Pascal: What do we know? Discussion, Latest News & Updates: 1000M Series GPU's

________________________________

From the looks of it, Nvidia is going to supposedly release Pascal in June/July, to time it with the AMD Polaris release, which makes sense.
It also underscores why having AMD around as competition is a benefit for all of us, to help keep Nvidia on their toes to not just push out tech advancements, but to do so sooner.
It's a win/win for the consumer, and in our case, the PC gamer.

 

I don't think that is GeForce series hardware we're seeing, so I'm not worried.

Performance will be much better than Maxwell. Not +10%, lol.

 

So if I have this right the "gaming" orientated variety will have the FP/compute portions nixed for better performance?

I'll be happy with over 50% improvement over a 980m as its already a ~300% improvement over my 765m lol.

Sent from my SGH-M919V using Tapatalk

 

Official GP100 vs GM200 FP32 performance is about 74% better (Tesla P100: 10.6 FP32 TFLOPS, TitanX: 6.1 FP32 TFLOPS).

So that kinda puts upper bound about what to expect for gaming Pascal GPUs. It would probably be less, Nvidia can afford to go more nuts with $10,000 per-unit Teslas than $1,000 high-end consumer GPUs.

 

I'd say it gives them much more room for improvement on the GeForce series cards. The profit margin on "gaming" GPU's is very high. Tesla's are way different and shouldn't be used to compare performance to gaming. They are actually less "powerful."

Bandwidth should be near 1TB/s on flagship cards and 4x more efficient than Maxwell. But the first iteration we see will likely not show this.

 

I don't know if you can tell that much from the GP100. If you look at the table comparing it to its predecessors (there's a nice one at AnandTech), it has a higher TDP than the GM200 (300W vs. 250W), but many aspects have improved much more than the single precision (i.e. FP32) performance (e.g. the memory bandwidth is up by a factor of 2.5 and the double precision increased by 25). Unfortunately, some of these are not relevant to gaming (the FP64 performace) and others are not likely to make it to consumer products this time around (the memory bandwidth is a result of HBM2 whereas consumer cards will be either GDDR5 or at best GDDR5X).

Also, for laptops, it matters a great deal whether Pascal works better at 300W or at 30W. Every modern architecture can be operated at both points, but performance usually does not scale linearly with power. It could be that the 1060M (or whatever they call the 60W card) is better than the 965M by a significantly larger margin than the desktop 1080 Ti is better than the 980 Ti.

 

I don't think we can tell much of anything for mobile Pascal from GP100 except that it's more powerful. Anything beyond that is speculation.

 

I agree

But speculation will keep me from getting the actual ones and that is good for my later fun-to-have, as my GTX860M is able to pull decent FHD performance.

 

???

Compare die to die performance, GP106 being 10% faster than a crippled GM204? Probably. Will we see full GP104 on mobile soon? Not likely.

Do you really think nvidia would not milk Pascal all the way down to the bone? Considering we will be on 20nm for a long time and Volta wont be ready until 2018.

HBM2 will be arriving on 1080 TI and probably the Titan Pascal. So HBM2 is coming on consumer, just probably not until the end of the year. Depending on 1080 launch date, it may use GDDR5x or GDDR5.

 

nvidia be like: yummmm, so much milking potential!

1st gen pascal: crippled core/gddr5/lower clocks
2nd gen pascal: crippled core/gddr5x/lower clocks
3rd gen pascal: crippled core/gddr5x/proper clocls
4th gen pascal: full core/gddr5x/proper clocks
5th gen pascal: full core/hbm2/slightly higher clocks
6th gen pascal: well, u get the picture...

Sent from my Huawei Mate 8 NXT-AL10 using Tapatalk

 

I am surprised that no-one talks about the giant elephant in the room:

Dual GPU without SLI connector (MXM spec?):

Spoiler

8 MOSFETs, 3+1 large power phases and liquid cooling for a mid-range 4GB card:

Spoiler

 

Haha, I know you're joking with the exaggeration of number of generations, but they do love to milk! I can't imagine there being more than 2 or 3 iterations/generations of Pascal.

 

That's the GM206 in an over spec PCB for the car market.

 

Arn't those getting stuck in a car? I remember something about that and seeing that they would be stuck somewhere with less than stellar air flow. Like a trunk or in the dash.

That being said I think eathram (sorry if I misspelled) said they are going to run super hot.

Sent from my SGH-M919V using Tapatalk

 

http://www.engadget.com/2016/01/04/nvidia-drive-px2/

 

Supposedly it's "blisteringly hot"... which doesn't bode well for mobile cards (as AMD has shown us *cough*). We might have to wait 2 gens to get the full power of pascal on mobile... Maybe the 1280M? Almost definitely looks like the 1080M will not be worth upgrading to at this point for people who want single GPU 4k performance. But then there's the whole BIOS and MXM slot compatibility. What are the chances that the 1280M will be compatible with clevos from 3 years earlier?

 

True, however we have only seen the compute orientated ones I do think. So they might have something different in store for consumer cards.

Sent from my SGH-M919V using Tapatalk

 

"Meaker@Sager", when did this happen?

 

Any source for the reports of the GPUs running (extremely) hot ? Can't find anything about this

EDIT: Plus even if the GM206 in these specific boards will run very hot, how does this relate to the new Pascal GPU's? Isn't GM206 the same used in the GTX950/960 ?

 

It was one of our guys that supposedly has connections, so I'd take it with with a wagon of salt.

Sent from my SGH-M919V using Tapatalk

 

I think we might be mixing things here.
Maybe @Ethrem pointed out that Pascal in general runs (very) hot and that we linked it here to the PX2 which isn't even a Pascal based GPU...

 

Yeah your right.

Plus I haven't exactly been around to pay attention. I'll just have to wait for down the road to see what happens.

Sent from my SGH-M919V using Tapatalk

 

that happened when powernotebooks went out of business and sager snatched him up

 

Look again...
This time it's not their CES January fake where they used two GTX980M/M5500!

They had no Pascal board ready back at CES:

Spoiler

Now at GTC they did:

Spoiler

GP106 Vs GM206:

Spoiler

 

Interesting, quite a bit smaller die than GM204, more in line with GM206 size which is why it caught me out.

 

Prema beat me to it. The PX2 is powered by 2 Pascal GP106 GPUs and 2 Tegra P1 SoCs. It is liquid cooled because not only the Pascal GPUs but the Tegra chips both run hot.

The milking of Pascal will be inevitable due to memory technology availability. Also with Volta being delayed despite the fact that TSMC will have 10nm up and running at production capacity by 3Q with products expected to ship in 2017, possibly even a certain SoC this year, we will be seeing Pascal for awhile...

 

It would be wonderful to get a Titan (GP100) "desktop" gpu for the P870DM similar to the current gtx 980.

 

want to upgrade my gt80 to pascal or get a pascal gt80... how long till the nb chips come out?

Is it worth ordering a VIVE are they any good?

 

That will never happen. Big die is just not going to end up in notebooks again, not after the Fermi disaster that was the 480M. It's simply not possible to cram the big die into a laptop, nVidia was just desperate.

 

That and it seems they are barely going to get mid range into the laptops.

Sent from my SGH-M919V using Tapatalk

 

I don't know the story with that

Can you please post it, Pascal release will not be delayed by a small discussion XD.

 

Yes but GTX 480M was limited to the 100W MXM 3.0b form factor. I'm saying if they make a gtx titan like the 200W GTX 980 or 180W GTX 980 MXM cards we have.

Actually now that I think about it, a desktop 1080 card in that form factor would be more than enough, probably faster than 980M SLI easily.

Desktop GTX 480: GF100 core, 480 CUDA Cores , 700mhz core clock, ran hot, sipped lots of power
Laptop GTX 480M: GF100 core, 352 CUDA Cores, 425mhz core clock, ran hot, sipped lots of power

Basically they were able to get out better performance out of the GF104 based GTX 485M, sporting 384 CUDA cores, 575mhz core clock and it ran cooler.

 

Nvidia has no motivation to do that though. Big die tend to have lower yields than the smaller chips so they would either have to raise the price on the cards or cut into their margins to do it. As I said, nVidia was just desperate with Fermi. The situation is totally reversed now.

 

Nvidia is very desperate on holding their HTC market from knightshill. Hence the massive 610mm^2 on 16nm finfet where knightshill would achieve the same performance on 10nm.

 

I do wonder if we'll get to see another desktop class card in laptops at all. Or if the 200W 980 was a fluke that will not be repeated again for a long time. Good thing the 870DM is compatible with regular mobile cards as well heh, or its lifespan would be cut quite short.

 

GTX 980 desktop are being sold at a roughly 3 x the cost of the same die on desktop. Nvidia like green papers. Why milk the same die only once when you can do it twice and milk it extra hard for the second time?

 

Desktop 980 was a bad move in my opinion. It only showed the little performance difference between a 980m, and it consumed more power, and its unusual specs/size meant it only fit on very specific models, plus a bunch of different MXM appeared with different power targets etc etc.

 

How else could we get those rocket high bench scores and be able to use software that is impossible to use with SLI setups? Especially pro software, like Unity, or Maya, or so.

Also, Desktop GTX 980 in laptops plays decently and has very good performance. Though it might have a bad resell price. Nvidia might had made a good move for themselves, and bad for customers, if you think this in perspective that once you got one of that things, you are locked with it for a while .

 

What are you talking about? A 980m is within 20-25% of the performance. Even a mild overclock closes the gap enough so that the difference is less pronounced. Desktop 980 is nowhere near SLI 980m.... SLI 980m are on 980 ti territory.

There are very very few laptops with Desktop 980 due to the limits I mentioned. And some like the GT80 come with their caveats (Straining the GPUs causes battery drain because PSU is not enough). As for the PSU not being enough, if you don't have a 330W PSU, your overclocks are also limited due to the PSU.

Since there are very few machines that actually support it, and only support one of the versions of the 980 available, its price is very high for very little performance gain.

 

Speaking frankly, and we have Digital Foundry videos to boot, the 980M SLi outright beats the Titan X in quite a few SLi supported games where scaling isn't an issue.

In my opinion, a 50% performance improvement is the bare minimum for something to be deemed a significant upgrade. On that point, the notebook GTX 980 is rather uneventful in comparison to the 980M. Of course, there is an appreciable difference, but the two cards are very much within the same territory.

 

+30-35% faster than the 980M to be exact. i agree with the 50% though, i made it a rule not to upgrade to any hardware thats not at least 50% faster at stock than my best overclock

Sent from my Huawei Mate 8 NXT-AL10 using Tapatalk

 

My most used software does not support SLI, so for me that 35% more is important, but yeah, the problems are there. Though, it is an ultimate GPU, and provides an extreme level of performance, and on P870 from Clevo, it really has a LOT of power, and can show it's teeth.

 

To be truly honest, I can't consider the 980M and the 980 to be in the same class if you're going to mention overclocking at all. The 980M has problems hitting or passing the 1350MHz mark for constant usage (and I don't know about you, but if I'm going to game, and I OC to game, I expect minimum 24-hour gaming stability from the OC), and its memory is honestly pretty poor of an OCer (whenever that matters; which it doesn't seem to matter too much right now, but some games do like memory overclocks). The 980 on the other hand can hit 1400MHz easily stable and 8000MHz on the memory (with 980M topping out under 6000MHz on average) while innately being 25% stronger in terms of core count.

As far as I can see, the two aren't even in the same class. BUT the price of the 980 and its limited usage scenarios make it a difficult sell for people. For gamers who play a variety of games? Buying it and OCing it is best. For gamers who play a smaller subsect of games can just check whether or not their most played games will use SLI, and make their decision then.

Now, more than I've ever remembered, getting a dual-GPU setup is "you better know what you're doing" than ever before, and you simply need to accept the fact that on launch, games are very unlikely to work with multi-GPU, and are very likely to work badly if they do.

Note that I still feel the 980 is way overpriced. Even making allowances for the overpriced nature of notebook GPUs, the 980 should've taken over the ~$700 pricepoint of the 980Ms and had the rest of the notebook GPUs drop in price. a $500 980M and a $350 970M would be pretty much bliss for notebook gamers as far as things currently are.

 

well THERES your problem

 

This thread needs a "milk" counter.

 

ikr? How dare we not have to overspend!

 

Emulating Nvidia GPUs - 4/10/2016

 

Yes, those numbers are pretty good, unfortunately it doesn't meet my qualifier that I require at least a bare 50% increase -- a whole third of a performance leap -- O.C. and otherwise to consider anything a substantial upgrade, and that doesn't appear to be the case, even at its most generous disparity in what one can milk out of the notebook variant of the GTX 980. It would appear that our respective definitions of territory/class differ in this case and I have no issues accepting this margin of disagreement because you do raise fair points. As I see it, the notebook GTX 980 just isn't that huge of a leap over the 980M; of course, no one contends that it isn't a noticeably better card with better headroom for overclocking.

Throw in the price difference and the point highlighted above is only further compounded, but that is a distinctly different issue that doesn't speak exclusively to the strengths of the card. Hell, I find that even the 980M is quite overpriced.

As for SLi . . . Yes, there are issues at the moment and things aren't looking too great for its future, although that too is an entirely distinct issue at this stage, because one could apply that to the GTX 980 SLi versus the next beefed-up single GPU.

 

Sli support is awful these days, for professional applications.

IF the next GTX1080M or what the notebooks will have inside will not be stronger than GTX980Desktop, than it is already a dead end.

It is all just a speculation, but I think that I will be able to get a Clevo P870DM-G (the king of notebookes) even after Pascal is first released. I mean, if desktop Pascal will be 100% or 50% better than their predecessors, it is enough of a proof to wait. The only real downside of P870DM-G is the weight, which is a bit high at the moment, but I would have that weight any day over a lower performance.

 

Cool, didn't know they emulated the chips using super powerful computers & then just connected that emulated GPU to a regular PC motherboard to see how their latest design of NVidia GPU functions without even fabricating a GPU - good idea!