specs and play around with the number. In reality it should have 2x perf/watt of Maxwell + 4x FP16 performance + 3x memory bandwidth of Maxwell. Tegra X1 was the first step towards stronger FP16 performance. Its built for deep neural networks that use FP16 more than FP32/FP64. Anandtech is speculating as well that GK110 might have been the last time we saw a big nVIDIA chip sharing Compute (FP64) and Graphic duties. From now on nVIDIA might make FP64 processors exclusively for Compute.

[SickBeast]

The Titan X has value as a pure gaming card. I'm sure it can come pretty close to keeping up with a pair of GTX 970s which wouldn't cost much less.

[SteelCrysis]

2 970s at $330 = $660. Titan X = $999. Tom Cruise Laughing.gif

[railven]

2 970s at $330 = $660.  Titan X = $999.  Tom Cruise Laughing.gif

Yeah, what I was thinking. Sick must be some kind of balla if >$300 difference is "wouldn't cost much less."



I'm loving everything I'm seeing for Titan X except the price. Is peaking my interest more towards GTX 980 Ti, whatever that may be, but frankly at this point SLI 970/CFX 290's looks like might be the better deal until we see what 390X can/can't do.

In the last two year the chips went:
GF110 >> GK110 >> GM200
$550 >> $700 >> $1000

It makes me sad :( My poor wallet.

[RolloTheGreat]

The Titan X has value as a pure gaming card.  I'm sure it can come pretty close to keeping up with a pair of GTX 970s which wouldn't cost much less.

Huh?! Two MSI GTX970s are $628 on newegg right now. (and you can get two of most models for under $700)

[SickBeast]

I would have to see benchmarks but a 30 percent premium sounds worth it to me so long as the performance is similar. SLI is not ideal. 12gb of vram is a nice added bonus as well. By the way, the GTX 970s are much more expensive in Canada which is why I made that comment. I did not realize that they were so cheap in the US.

[Picao84]

Like I speculated in the old forum, Big Maxwell, unlike Big Kepler/Fermi, is not strong in DP: only 0,2 TFlops.
One more for the record, following my speculation that Maxwell would be most if not all done in 28nm, several months before the release of GM204. As usual, I was faced at the time with some fierce opposition to this idea, never mind how many arguments and facts I thrown at it.

On a side note, the keynote is very very interesting. Lots of information about Pascal and performance estimates: 10x faster than Maxwell!!

http://www.ustream.tv/channel/gpu-techno...rence-2015

You are spouting bullshit here, Picao84!

You cannot infer GM200's true FP64 capabilities from Titan X, since they are artificially limited to 1/32 of the GPU's true capabilities in Titan X, just like regular consumer geforces.

It does make you wonder precisely what the point of Titan is nowadays though, they used to limit Titan to 1/3 the true fp64 rate and tout the card as something a small/home developer would like since you could use it like a slower Quadro/Tesla or a geforce. The 1/32 fp64 rate makes it no better than any other garden variety geforce for CUDA though.

I guess the true purpose of Titan nowadays is to separate fools from their money (just like Intels x99).

*sigh* No I'm not. Maximum of GM200 is really 1:32 DP. Before lashing out at me go see the reviews with technical information on them like Anandtech. Why do you think they are speculating GK110 was the last time we have seen a big nVIDIA GPU share gaming and DP capabilities? Even nVIDIA themselves said that for compute there is GK210. I wont quote now cause I'm on the mobile phone, but if you don't believe me go check for yourself. Plus nVIDIA is now betting heavily in deep neutral networks, which don't need FP64 and are happy with FP16 and FP32. Those are in fact big news for gaming as well. Less space occupied by DP that can be b used for gaming functions. GM200 is already a sign off that shift.

---

2 970s at $330 = $660.  Titan X = $999.  Tom Cruise Laughing.gif

Yeah, what I was thinking. Sick must be some kind of balla if >$300 difference is "wouldn't cost much less."



I'm loving everything I'm seeing for Titan X except the price. Is peaking my interest more towards GTX 980 Ti, whatever that may be, but frankly at this point SLI 970/CFX 290's looks like might be the better deal until we see what 390X can/can't do.

In the last two year the chips went:
GF110 >> GK110 >> GM200
$550 >> $700 >> $1000

It makes me sad :( My poor wallet.

Yeah.. But nothing changes if nVIDIA has a monopoly right? Right? *rolls eyes* Happening in front of our eyes but still some deny it. Incredible.

[Picao84]

Like I speculated in the old forum, Big Maxwell, unlike Big Kepler/Fermi, is not strong in DP: only 0,2 TFlops.
One more for the record, following my speculation that Maxwell would be most if not all done in 28nm, several months before the release of GM204. As usual, I was faced at the time with some fierce opposition to this idea, never mind how many arguments and facts I thrown at it.

On a side note, the keynote is very very interesting. Lots of information about Pascal and performance estimates: 10x faster than Maxwell!!

http://www.ustream.tv/channel/gpu-techno...rence-2015

You are spouting bullshit here, Picao84!

You cannot infer GM200's true FP64 capabilities from Titan X, since they are artificially limited to 1/32 of the GPU's true capabilities in Titan X, just like regular consumer geforces.

It does make you wonder precisely what the point of Titan is nowadays though, they used to limit Titan to 1/3 the true fp64 rate and tout the card as something a small/home developer would like since you could use it like a slower Quadro/Tesla or a geforce.  The 1/32 fp64 rate makes it no better than any other garden variety geforce for CUDA though.

I guess the true purpose of Titan nowadays is to separate fools from their money (just like Intels x99).

*sigh* No I'm not. Before lashing out at me go see the reviews with technical information on them like Anandtech. Even nVIDIA themselves said that for compute there is GK210. I wont quote now cause I'm on the mobile phone, but if you don't believe me go check for yourself.

Yes, I have read the technical information.  The 1/32 fp64 is an artificial limitation, imposed upon titan so as not to encroach upon Quadro's and Tesla's based upon the same chip that will have a 1/3 fp64 rate.

No it's not!!! Quote then where you read that bullshit!

Here is from Anandtech:

"For GM200 NVIDIA’s path of choice has been to divorce graphics from high performance FP64 compute. Big Kepler was a graphics powerhouse in its own right, but it also spent quite a bit of die area on FP64 CUDA cores and some other compute-centric functionality. This allowed NVIDIA to use a single GPU across the entire spectrum – GeForce, Quadro, and Tesla – but it also meant that GK110 was a bit jack-of-all-trades. Consequently when faced with another round of 28nm chips and intent on spending their Maxwell power savings on more graphics resources (ala GM204), NVIDIA built a big graphics GPU. Big Maxwell is not the successor to Big Kepler, but rather it’s a really (really) big version of GM204.

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. As a result, all of that space in GK110 occupied by FP64 ALUs and other compute hardware – and NVIDIA won’t reveal quite how much space that was – has been reinvested in FP32 ALUs and other graphics-centric hardware."

Note that it says GM200 and not Titan

From Tech Report:

"The GM200—and by extension, the Titan X—differs from past full-sized Nvidia GPUs in one key respect, though. This chip is made almost purely for gaming and graphics; its support for double-precision floating-point math is severely limited. Double-precision calculations happen at only 1/32nd of the single-precision rate.

For gamers, that's welcome news. I'm happy to see Nvidia committing the resources to build a big chip whose primary mission in life is graphics, and this choice means the GM200 can pack in more resources to help pump out the eye candy."

How about you learn to read then?

[Picao84]

http://www.anandtech.com/show/9059/the-n...n-x-review

Big Maxwell is not the successor to Big Kepler, but rather it’s a really (really) big version of GM204.

Can you read properly???? Why do you think he is saying that GM200 is not a successor of GK110? Because it has slow DP! He is not saying Titan vs Titan X. He is saying Big Kepler vs Big Maxwell! The chip is itself limited to 1:32DP, Titan X or otherwise!

[Picao84]

Sure can.  Look at the FP64 column, first left - 1/32 FP32

And where does that negate what I said? Slow DP! It's not an artificial limitation!

[Picao84]

Of course it is an artificial limitation. They can clock FP64 at ratios other than 1/32 if they want. Pro cards are usually clocked at 1/3 fp32 rate.

They do this with scheduler timing.

It's not about the clock! It's about the number of FP64 units!

In the same review:

"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. As a result, all of that space in GK110 occupied by FP64 ALUs and other compute hardware – and NVIDIA won’t reveal quite how much space that was – has been reinvested in FP32 ALUs and other graphics-centric hardware."

I must be debating with a wall here.. Or you are just trolling.

[Picao84]

You can increase FP64 performance purely by timing.  How else do you imagine that a Quadro/Tesla based on GK104 differentiates itself from a geforce based on GK104.  The number of FP64 units per SM stays the same, the scheduler timing is what varies!

Are you going to tell us geforce GK104's are just as powerful as Quadro/Tesla GK104's when using GK104?!

There are no strong DP GK104 Teslas! GK104 was used in strong SP Teslas, not DP!!!! DP was always the realm of GK110 (and now GK210). The FP64 units dictate the DP power, not the the timing! You have no idea what you are talking about!

---

Here: http://www.nvidia.com/content/tesla/pdf/...asheet.pdf

A K10 GK104 Tesla has the wonderful performance of 0.19 TFlops FP64. *rolls eyes*

[Picao84]

I could use big kepler as another example or any fermi.  The principle is exactly the same for all of them.

Per SM, geforce, Quadro, Telsa all have the same units available to them.  The difference with fp64 between geforce and the professional products is the scheduling ratio.

The first Titan also had its fp64 scheduling ratio configured like a pro product.  nvidia has now reneged on that and given Titan owners consumer geforce fp64 scheduling ratios instead.

Look this is very simple:

GK104 - few DP units. GeForce has 100% of Tesla DP performance.
GK110 - many DP units. GeForce non-Titan has less DP performance because most DP units are disabled/fused. Nothing to do with clocking.

GM204 - few DP units. GeForce has 100% of Tesla DP performance (so far no Tesla has shown up with GM204 though).
GM200 - few DP units. NO DP TESLA!!!! DP is exclusively the realm of GK210 on Tesla K80.

You know I'm right and you are just arguing for the same of it. You just hate to be wrong. After all every single review is saying the same thing: GM200 is a gaming oriented GPU that traded FP64 performance for FP32. That has nothing to do with clocking but everything to do with the GPU structure. Unless you want to tell me that you know better than nVIDIA's own reviewer guide.

[Picao84]

From Anandtech:

"For that reason I wouldn’t be too surprised if we a Tesla card using it aimed at FP32 users such the oil & gas industry – something NVIDIA has done once before with the Tesla K10 – but you won’t be seeing GM200 in the successor to Tesla K40."

Tesla K40 is GK110 based.

---

What?! GTX780 and 780 Ti have just as many fp64 units in one SM as a quadro/tesla does. Yes, 780/Ti get SMX's fused off but that isn't where the fp64 differential comes from, it arises primarily from the 1/32 consumer vs 1/3 pro scheduling ratio!

Nooooooo! That's precisely where the 1/32 ratio comes from! The FP64 ALU units are responsible for... DP performance! If you have less of them enabled, you have less DP performance!

What do you think 1:32 ratio means??? It means that there is 1 FP64 unit per 32 FP32 units! A ratio of 1:3 means there is 1 FP64 unit per 3 FP32 units! That's what it means. Nothing about scheduling!

[Picao84]

Who knows what future Quadro's and Tesla's might or might not be based off of?

Not me, certainly.  And not you either unless you work for nvidia in that area.

Having said that my gut feeling has always been that nvidia would skip over Big Maxwell to concentrate on Pascal and Volta.

Big Volta in particular is required for the upcoming supercomputer nvidia is helping to build.

The first paragraph was not mine but Anandtech. Forgot to identify.

So typical of you to try and change the subject when you know are wrong, instead of admiting it :p Every time I have a discussion with you I picture you as an angry Donald Duck, don't know why :p

[RolloTheGreat]

Yeah.. But nothing changes if nVIDIA has a monopoly right? Right? *rolls eyes* Happening in front of our eyes but still some deny it. Incredible.

Ten years ago the 7800GTX 512 launched with a $700 MSRP, and people were paying $800-$900 for them at etailers.

Nvidia has been selling expensive gaming cards to those who will buy them for a long time. They also sell cards everyone can afford, and those who wish those levels of performance at reasonable prices just need to wait a year.

Nvidia can't exist selling $1000 parts, you mistake them wisely taking the early adopter toll from those who will pay it for monopoly pricing trends.

[Picao84]

I'm not wrong.  and I'm certainly no angry Donald Duck, but I'm beginning to suspect you may be Daffy Duck based on how you think FP64 differs between consumer and pro products of the same GPU die.....

No I'm not wrong. FP64 units do matter. They are disabled pretty much in the same way a GTX970 has 384 FP32 units disabled to differentiate it from a GTX980. A GeForce GTX780Ti has only 90 (2880/32) out of the 960 (2880/3) FP64 units enabled on a Tesla K40.

The calculation of theoretical performance of NVIDIA GPUs goes along this: Units * Clock * 2 (nVIDIA GPUs can process two instructions per clock).

Example:
GK110 Tesla K40
Single precision - 2880 * 0.745 * 2 = ~ 4.29 TF
Double precision - 960 * 0,745 * 2 = ~ 1,43 TF

[Picao84]

You are cracked.

You cannot fuse off parts in an SMX.  You can disable an entire SMX or more.  780 Ti has all 960 DP units (64 per SMX) that pro GK110 dies do.

The difference lies in the rate at which the GPU's scheduler dispatches work to those units.

I know
You just took my bait and contradicted yourself. If the units are not there, just like they aren't in GM200 (that's a fact!!), how can the scheduler dispatch to them? Therefore FP64 units are ESSENTIAL for Double Precision, unlike you have been saying all along. GM200 only has 4 FP64 units per SM and that's a fact! The scheduler cannot get FP64 units out of thin air, therefore the 1/32 rate of GM200 is not an artificial limitation. It's physical, the chip cannot do more! End of story!

[BjorgenFjords]

It's a safe bet that since no actual FACTUAL proof whether Big Maxwell is artificially gimped that neither of you truly know. Just conjecture and each of you can't take the other being right. So.... Stop.
Nvidia can probably do anything they want with the same hardware across GeForce, Quadro, Tesla. All via software.
So guys, it's really not that important. Titan X is a gamers card. Leave it at that.

[RolloTheGreat]

It's a safe bet that since no actual FACTUAL proof whether Big Maxwell is artificially gimped that neither of you truly know. Just conjecture and each of you can't take the other being right. So.... Stop.
Nvidia can probably do anything they want with the same hardware across GeForce, Quadro, Tesla. All via software.
So guys, it's really not that important. Titan X is a gamers card. Leave it at that.

What's interesting to me about all the "DAAAAAAMMMNNNNN Eeeen-VE-DE-AA for milking profits from honest gaming fans!" threads is:

What would you prefer?

An AMD like situation where the company is always one step ahead of the Grim Reaper, low on R&D funds, low on staff, long time between releases and corners cut like the leftover HSFs on the 290 reference models?

Or making big money and putting out nice parts pretty regularly?

I'm sure some internet pundits "who could surely do it better" will say "But tRollo, somewhere between the profiteering and the flirting with bankruptcy is a happy medium where they make a fair profit and we get cards!".

To such naivete' I can only respond:

Sorry, Nvidia exists to take as much of your money as they can by any legal means. They are not your special friends, they are not concerned with anything but your wallet. AMD would do the same thing if they could.

[Picao84]

It's a safe bet that since no actual FACTUAL proof whether Big Maxwell is artificially gimped that neither of you truly know. Just conjecture and each of you can't take the other being right. So.... Stop.
Nvidia can probably do anything they want with the same hardware across GeForce, Quadro, Tesla. All via software.
So guys, it's really not that important. Titan X is a gamers card. Leave it at that.

There is enough FACTUAL proof: the reviewers guide followed by every single reviewer. They all explain the GM200 chip, just like they explained every single GPU before, whether a full chip or not. Why would nVIDIA lie about the quantity of FP64 units? Because they very explicitly said that there are 4 units per SM, down from the 64 on GK110. What gstandford can't seem to get around is that less FP64 units = reduced double precision performance. It's basic math:

4 units * 24 SM = 96
3072 cores / 96 = 32

This where nVIDIA gets their 1:32 rate from. Therefore GM200 cannot go over that rate. It's physically limited. Unless nVIDIA deliberately lied, which I don't believe. Giving that die size increased only 9% and they had to include things like double the ROP count of GK110, double the cache of GK110, among others, I tend to believe that trade off was needed.

---

It's a safe bet that since no actual FACTUAL proof whether Big Maxwell is artificially gimped that neither of you truly know. Just conjecture and each of you can't take the other being right. So.... Stop.
Nvidia can probably do anything they want with the same hardware across GeForce, Quadro, Tesla. All via software.
So guys, it's really not that important. Titan X is a gamers card. Leave it at that.

What's interesting to me about all the "DAAAAAAMMMNNNNN Eeeen-VE-DE-AA for milking profits from honest gaming fans!" threads is:

What would you prefer?

An AMD like situation where the company is always one step ahead of the Grim Reaper, low on R&D funds, low on staff, long time between releases and corners cut like the leftover HSFs on the 290 reference models?

Or making big money and putting out nice parts pretty regularly?

I'm sure some internet pundits "who could surely do it better" will say "But tRollo, somewhere between the profiteering and the flirting with bankruptcy is a happy medium where they make a fair profit and we get cards!".

To such naivete' I can only respond:

Sorry, Nvidia exists to take as much of your money as they can by any legal means. They are not your special friends, they are not concerned with anything but your wallet. AMD would do the same thing if they could.

You cannot say that and still say that nVIDIA having a monopoly is indifferent. It's all fine to follow nVIDIA logic as a company. But you cannot say that the status quo is not changed at all with a monopoly. Its already changing. You prefer like that? That's your stance. What you cannot say, for the third time, is that the price equilibrium that exists/existed does not change.

[RolloTheGreat]

My last 3 nvidia cards (460/670/970) have been some of the most reasonably priced I've ever purchased from nvidia, especially given the performance on offer.

Same with my 2600K and 4790K's

The value on offer with these blows periods where nvidia and ATi or AMD and intel competed clean out of the water.

Don't be a fool Greg! intel is lulling you into a false sense of security since they became the only CPUs worth buying in 2006. They are waiting for the right moment to spring their monopoly trap!

It is only when there are only worthless used AMD CPUs for sale that the other shoe will drop and prices will go through the roof!

As long as there are a couple worthless new AMD CPUs to be had intels monopoly has held at bay by the fierce competition they represent.

BTW- I hear they're putting a new heatsink on the 2011 motherboard chipset they're still using. Nothing has happened in motherboard tech in the last 4 years- you'll be fine.

[SickBeast]

It will be interesting to see what nVidia does without AMD. If Captain Jack bombs (which it very much could), nVidia is going to find itself in a situation where they really don't have any competition in terms of midrange and high end cards. We will see what they do with their pricing on their future cards.

Their prices aren't bad right now but they are more expensive for similar performance compared to AMD. Most people are willing to pay the premium because the nVidia cards run cooler with far lower PSU requirements and less fan noise.

[ocre]

Like I speculated in the old forum, Big Maxwell, unlike Big Kepler/Fermi, is not strong in DP: only 0,2 TFlops.
One more for the record, following my speculation that Maxwell would be most if not all done in 28nm, several months before the release of GM204. As usual, I was faced at the time with some fierce opposition to this idea, never mind how many arguments and facts I thrown at it.

On a side note, the keynote is very very interesting. Lots of information about Pascal and performance estimates: 10x faster than Maxwell!!

http://www.ustream.tv/channel/gpu-techno...rence-2015

Pico84, i was one of the people that thought that the gm200 would have a higher DP ratio that the gm204. I even suggested maybe the gk210 was a result of a special contract that needed the specific changes that the chip offered over the gk110.

I really did expect a higher DP ratio. I wasnt so sure the DP ratio would be as high as kepler's but i really felt that the full gm200 would still best the gk110 in DP, even if by small tad. I fully expected it to better the gk110 in DP performance, at the very least.

So my expectations were wrong.

YOU WIN!!!!

If that makes you feel better, I have no problems admitting my imagined version of the gm200 was wrong on my personal DP performance expectations of the gm200. But..........

You know you have been steadily changing your claims from the beginning. You first came off saying the GM200 would be poor in compute. That was your claim and i challenged that claim. The gm200 is actually the most powerful compute graphics cards there are. You used this term "compute" in just the same fashion as so many of the pro AMD defenders in the time of Tahiti after the gk104 launched. Just so you know, compute is not DP performance. The truth is, when people typically talk about a GPUs "compute" performance, they refer to benchmarks and task that do absolutely no double precision calculations at all. These OpenCL benchmarks, task, bitcoin mining, etc have nothing to do with DP performance and they were always pointed to when someone wanted to show how strong Tahiti was in compute. Well, turns out that the GM200 is an amassing compute card.

Maybe you just didnt know what you were talking about or maybe you are changing the goal post. Either way, there is an issue here and you were not entirely correct in your original claims.

Anyway, i have another question for you. I dont care to get into your debate with gstanford, but there is another problem among your claims. The gk110 titan did not have this static DP performance that was directly tied to the HW layout. It doesnt work like you claim at all. Actually, out of the box you only get something like 1/24th the DP rate. But you say this is hardwired to the HW layout and this is where your claims get busted. The titan will not run at 1/3rd DP rate unless you toggle it on through the control panel. It is a software switch in implemented through driver. With it off, the Gk110 has the same DP rate as the Gk104. Toggle it on and it goes to a 1/3 DP rate.

http://hexus.net/tech/reviews/graphics/5...-overview/

This fact is further substantiated by the knowledge that, just like Tesla K20X, TITAN can run double-precision compute at 1/3rd of single-precision speeds, leading to over 1TFLOPS DP throughput. However, being a gamer's card at heart, TITAN's DP rate is set to 1/24th of SP, just like GTX 680, as no games use double-precision calculations. The full 1/3rd ratio can be set via the control panel, yet doing so forces the GPU's clocks down. And no gamer wants that, right?

So i guess the GK110 has a shape shifting SMMs or gstan knows something about what he is talking about.

I am not saying that the gm200 is capable of a higher DP rate, i actually accepted the fact that i was wrong. I am just saying that you may not know as much as you think about the what nvidia disables DP. If we look at the gk110, it is obvious that Gstan is correct in suggesting the clock rate goes up. I am actually glad he was willing to add some valuable information to the discussion.

No it is up to you to figure out why Nvidia has limited the Gm200 to the same DP rate as the gm204. Cause you both cant be right

[Picao84]

You know you have been steadily changing your claims from the beginning.  You first came off saying the GM200 would be poor in compute.  That was your claim and i challenged that claim.  The gm200 is actually the most powerful compute graphics cards there are.  You used this term "compute" in just the same fashion as so many of the pro AMD defenders in the time of Tahiti after the gk104 launched.  Just so you know, compute is not DP performance.  The truth is, when people typically talk about a GPUs "compute" performance, they refer to benchmarks and task that do absolutely no double precision calculations at all.  These OpenCL benchmarks, task, bitcoin mining, etc have nothing to do with DP performance and they were always pointed to when someone wanted to show how strong Tahiti was in compute.  Well, turns out that the GM200 is an amassing compute card.

That is totally not true. I never changed any goal posts. My point was always about DP.
See the following post on the old forums, still available ( I tried to copy the link but did not find it):

I was looking for an old thread and then I remembered that many were lost weeks ago.

Anyway, looks like my expectations are not so different from what may happen. No Maxwell HPC chip at all, so GM200 should indeed be a graphics oriented chip...

http://wccftech.com/nvidia-planning-dit ... ving-2017/

The reason behind not deploying Maxwell in Tesla Accelerators is said to be the lack of FP64 Floating Point Units and additional double precision hardware. And the reason behind not including DP FPUs in Maxwell might have to do with the extremely efficient design that NVIDIA was aiming for. This however means that NVIDIA’s upcoming Maxwell core, the GM200 which is the flagship core of the series might just remain the GeForce only offering unlike Kepler which was aimed for the HPC market first with the Titan supercomputer and launched a year later after the arrival of the initial Kepler cores as the GeForce GTX Titan.

Since the DP FP64 FPU hardware blocks will be removed from the top-tier cards that are rumored to arrive next year, they will include several more FP32 FPUs to make use of the remaining core space and that means better performance for the GeForce parts since games have little to do with Double precision performance.

GM200 did not have that much space to grow in 28nm, given that GK110 was already huge. Could they make it faster at DP? Probably yes. But how much? Would it be worth it for say 20% more performance? HPC market is very different from graphics market. HPC is sold in massive quantities, not individual cards. No company is going to massively upgrade their system for 20% extra performance. Jumps in performance must be much higher than that. They kinda did that sort of jump with GK210 double chip cards. They choose to optimise an existing chip, GK110, to reduce power consumption heavily in order to offer a bigger jump in performance than a single GM200 could offer. It was a very intelligent move on their part, IMO.

I even said it was a very intelligent move on their part, so take back your claims about AMD bla bla bla.

It is not a GEFORCE only part, but also a QUADRO part
Just not a high end TESLA part.

More information:
http://wccftech.com/nvidia-gm200-based- ... s-compute/

Hmm, 6,07 TFLops single precision versus GK110 and GM204 5,2 TFlops. Not much of a big jump, is it? Then again, for graphics, nVIDIA architectures were always very efficient per TFlop. However, this makes me question things a little bit.. With weak DP performance, I would have expected a larger jump in single precision. Then again, the possible extra TMUs and ROPS (2x the ROPS of GK110!), plus double the cache of GK110 (I've seen a speculation from somewhere else for 3MB, up from GM204 2MB) might have taken quite a bit of the space left vacant by absent FP64 units.

I even said that the rumored performance, 6,07 TFlops was lower than what I was expecting. So no, you cannot say that I expected less Single Precison performance, or that I don't know the difference between them!!

Maybe you just didnt know what you were talking about or maybe you are changing the goal post.  Either way, there is an issue here and you were not entirely correct in your original claims.

I was correct in my original claims. Looks like you just have poor memory!

Anyway, i have another question for you.  I dont care to get into your debate with gstanford, but there is another problem among your claims.  The gk110 titan did not have this static DP performance that was directly tied to the HW layout.  It doesnt work like you claim at all.  Actually, out of the box you only get something like 1/24th the DP rate.  But you say this is hardwired to the HW layout and this is where your claims get busted.  The titan will not run at 1/3rd DP rate unless you toggle it on through the control panel.  It is a software switch in implemented through driver.  With it off, the Gk110 has the same DP rate as the Gk104.  Toggle it on and it goes to a 1/3 DP rate.

http://hexus.net/tech/reviews/graphics/5...-overview/

This fact is further substantiated by the knowledge that, just like Tesla K20X, TITAN can run double-precision compute at 1/3rd of single-precision speeds, leading to over 1TFLOPS DP throughput. However, being a gamer's card at heart, TITAN's DP rate is set to 1/24th of SP, just like GTX 680, as no games use double-precision calculations. The full 1/3rd ratio can be set via the control panel, yet doing so forces the GPU's clocks down. And no gamer wants that, right?

   

So i guess the GK110 has a shape shifting SMMs or gstan knows something about what he is talking about.  

I am not saying that the gm200 is capable of a higher DP rate, i actually accepted the fact that i was wrong.  I am just saying that you may not know as much as you think about the what nvidia disables DP.  If we look at the gk110, it is obvious that Gstan is correct in suggesting the clock rate goes up.   I am actually glad he was willing to add some valuable information to the discussion.  

No it is up to you to figure out why Nvidia has limited the Gm200 to the same DP rate as the gm204. Cause you both cant be right

My debate with Gstandford IS about the physical limitation on GM200, something that he continues to say its not true! I never said that GK110 on GeForce GTX780Ti was not artificially limited or hardwired (when I said they were fused, I was leading him into a trap.. from the moment he says it is software based he had to understand that if GM200 does not have more than 4 FP64 per SM... he had to understand GM200 is not artificially limited - guess he didn't connect the dots). Of course it is! I even showed the difference between them in active cores (960 vs 90). What you don't seem to understand is that the DP rate is not some internal clock. Its related to the amount of enabled FP64 units (software or otherwise its irrelevant, if the units are not there, there is nothing you can do). Gstanford seems to believe that nVIDIA can, if they want, enable DP rates higher than 1:32 on GM200. They cant! This conversation is all around that. Its silly!

Honestly, I'm out of this forum. There is a pack mentality here that if someone does not share the same vision, they are hunted and accused, directly or indirectly, of being pro AMD. This forum lacks rationality and free thinking. It is even worse than other places you like to bad name. At least on Beyond 3D its possible to have educated discussions and really learn with other people. I have no idea what BTR is, and I certainly started to dislike and distrust Appopin, but you guys... are not any better. Just look at yourselves in the mirror. Your self image might not be who you think you are. Over and out.

[Picao84]

doesn't matter if GM220 has 1, 4 or 4000 FP64 units on it!

They can still be ran at different ratios of full speed and if you look at the anandtech chart I posted before you will see they are running at a 1/32 rate.

If you are too thick to understand that, then I'm afraid there is nothing more that I can tell you....

Do you know what native means??? You are the thick one here that cannot see the obvious!

image upload without registration

http://www.anandtech.com/show/9059/the-n...x-review/2

[Picao84]

Perhaps "native FP64" means what "4 gb", "64 rops", "256 bit" mean on GTX 970.....

So your escape pod is "perhaps" and "nvidia is lying"? Stop the bullshit. You are a perfect picture of what is wrong in this forum. Everyone wants to be right at all costs and twists reality when faced with the truth. In your case its ridiculous really, since you brought this upon yourself by lashing out at me before having all the facts straight. Like I said, you didn't read any reviews properly before opening your mouth, otherwise it would have been shut.

[RolloTheGreat]

It will be interesting to see what nVidia does without AMD.  If Captain Jack bombs (which it very much could), nVidia is going to find itself in a situation where they really don't have any competition in terms of midrange and high end cards.  We will see what they do with their pricing on their future cards.

Their prices aren't bad right now but they are more expensive for similar performance compared to AMD.  Most people are willing to pay the premium because the nVidia cards run cooler with far lower PSU requirements and less fan noise.

My guess is it will be a lot like it is now.

Nvidia "could" say the new price point for mid range cards is $500, but:
A. They risk pushing people out of computer gaming altogether and onto comparable goods. (consoles, tablets, phones, handhelds, intel based laptops)
B. They would sell far less of all products in the stack if they adjust prices upward by any significant amount.

It's true that some people would just pay the higher prices, but they are offset by those who would give up high end computer gaming altogether.

Nvidia can only survive selling people graphics cards every year or two at a price they want to pay. They've already lost a huge chunk of their business to consoles, and intel laptops. I guarantee you they don't want to lose more just to cash in temporarily. Consoles already have a slew of huge advantages: $400-$500 acquisition cost for whole unit and it's the only hardware you buy for 8 years, ease of use, huge installed user base that likely includes most of your gaming friends, more comfortable to sit in a recliner in your living room than in front of a pc, console game exclusives, etc..

If Nvidia attempts to pillage us, they will see very quickly how many think playing the same games at 720p or 1080p on a 50"+ screen is probably "good enough" in the face rising PC gaming costs. They would basically be ending their own business.

[dmcowen674]

Honestly, I'm out of this forum. There is a pack mentality here that if someone does not share the same vision, they are hunted and accused, directly or indirectly, of being pro AMD. This forum lacks rationality and free thinking. It is even worse than other places you like to bad name. At least on Beyond 3D its possible to have educated discussions and really learn with other people. I have no idea what BTR is, and I certainly started to dislike and distrust Appopin, but you guys... are not any better. Just look at yourselves in the mirror. Your self image might not be who you think you are. Over and out.

There is no pack. The Duopoly is equally discussed here therefore there is certainly plenty of "rationality and free thinking".

You know the saying "If you can't take the heat, get out of the kitchen".

[BjorgenFjords]

What's funny is, free thinking people is exactly what he cannot seem to tolerate. I think he wants to be surrounded by people who agree with him, or as he puts it "able to have educated discussions". Because if you don't agree with him, you must be uneducated.

Sheesh man.

[RolloTheGreat]

What's funny is, free thinking people is exactly what he cannot seem to tolerate. I think he wants to be surrounded by people who agree with him, or as he puts it "able to have educated discussions". Because if you don't agree with him, you must be uneducated.

Sheesh man.

I pretty much put that guy on "ignore" when he threw Nvidia Focus Group conspiracy theory out at us. You're the only member now aren't you?

Oh noes, don't unleash the PR disinformation juggernaut of all one of you upon us Keys! How can we defend ourselves against such a corporate influence?!?!

[RolloTheGreat]

In light of you being the Nvidia Focus Person and a person can't be a Group by definition, I say Nvidia has no transparency and is competing unfairly with AMD!

Without the treachery of the Nvidia Focus Person and intel's buyer loyalty rebates, AMD would probably have us all living at Star Trek-like levels of technology.

[SickBeast]

Picao seemed to have too much of a temper. Rage issues. It's too bad because he was pretty knowledgeable.

[RolloTheGreat]

They already do - their shills at least.  Star Trek is fantasy and the shills live in fantasy land..........

I think I know why Captain jack is taking so long.

They're only allowed to use AMD CPUs designing it.

ATi Engineer:"Boss I would have had Captain Jack done 6 months ago if I had a Haswell! What's with this "Vishera" crap?!"

Manager: "Shut up and be glad you have a job! AMD was nice enough to take us in when NVIDIA had pushed us to the edge of bankruptcy, you'll use their CPUs and like them!"

ATi Engineer: "But boss! My freaking PHONE is faster than this "top of the line" desktop! I remember when they launched this chipset when I was in High School!"

Manager: "I know....just pray they don't sell us to Kia to make dash displays for subcompact economy cars for the rest of our lives......"

[ocre]

That is totally not true. I never changed any goal posts. My point was always about DP.
See the following post on the old forums, still available ( I tried to copy the link but did not find it):



I even said it was a very intelligent move on their part, so take back your claims about AMD bla bla bla.




I even said that the rumored performance, 6,07 TFlops was lower than what I was expecting. So no, you cannot say that I expected less Single Precison performance, or that I don't know the difference between them!!




I was correct in my original claims. Looks like you just have poor memory!


My debate with Gstandford IS about the physical limitation on GM200, something that he continues to say its not true! I never said that GK110 on GeForce GTX780Ti was not artificially limited or hardwired (when I said they were fused, I was leading him into a trap.. from the moment he says it is software based he had to understand that if GM200 does not have more than 4 FP64 per SM... he had to understand GM200 is not artificially limited - guess he didn't connect the dots). Of course it is! I even showed the difference between them in active cores (960 vs 90). What you don't seem to understand is that the DP rate is not some internal clock. Its related to the amount of enabled FP64 units (software or otherwise its irrelevant, if the units are not there, there is nothing you can do). Gstanford seems to believe that nVIDIA can, if they want, enable DP rates higher than 1:32 on GM200. They cant! This conversation is all around that. Its silly!

Honestly, I'm out of this forum. There is a pack mentality here that if someone does not share the same vision, they are hunted and accused, directly or indirectly, of being pro AMD. This forum lacks rationality and free thinking. It is even worse than other places you like to bad name. At least on Beyond 3D its possible to have educated discussions and really learn with other people. I have no idea what BTR is, and I certainly started to dislike and distrust Appopin, but you guys... are not any better. Just look at yourselves in the mirror. Your self image might not be who you think you are. Over and out.

There is no reason for you to get so over dramatic. I openly stated I was wrong in my expectations for the DP performance of the GM200. I clearly have no problem admitting that much. But when i challenged you to explain how the GK110 can have a software switch from gk104 like DP ratio to the 1/3rd ratio of the original titan, you get all crazy.

I have no idea what your talking about when you say your being hunted for being pro AMD. Do W-H-A-T-?
That is when i realized maybe you are really are completely irrational.

I could be wrong though. And i wont have an issue admitting it.........I actually hope that i am. I thought you were an alright guy that has strong beliefs and will fight for them. But really, your getting crazy and worked up with a lot of stuff that is 100% all in your head.

For starters, I am not on gstanfords side on this at all. The DP performance of the GM200 is limited. I conceded, I see that.

But that doesnt mean you were correct about the clock rate. Because obviously, the gk110 had some way of switching between the 1/24 rate and the 1/3 rate. This doesnt mean that the gm200 can increase its DP clock/rato. But you originally claimed that DP was a result of the hardware physical FP64 units and there was no such clock or timing. There obviously is a way with the gk110.

My personal feeling is that the special ability to increase the clock/timing on kepler was special and only the gk110 had this ability. That is couldnt be done with the gk104. But that is my guess. But with that, i imagine that the gm200 is missing that ability as well. Perhaps this takes up much needed space in the die.

So it is not that i ever challenged the DP rate of the gm200. I believe it is and always will be 1/32. It is just that you must also see that there is some special way the gk110 can go from 1/24 to 1/3. That perhaps there is something you can learn more about this. Perhaps your assertion that gstanford "doesnt know what he is talking about" is way to harsh. Because obviously there is more to this. And i think that is why you got so angry.

Have no idea why you would go off like that. There was only one person claiming GM200 can do more that 1/32 DP and that was Gstan. No one else here is challenging that. Every body else believes you, they believe that the gm200 is as DP limited as the gm204.
I just wanted you to think further into it. To realize that there is some way of changing the rate with a gk110 and it is no a hardwired/fixed result.

I also dont think you made up that the 780ti had fused off DP in each SM just to lead him into a trap. I think you mistakingly thought that is how it worked and then you try to twist your statement into something else. It is alright, i had no idea how nvidia cut back DP ratios with kepler. There are probably very few people that do know how it works. It is okay if your not 100% right, all of the time. dont get mad.

But i do believe you on the GM200. I believe it is limited to 1/32 and cannot be changed like the gk110 can. But it is obvious there is some sort of clocking/timing/ or multipliers at work in big keplers DP capabilities.

[BjorgenFjords]

Just to be clear, I don't agree or disagree with Picao or Gstanford in this matter. All I can say without any doubt it, I don't know if the 1/32 DP ratio is a hardware limited factor of GM200 or a software limited factor. It could very well be that Nvidia decided, for their own in house reasons, to restrict DP to 1/32 on "GeForce" Titan this time around. Perhaps the Quadro, and/or Tesla versions of GM200 (if there are any) will, through software, allow a significantly stronger DP ratio.
Picao maintains (without any real proof) that GM200 is hardware limited. It does not physically have the hardware for anything lower than 1/32. Gstanford maintains (without any proof) that GM200, like GK110, could through software, attain a lower ratio.
I don't agree with either of them, or disagree with either of them, because I can't.

[Picao84]

[quote='Picao84' pid='1038' dateline='1426757061']

I have no idea what your talking about when you say your being hunted for being pro AMD.  Do W-H-A-T-?

I'm not posting more here, but this point needs clarified. I am NOT pro AMD or an AMD fan, nor have ever been. I'm neutral with a slight bias towards nVIDIA (did you play Bioware/D&D games? Same logic as "neutral good"). Here is the the point that here some cannot understand. Here, if you post something that is not along the lines of "nVIDIA is good/AMD is bad", you are seen as pro AMD. Rubbish. Yes, I did imply Rollo and Keysplayr are overly pro nVIDIA, which is not exactly unknown to anyone. And I only said that after Keysplayr got in the way of my debate with Rollo, not to add something to the conversation, but to, out of the blue, say that my post was overly emotional. His first post on this thread. That is very significant and set the new tone. Something that he continued doing as the thread progressed. That is my rational for the "pack" comment, since most of his posts targeted me instead of adding to the conversation. Had he not done that, things were probably different. Clarification over.

---

You can't design a modern FP64 unit to be as slow as 1/32 natively.

Its not the unit itself that it is slow. But I've already shown you the math behind the "1/whatever" rate. You ignored it. I actually learned it from real CUDA developers that know how the chip works and try to extract every bit of performance from the chips. But hey, what do they know right?

nvidia doesn't fundamentally alter the ALU's in a SMX between big and small die either.  The Scheduler may get more capable between big and small dies (as was the case with GK110 vs GK104).

Basically you design the SMX once, and that is the basis for everything from the smallest Tegra die to the monster Tesla die.

The only thing that differs is the number of SMX's on the chip and how complex the scheduler is.

*sigh* Wrong, wrong, wrong.

GK104 SMX:


GK110 SMX:


Impressive how many here go "la,la,la,la" and say there is not factual evidence when nVIDIA is so open about it. Instead they rely on "ifs" and "buts" and hypothesis that nVIDIA is not saying the full truth, bla, bla, bla. They rely on "wishful thinking" and their own flawed interpretation of how things work, and accuse the person who presents facts of "spewing bullshit". 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.