nVidia 2015 mobile speculation thread

Unlikely GM200 will even be able to fit on a mobile die unless it's soldered and has some custom socket. I wouldn't count on it nor would I want it. And if that's the case, the 990M is going to be a disposable card - never upgradeable. You may as well give your money to a homeless man or throw it in the trash. Depreciation will eat you alive when Pascal hits the market. Nobody will want your expensive junk.

 

yea, im gonna skip 9xx and jump to pascal most likely would be a huuge bump up from a 765m imo. my questions would lie in whats gonna happen with mxm and relavent tech. it seems to me the market has turned a little pessmistic as if late. (cant blame them. )

my hope would be low end and if the 990 is a indicator of super high end would be soldered and then the "normal x60-x80 cards would go mxm"

 

It's certainly hard to justify a 990M at the rumoured price when Pascal is no more than a year off. If it keeps the 980M pricing then sure, it'll be tempting.

I may or may not go for it. All depends on pricing. The lack of anything AMD makes me nervous about the cost of this new GPU though.

Sent from my Nexus 5 using Tapatalk

 

My wishlist for pascal is gtx 1070m or 1080m should be equal to gtx 780m sli maxed oced i had and have mxm form so i can dump my 780m and use it and keep it running cool. By the way within july 2016 can we see mobile pascal??

 

lol. Sounds to me like you have never seen a MXM card and a heatsink inside a gamer notebook before.

GM200 can fit on a MXM card and the heatsinks can take 600mm2 die without any problem.

Here is a 400mm2 980M. You see all that unused room?



Here is a comparison what a GM200 die would look like. No problemo. Maybe now you will understand. Please stop with the "its too big". Its not.

 

398mm to 601mm is quite a leap for one mobile GPU. Looks like it will also require a larger heat sink. Seems like a pointless investment.

They'd be better off just sticking with GM204 and having 2048 cores.

 

I`m done. I just showed you that its possible. literally two posts above...
Atleast try to open up your mind and read what other people are writing.
There is a lot of unused space on a heatsink as well. Just like the GM204 and GM200 comparison shows. If you had seen a 100W heatsink, you would know. You have a 980M dude, take that notebook apart and look at it. There is black tape surrounding a big part of the copper. Unused part of the heatsink that can accomodate a 600mm2 die.

As long as the TDP doesnt exceed 125W or whatever, it doesnt matter if its a GM200 or a GM204. A heavy overclocked 2048 core or a lower clocked GM200. The power draw might be similar anyway, but there could be financial benefits by going one way, or for example stability.

 

Yeah, I agree, in some respects GM200 having that massive surface area will help in transferring the heat that IS produced quickly to the heatsink - especially if they bring down the core frequencies to fit within a sensible TDP - that way the amount of heat that needs to be dissipated per mm squared will be less than the 980M for instance.

(but I still think it'll probably be GM204 instead)

 

Doesn't make sense to change manufacturing for a single mobile GPU. GM200 is gone after 2015. It's a dead end. Even the GTX 990 (desktop variant) will be a dual-GPU based on GM204.* Still don't understand why you're so defensive of this 990M. It's going to be obsolete by June, you said so yourself.



The definition of insanity is...? Just kidding... We all love speculating.

 

So @Cloudfire when is it coming middle to end october? So After 1 month we will start seeing leaked benchmarks then??

 

leaked ETA said week 40 this year, so that would be starting on sept. 28th

Sent from my Nexus 5 using Tapatalk

 

buying this before a 16nm release is kinda like wasting money away especially we'll have problem cooling even if it comes with mxm type. right now we can just OC 980m hella high and still have problem try to keep em cool.

 

Yes thats a good point. Another reason to hope for GM200.

GM204: 0.314W per mm2 die
GM200: 0.207W per mm2 die

I tell you, GM200 is the best scenario we can hope for. Easier to remove heat like you say, but also we get 2500 cores with greater stability because of lower clocks. There will also be much better overclocking potential than a heavy overclocked GM204.

 

Yes, actually that does sound good - especially as you put some numbers to the W/mm2 - it's a big difference! I think I worked out that you based your calculations above on a 125W TDP: 125W/600mm2 = 0.208W/mm2
(Well, about 125W TDP anyway). Actually, thinking about it, I think this would be a more efficient prospect to have more cores at a slightly lower frequency - it's possible that the chip would then be operating at a more efficient level in terms of Calculations per Watt - I'm not sure what frequency is the most efficient for Maxwell? Sure, it's more expensive to produce a GM200, but maybe they'll charge a ridiculous amount for it to cover their costs! And besides, if it is a low frequency it's running at, then it's gonna be a chip to be loved for overclockers if they can tweak the bejesus out of their cooling systems! Right, now I'm 50:50 on whether 990M is gonna be GM204 or GM200.

 

FTFY

right now everyone's saying a waste of money, but once it comes out and blows everyone's mind with its performance numbers, people are still gonna go ahead and buy it its always like that

Sent from my Nexus 5 using Tapatalk

 

But how do you propose we take a 250W part, and have it run at half its TDP (125W) without it running like total ass? (or a 3-legged ass)

Just a reminder that the full GM204 with 2048 cores is a 165W part, so if you want to stuff in MOAR COARZ yet end up with a lower TDP, well that ain't happening unless core and mem clocks are pulled so far back, that you would've been better off designing something around GM204 in the first place (again I present you the example of 480M vs 485M).

 

990M if it exists is GM204. I'm not a fool, you can't possibly convince me it's GM200. Nvidia would sooner commit seppuku than do that.

 

I think it comes down to the answer to the question of 'What frequency voltage does Maxwell run at to maximise Calculations per Watt?'. If the answer is below 1Ghz, then perhaps GM200 does make sense - although expensive to manufacture & buy I would have thought.

 

Would be pointless to clock below 1Ghz. You guys are becoming so desperate for GM200, you're starting to make up stuff to force it to be possible.

I'd love to see GM200 in mobile before Pascal, but I won't count on it. What I want more than anything is NVIDIA to get their heads out of their a**** and stop failing with driver support, making them degrade performance. It's like AMD is providing NVIDIA with driver support.

Where are them Chinese benchmarks at, @Cloudfire? We should see them by now if they plan to launch in October. Get your Chinese on!

 

Again.

GTX 980 Ti: 2880 cores running around 1170MHz
Theoretical desktop chip: 2432 cores running around 1000MHz. GDDR5 downclocked and two memory bus disabled.

That should put TDP down from 250W to around 190W. Easily. For a desktop card.

I`m shocked that you don`t know that mobile graphic cards run at lower voltage either. You have been here for so long man. The whole point of offering similar specs between mobile and desktop while TDP being so different is because they reduce voltage and bin the chips. Which in turn reduce TDP.

Which is why GTX 680 had a TDP of 195W while GTX 880M had a 125W TDP. With similar specs

 

I just persuaded myself to wait another month until the end of August for the DM, now the 990M is coming one month later?

 

Wait until you die, then you won't need to wait anymore.

People gonna be like, "Was waiting for the 990M and now the 1080M is coming out in a few months? "

 

Might as well, been waiting nearly 6 years already!

I suppose I got lucky with the build quality on this hp. Seems worth waiting for DDR4 and maybe this mysterious 990M, if there is a chance whatever I buy next will last until 2020.

 

Dont count on mobile Pascal anytime soon.

Runors has it that GP100 will arrive first, then its HBM2 yields and availability, then its the question if MXM even supports HBM which may mean soldered GPUs. Lots of ifs

 

you mean mobile pascal? ^^

edit: yes you did ^^

Sent from my Nexus 5 using Tapatalk

 

Cloudfire know Chinese? So am I.

 

Well if GM200 can fit in mobile form factor why not GP100?

 

lol yes I'm aware that mobile cards run on lower voltages, but again, how do you half the TDP whilst not completely butchering efficiency and performance? You did base your W/mm² calculation on 125W TDP no? If you're going to talk about a hypothetical cut down GM200 that runs off 190W TDP, sure that's much more believable. But no laptop up to now has ever been designed to handle such a part, so you'd basically have to design a laptop around this particular part.

Also, my point in bringing up the 480M vs 485M example is that when you cut down a big die too far (480M), your efficiency drops off a cliff, and your performance would be worse than if you had started with a smaller chip instead. I'm sure Meaker can back me up on this.

 

I mean isn't that old Clevo with the extreme desktop cpu big enough to practically fit a full desktop card inside of it?

 

yup yup, pretty much any clevo sli machine could take a 250W desktop gpu easily

Sent from my Nexus 5 using Tapatalk

 

BS. Can barely handle 880M as it is.

 

If by handle, you mean, it won't burst into flames. Sure, I'd say that's a fair statement.

Assuming all this TDP talk comes true, you're going to need a 330W PSU for a single GPU laptop with the 990M. Overclocking the 980M is the smartest thing to do. Dropping €1500+ on a 990M is a bit ridiculous, even for a wealthy person. But I'd probably do it, too.

 

I dont think you understand what I`m trying to say.

The hypothetical desktop GM200 GPU based on 2432 cores, lower GDDR5 speed, and two memory bus disabled would have ~190W TDP. Our mobile chip would be far lower due to lower voltage and binning.
Down to 120W maybe.

GTX 880M was a 125W TDP chip. Why wouldnt a GM200 with same or slightly lower work this time? I`ve already covered that GM200 can fit inside a MXM card. Heatsinks can too. PSUs could deliver enough power to GTX 880M. They can too for GM200 then.
There is nothing else that would stop a GM200 from happening. Except what Nvidia ultimately decide.
For gamers it would be much better to have more cores to play with. With more surface area to remove heat.

You mention efficiency. GTX 980 is 8% more efficient than GTX Titan X (3072 cores). GTX 980 is 6% more efficient than GTX 980Ti (2816 cores). They seem to equalize the more you go down in cores. 2432 cores would probably be extremely similar to GTX 980 in efficiency. Which isnt shocking considering it will only have 384 more cores.

 

nothing to discuss there, sli laptop has 2 gpus, 100-125W each, thus can handle 200-250W. period

Sent from my Nexus 5 using Tapatalk

 

Two 125W GPUs and a single 250W monolith are very different things

 

If I should guess, HBM2 and TSV availability.
I feel that Nvidia is launching GP100 instead of GP104 first this time to have enough HBM2 chips to cover the GPUs they will sell.

Earlier we had Samsung, Elpida, Hynix all producing GDDR5. Availability for this chips was a non issue. Its different this time around. Only Hynix manufacture HBM at the moment. AMD had to go to an additional new source for TSV for the HBM. They couldnt get enough. With only 2 products using HBM they had huge issues with delivering enough cards. Now imagine AMD using HBM for 6-7 products and Nvidia the same.
Thats why I think Nvidia is safe betting their launch with the big boy GP100 first. And I think desktops are a priority. I dont think they will have enough TSV and HBM chips to cover many products for desktop and mobile in the beginning. Take GK104 for example, from Kepler. We had over 10+ products in total using that chip for mobile and desktop.
I think Nvidia will be waiting for HBM2 production to increase and mature before launching mobile GPUs.

I also think that MXM will need a new specification to use HBM. So that could take some time, if it even arrives. Who knows, Pascal may or may not be soldered on the motherboard only

 

I can understand 195W GK104 being cut down to 125W and stuffed inside a notebook (with thermal and TDP throttling). I can understand 165W GM204 being cut down to 100W and stuffed inside a notebook.

I cannot understand cutting 250W GM200 down to 120W. To say that GM200 can be reduced to 190W, and then somehow magically reduced again to 120W, is rubbish. Your postulating makes no sense.

 

Thats because you are not reading what I`m writing. Just like n=1.

The TDP of the GM200 it will be based on is not 250W.
GM200 isnt locked on 250W. Neither is GM204 locked on 165W

 

Because what you're writing doesn't make sense. Full GM200 is 250W. Full GM204 is 165W. Clearly the full chip is most efficient. 980 Ti has 2 fewer SMs than Titan X but still 250W, Titan X has higher perf/W. 970 has 3 fewer SMs than 980 but only drops 20W, 980 has higher perf/W. 780 and Titan have various numbers of SMs chopped off compared to 780 Ti and Titan Black but still 250W, the latter two have higher perf/W.

 

sure theyre not exactly the same, but in the end it comes down to how much copper, how many heatpipes and how much airflow u can squeeze into one laptop. combining the two separate gpus would also mean combining their two cooling systems into one massive system with two attached fans. in the end, ull still be able to handle 200-250W TDP. im aware thar such a laptop doesnt exist (yet?), but my point is that it IS technically possible

Sent from my Nexus 5 using Tapatalk

 

Possible? Maybe.

Practical? No.

 

If they make a variable TDP as rumored by NotebookCheck, there would be no problem creating a GM200 mobile chip as Cloudfire has explained. The larger die adds more surface area to remove heat with. Stick a fan directly over the heatsink... Seems plausible to me but it would most likely be soldered unless Clevo builds another machine as thick as the P570WM to house such a design.

 

Variable TDP doesn't make much sense to me. How can you have a single GPU product (990M) that varies in performance by that much? It'd be very difficult to market. What if a customer buys one of the lower TDP ones and then is sorely disappointed because they've seen the benchmarks from the higher TDP versions? It wouldn't be like previous NVIDIA mobile GPUs that have varied in performance by a maximum of ~10% (e.g. 860M Kepler and Maxwell).

 

There is rumor regarding new version of P570WM. And there are tease for G751's follower plus a super awesome flagship model.And benchmark on Quadro M5000M

 

What I said about efficiency was if you overly cut down a large die, at some point it goes over the perf/watt cliff and efficiency goes to hell. You're comparing a full GM204 to a full GM200 or slightly cut down GM200 which is totally different. I'm saying by the time you cut down a GM200 part and downclock it enough to allow it to fit within half its nominal TDP, the perf/watt will completely suck, and you would've been better off starting with a smaller and lower TDP die. This is best exemplified on the desktop side by GTX 460 vs 465.

Desktop GTX 465 is a very cut down 256-bit GF100 with the following specs:
352/44/32 core/TMU/ROP
607/1215/3206 MHz core/shader/memory
200W TDP ( AnandTech)

Then you have the full GF104 based original GTX 460 1GB
336/56/32 core/TMU/ROP
675/1350/3600 MHz core/shader/memory
160W TDP ( AnandTech)

Pretty glaring isn't it, slightly fewer cores but more TMUs, same ROPs, 10% higher clockspeeds, and yet the 460's TDP is still 20% lower. (160W vs 200W) And if you check out the links I posted, you'll see the 460 is neck to neck with 465, edging it out in most titles, and not more than 10% behind when it does lose. All the while having 40W less TDP. In particular, AnandTech made this statement:

So this reinforces my point about dropping off the perf/watt cliff if you cut down too far, and why a heavily cut down and downclocked GM200 with special magical binning to fit within 120W TDP is going to suck balls compared to a proper full GM204.

If you're still not convinced, well this graph should say it all:

 

Certainly poppin the popcorn for this reveal.

 

new version p570wm? i'd be interested in that, and see new 990m or pascal gpu in it, along with haswell-E / broadwell-E

 

This is why I should stop writing on these forums. Sigh

GTX 980Ti have the same TDP like GTX Titan X despite having 256 less cores because 980Ti runs in average 60MHz higher than Titan X.

I am talking about a desktop chip with 640 less cores than Titan X and around 120MHz LESS on the cores. And lower clocks on the VRAM and two memory buses disabled.

 

980 Ti runs at higher voltage too. It doesn't change what I said, which is that the un-neutered chip has higher perf/W. Again, if the full die operates at peak efficiency, how can you say that it can be scaled it down while maintaining that same level of efficiency (or even better efficiency according to you)? Very hard to believe, I have yet to see it happen. @n=1 explained it well.

 

No it changes everything you said. If n=1 stopped using the really old Fermi as his examples all the time, and started researching a little better he would understand that his example is wrong.

Take full GK110 like GTX 780Ti with 2880 cores and compare that to cut down GTX 780 with 2304 cores and look at efficiency. Thats 580 cores removed and the efficiency is down by whopping 1%....



The GM200 chip Im referring to will have around 640 cores less than the full GM200.

There goes the efficiency argument.

Power consumption of GTX 780 is 50W lower than GTX 780Ti.
A hypothetical desktop GM200 chip with 640 less cores and 100+MHz lower clock would be a drop down to GTX 680 power consumption which we got our GTX 880M from.

Nothing could stop a GM200 from happening on mobile. Except the secret plans from Nvidia. If they care and want the best for our gamers, they would offer that GM200 chip.