NVIDIA Maxwell Speculation Thread

[iMacmatician]

From Videocardz: "Maxwell GM206-250: GeForce GTX 950 (Ti) incoming?"

Both GTX 750 Ti and non-Ti may soon find their replacements, as we’ve just learned NVIDIA has just finished its new processor — Maxwell GM206-250.

Here is what we know for sure:

• the new processor is codenamed GM206-250
• it’s a GeForce product
• it’s a GTX product (which means it’s not GT x40 and below).

If this report is true, then given the potential timeframe for the update, I expect to see Pascal arrive in this segment later rather than sooner.

 

[McHuj]

From Videocardz: "Maxwell GM206-250: GeForce GTX 950 (Ti) incoming?"

If this report is true, then given the potential timeframe for the update, I expect to see Pascal arrive in this segment later rather than sooner.

If that's true, which it likely is, then it's going to suck not having an new exciting product in the mid tier segment this fall. We could be waiting on Pascal in to Q4 2016.

 

[rcf]

GM206-250 isn't enough, Nvidia needs to get rid of Fermi and Kepler in the low-end.
A fanless Maxwell2 card with h265 decoding would be nice.

 

[mczak]

GM206-250 isn't enough, Nvidia needs to get rid of Fermi and Kepler in the low-end.
A fanless Maxwell2 card with h265 decoding would be nice.

I don't disagree it wouldn't be nice, but it's not like the competition has some brilliant up-to-date lineup in that area...
Quite the contrary, Maxwell 1 may not have quite all the newest features, but at least it has the same efficiency, so it's not really missing much.
Nvidia has a pretty complete Maxwell lineup, frankly I don't know why they insist on selling old crap (though gk208 isn't too bad, except it needs to be paired with gddr5 for acceptable performance, gk107 isn't terrible neither, however larger but can be paired with ddr3 for acceptable performance instead) instead of GM108 (which looks like a perfect low-end solution to me) - unless somehow it's impossible (I don't know if it's a "complete" chip, that is it actually has display outputs and such, hell noone even measured the die size of the damn thing).

 

[silent_guy]

GM206-250 isn't enough, Nvidia needs to get rid of Fermi and Kepler in the low-end.
A fanless Maxwell2 card with h265 decoding would be nice.

If they are still selling Fermi and Kepler stuff, it's probably because there's still demand for them? Like laptop vendors who don't want to respin their design or banks or medical companies with certified hardware that don't want to upgrade.

 

[swaaye]

My guess is things are going to change with Windows 10 because DirectX 12 will become a new, critical bullet point. Unless Fermi and Kepler can support that in some meaningful way?

 

[silent_guy]

My guess is things are going to change with Windows 10 because DirectX 12 will become a new, critical bullet point. Unless Fermi and Kepler can support that in some meaningful way?

I assume that DX12 is a non-issue for the applications that these things are used for.

 

[swaaye]

I assume that DX12 is a non-issue for the applications that these things are used for.

Oh but I am referring to the all powerful feature bullet for the needs of advertising a new machine. Practicality does not enter into this equation.

 

[silent_guy]

Oh but I am referring to the all powerful feature bullet for the needs of advertising a new machine. Practicality does not enter into this equation.

Fermi is supposed to get DX12, so that checkmark is taken care of. I don't think the target audience will be upset by the lack of some obscured feature levels.

 

[rcf]

If they are still selling Fermi and Kepler stuff, it's probably because there's still demand for them? Like laptop vendors who don't want to respin their design or banks or medical companies with certified hardware that don't want to upgrade.

Right now Intel's Braswell is the only sane option for watching HD h265 videos. There's also the GTX960, but it's primarily a gaming card. AMD's low-end and mid-end cards are just rebrands without h265 support. Intel's Skylake will support it but I guess only the high-end SKUs will be initially released.
Fanless, low-end Maxwell2 cards would sell better for HTPC than Fermi and Kepler sell for such banks and medical companies.
And h265 decoding is useful today while DX12 games will really show up only during Pascal's timeframe.

 

[mczak]

Oh but I am referring to the all powerful feature bullet for the needs of advertising a new machine. Practicality does not enter into this equation.

Unless I'm mistaken nvidia already promised support for DX12 for chips back to Fermi months ago, so I can't see why this would be a problem (sure the feature level won't quite be there but that's not much of a disadvantage for marketing reasons).
At least for mobile, nvidia ditched Fermi anyway in current series (the slowest part in the 900M series is a 920M, which is gk208). It is true though on the desktop they have plenty of at least Kepler and some Fermi parts still (due to GM108 being mobile only at least for the time being) - well they never bothered with lower end 900 series parts.

 

[trinibwoy]

Oh but I am referring to the all powerful feature bullet for the needs of advertising a new machine. Practicality does not enter into this equation.

Fermi will get the DX12 checkbox.

With the GTX 660 EOL nVidia doesn't have anything in the $150 segment to compete with the 370/265/7850 so the 950 Ti should slot in there.

 

[swaaye]

Thanks everyone for telling me 3 times that Fermi WILL GET DX12!!!!!!!!

The somewhat non-obvious feature level ambiguity thing is an interesting facet to DX12. It will extend the life of these chips even more. Just what chips are actually fully DX12?

 

[mczak]

And h265 decoding is useful today while DX12 games will really show up only during Pascal's timeframe.

Nvidia would like to sell you something with a Tegra X1 in it for that .
Carrizo should be able to do that too (still not really sure of the limits of the h265 decoder there), you could CF that with Topaz for better graphics performance (though Topaz itself has no video decode/encode or even video outputs at all itself). The market is probably too small to really be worth it for an extra chip mostly useful for that - as you said Braswell will do it just as well (granted its performance outside of video decoding isn't quite stellar), and newer cpus will too.

 

[mczak]

The somewhat non-obvious feature level ambiguity thing is an interesting facet to DX12. It will extend the life of these chips even more. Just what chips are actually fully DX12?

What do you mean fully? Both feature level 12_1 and resource binding tier 3? That would be exactly zero (Maxwell 2 has 12_1, GCN has resource binding tier 3).
Feature level 12_0 (which includes resource binding tier 2) would be more - on the AMD side everything GCN 1.1 and up, on the nvidia side everything with Maxwell, and on the intel side skylake.
There was a nice table here: http://wccftech.com/directx-12-supp...12-1-gcns-resource-binding-tier-3-intels-rov/

 

[silent_guy]

...

Many arguments can be made that Fermi and Kepler chips aren't the modern, but the simple fact that these skus still exist means there's a profitable market for them. That's really all that matter, doesn't it?

 

[rcf]

Many arguments can be made that Fermi and Kepler chips aren't the modern, but the simple fact that these skus still exist means there's a profitable market for them. That's really all that matter, doesn't it?

The fact that old stuff is good enough for many people has never been an obstacle to releasing new stuff to sell them. It's more likely that Nvidia doesn't refresh the low-end as often as before because they are selling less due to IGPs getting better.

 

[pjbliverpool]

What do you mean fully? Both feature level 12_1 and resource binding tier 3? That would be exactly zero (Maxwell 2 has 12_1, GCN has resource binding tier 3).
Feature level 12_0 (which includes resource binding tier 2) would be more - on the AMD side everything GCN 1.1 and up, on the nvidia side everything with Maxwell, and on the intel side skylake.
There was a nice table here: http://wccftech.com/directx-12-supp...12-1-gcns-resource-binding-tier-3-intels-rov/

'Fully DX12" has nothing to do with resource binding tier 3 though. That's a feature which sits outside the required DX12 spec. I'm also not clear on it's usefulness (over tier 2) in comparison to CR and ROVs (who's advantages have been made clear). I'd be interested to better understand that though.

 

[fellix]

Does GK110 (Big Kepler) supports async compute scheduling in D3D12 and is it related to the dynamic parallelism feature in CUDA CC v3.5?

 

[Deleted member 13524]

This sounds like the card that will finally put 3.5 year-old Pitcairn (HD78x0+HD88x0+R9 270+R7 370) to sleep.
AMD tried really hard to put this 90-year-old man in a wheelchair to continue working on their coalmine, but the competition is now probably going to sweep it away.