NVIDIA Maxwell Speculation Thread

[gamervivek]

Inno3d already have a watercooled Titan X.

http://www.inno3d.com/products_detail.php?refid=192

So that is a departure from the earlier Titans.

 

[Deleted member 2197]

I believe Evga & a few others always came out with water-cooled solutions that included high-end products. Nvidia itself never felt the need to come out with water-cooled "reference" versions of their products.

 

[gamervivek]

Regarding power consumption, can Hawaii clock higher if they reduce it? A 1.2Ghz hawaii with tonga improvements(performance and power(?)) could knock off gtx980 with ease. They have been stuck around 1Ghz since 4890 while a good number of nvidia cards are now getting to1.5Ghz with ease.

 

[Deleted member 2197]

I think places like 3DMark would be a good place to see how cards do with regard to performance/power if interested in pushing cards to their limits.
http://www.3dmark.com/hall-of-fame-2/

 

[Deleted member 13524]

The majority of the time, we're looking at devices that are power-limited. Whichever part needs less power to reach the same performance bar, and the less heroics needed to keep it from burning to a crisp, is likely to have the most headroom.

I understand that from the point of an OEM, an IHV, a company that is building a cloud-gaming service, someone building a GPU-based HPC solution, etc.
But for the final consumer or the marketing directed to that consumer this obsession with Perf/Watt is getting to ridiculous levels.
I'm seeing people (and reviewers advising people) paying 100€ more for a graphics card that performs the same but consumes 100W less.
Why? They'd need over 10 years of normal usage to compensate for that price difference on the electricity bill.

 

[3dilettante]

I understand that from the point of an OEM, an IHV, a company that is building a cloud-gaming service, someone building a GPU-based HPC solution, etc.
But for the final consumer or the marketing directed to that consumer this obsession with Perf/Watt is getting to ridiculous levels.
I'm seeing people (and reviewers advising people) paying 100€ more for a graphics card that performs the same but consumes 100W less.
Why? They'd need over 10 years of normal usage to compensate for that price difference on the electricity bill.

Overclocking benefits from a design that isn't already hitting its head on a ceiling. I've seen anecdotes that Nvidia's chips do well in that regard.
A card consuming 100W less in the upper performance range is also probably going to be quieter and fit a lot more cases and power supplies. Needing a new case and/or power supply could make the price margin closer.

Marketing-wise, AMD probably would have been served better if its standard cooler was better, and pushing that close the upper power ceiling meant it was forced to use weaselly "up-to" clocks. Because it didn't, for multiple card generations, it lead to a product that behaved and sounded sloppier.

 

[swaaye]

I understand that from the point of an OEM, an IHV, a company that is building a cloud-gaming service, someone building a GPU-based HPC solution, etc.
But for the final consumer or the marketing directed to that consumer this obsession with Perf/Watt is getting to ridiculous levels.
I'm seeing people (and reviewers advising people) paying 100€ more for a graphics card that performs the same but consumes 100W less.
Why? They'd need over 10 years of normal usage to compensate for that price difference on the electricity bill.

There's probably more to the decision than just power consumption. Dealing with a 300W card comes with additional considerations beyond the power bill.

 

[lanek]

Regarding power consumption, can Hawaii clock higher if they reduce it? A 1.2Ghz hawaii with tonga improvements(performance and power(?)) could knock off gtx980 with ease. They have been stuck around 1Ghz since 4890 while a good number of nvidia cards are now getting to1.5Ghz with ease.

I push my 7970's without problem to 1400mhz ,, (h2o)..

Dont look at the corespeed reported, its completely bugged most of the time with 3Dmark.

 

[homerdog]

I'm going to hold out for Pascal (or Arctic Islands) anyway. Maxwell is good but doesn't feel like a true generational leap over my 670OC.

I went from a 670 (slight factory OC) to a 970 and I would definitely describe the difference as "generational". The 970 is clearly much faster than the 670 and I only have a 1080p monitor.

 

[pjbliverpool]

I went from a 670 (slight factory OC) to a 970 and I would definitely describe the difference as "generational". The 970 is clearly much faster than the 670 and I only have a 1080p monitor.

Yeah I don't doubt you're right, but I tend to only jump when I can get 2-3x the performance. Titan X gives a little over 2x the performance of a 680 right now so I'm hoping a "970 level" Pascal will be a bit faster plus go to full Tier 3 DX12 feature compliance (or further).

The 670 (at 680 performance levels) is actually holding up really well at the moment. I like to see things really start to creak before I move, it makes the new GPU feel a bit shinier to me.

 

[liquidboy]

in chinese sadly ....

Also shows Dx12/11 feature level stuff

WinHEC 2015: "DX12 : new gpu features of nvidia's maxwell architecture (presented by NVidia)"

 

[homerdog]

Yeah I don't doubt you're right, but I tend to only jump when I can get 2-3x the performance. Titan X gives a little over 2x the performance of a 680 right now so I'm hoping a "970 level" Pascal will be a bit faster plus go to full Tier 3 DX12 feature compliance (or further).

The 670 (at 680 performance levels) is actually holding up really well at the moment. I like to see things really start to creak before I move, it makes the new GPU feel a bit shinier to me.

Yeah I'm usually the same way but I had the itch . Before the 670 I was rocking a GTX260 (which I might add ran most things pretty well on the 1440x900 monitor I had at the time).

 

[silent_guy]

To whom?

To those who prefer to talk architecture instead of purchasing decisions.

 

[iMacmatician]

Something I found funny… if you look at the bottom of the GTX TITAN X page, the fake die picture of GM200 looks just like the fake die picture of GM204, with the same number of fake-SMMs and all.

 

[Deleted member 13524]

To those who prefer to talk architecture instead of purchasing decisions.

"Talking architecture" doesn't exist inside an impenetrable shell, isolated from the outside world.
"Architecture" for a consumer-oriented IC exists with one clear and very specific purpose: to influence purchasing decisions.
Because these wretched purchasing decisions happen to put all the influence on either the aforementioned architects get a big bonus or a month's notice to find another job.

 

[silent_guy]

"Talking architecture" doesn't exist inside an impenetrable shell, isolated from the outside world.
"Architecture" for a consumer-oriented IC exists with one clear and very specific purpose: to influence purchasing decisions.
Because these wretched purchasing decisions happen to put all the influence on either the aforementioned architects get a big bonus or a month's notice to find another job.

See the replies of other people. But I'm more than willing to repeat: one way or the other, perf/W seems to matter a whole lot for a whole lot of people.

For the engineers, it means that they don't need an as-beefy cooling solution, or it means that they can squeeze out more performance for the same cooling solution.

For the sales people, it means that they can make more money.

For the marketing people, it means they have something to brag about. It doesn't matter whether or not this makes a difference in real life, consumers like to have things that others don't have. All those gamers who bought the original Titan didn't need FP64, but I bet that a lot of them at least partially justified their purchase for that reason anyway. Just like I would argue, loudly, with my high school buddy about the merits of a regular keyboard over a rubber one, because fate had given me parents who bought a C64 instead of a ZX Spectrum, you'll find tons of people who'll argue about the superb perf/W of a GTX 980 for no other reason than that they own one.

For the tech review website, it means that they have another graph about which they can write stuff, preferably on a different page for extra hits. It's another parameter in which they can pit one against the other. Create controversy. Readers eat up that shit... and it influences their purchasing decision.

For consumers, it means that their fan may not ramp up as fast. That it can be used in a computer with a wimpy power supply. That it may overclock better than the competition. That it keeps the die temperature lower (some people care about that, for some dumb reason.) That the smaller cooler will fit into a smaller case. That they can be proud of their GPU and don't feel a fool for having an expensive product that others mock as Thermi.

For those who don't play games and don't buy GPUs, but like to waste their time on a GPU architecture forum, it's just another topic of interest.

Did you notice that lack of 'cost of electricity' in all of the above? Yeah...

 

[spworley]

Perf/watt is indeed the metric to focus on, and will be even more so in the future. CUDA developers are being taught this explicitly. In a GTC session yesterday, M Hunt from NVidia gave a good talk on optimizing SGEMM for power. You can write code that maxes out your flops and/or bandwidth resources, ignoring power. That's the fastest you can make the compute run at a given frequency. But the interesting twist: that's max at some fixed frequency... but now frequencies are functions of wattage! Newer GPUs have boost clocks which only apply when you've got the power budget. So a power-optimized algorithm will likely be less performant, slower per clock, than other approaches, yet may actually end up becoming fastest in an absolute sense once the boost clock applies itself. Future GPUs, being more and more power constrained, will offer higher and higher boost clock ratios. This video at around 18:00 starts talking about this kind of gain on the new GM200 Quadro M6000 (which has a significant 16% boost clock). The takeaway is that starting now, GPU algorithm designs should optimize for compute per watt, not about compute per clock, since the per-watt metric will actually end up more performant in practice.

 

[Alexko]

Something I found funny… if you look at the bottom of the GTX TITAN X page, the fake die picture of GM200 looks just like the fake die picture of GM204, with the same number of fake-SMMs and all.

I guess the Photoshopping intern must have finished his internship.

 

[Kaotik]

Not sure if it's been mentioned, but on Quadro spec comparison site NVIDIA confirms that GM200 supports DX12 FL 12.1

 

[lanek]

I guess the Photoshopping intern must have finished his internship.

Its not photoshopped, its a simple 3D modelisation with some composition added. r.. pcb is just a textures added without normal ( UV )/ or / displacement mapping. Maybe a little bit of bumpmapping. We all know, sockets part dont look like chrome or perfect polished aluminium.. it have look less faked with some anistrop roughness.

at first i will have think this have been made with Autocad3D, due to the render of metl part who is a bit typic, the "core greens" parts too, who is essentially too much linear, the reflect etc..but it could well be another software. With Autocad you cant do the type of error you see on the round corners... its more easy when you use "split edge" with Blender / 3Dmax in addition of subdivide surface ( tesselation )

But its just a representation.. nobody tell you it need to be real.

I will not show the transitors who is perfectly fake, using the same material of the socket.. just zoom on the initial image.

But i dont know why i write this, not important.

new GM200 Quadro M6000 (which has a significant 16% boost clock).

Are you sure of this ?, For what i have understand, M6000 dont have turbo clock speed. ( I can be wrong, but this was not the case on test version, fixed at 988mhz )