NVIDIA Maxwell Speculation Thread

[Alexko]

I'm not saying that gm204 isn't amazing in performance/watt (970 in the performance/price too), I'm saying that I preferred as it once was (before Kepler and TMSC problems), when came out the top gpu and then the other, according to the scheme "from top to low". if you look at it from the point of view of the "entusiasth"* perhaps you'll understand what I mean. Amd do still in this way.

Do they?

They launched GCN 1.0 with Tahiti, then smaller GPUs, true. But then they introduced GCN 1.1 with Bonaire, waiting quite some time before bringing it to the high-end segment with Hawaii. Then they introduced GCN 1.2 with Tonga, which is similar to Tahiti and somewhat high-end but still below Hawaii.

 

[xDxD]

Do they?

They launched GCN 1.0 with Tahiti, then smaller GPUs, true. But then they introduced GCN 1.1 with Bonaire, waiting quite some time before bringing it to the high-end segment with Hawaii. Then they introduced GCN 1.2 with Tonga, which is similar to Tahiti and somewhat high-end but still below Hawaii.

Come on, bonaire was not sell like the top one at 500$ it's not the same playground, same with Tonga.

 

[Alexko]

Come on, bonaire was not sell like the top one at 500$ it's not the same playground, same with Tonga.

They would have sold for that much, had they been fast enough.

 

[Blazkowicz]

The top GPUs are kind of an "enterprise" product now as well as something for gamers, it's possible that not rushing them allows to catch more hardware bugs. (makes the bins slightly better too)

If anything that's a sign of maturity.

 

[trinibwoy]

They would have sold for that much, had they been fast enough.

Yeah there's really no comparison between Bonaire or Tonga and GM204. The latter was the fastest thing available at launch.

 

[ams]

I'm very skeptical about this MFAA thing: anything that does temporal filtering is highly suspect. It's probably fine for static images but I don't see how it can work anything that moves.

You could be right, but far too early to tell at this point how effective MFAA can or will be. NVIDIA claims that MFAA will in some cases provide even better image quality than MSAA (and vice versa too I would imagine). MFAA requires Maxwell Gen 2 hardware with it's new more programmable and more flexible pixel and subpixel sampling pattern ability in the ROP's, so we won't be seeing this on Kepler or older gen hardware. According to Anandtech, when in motion, the image quality difference vs. MSAA is going to be highly dependent on the quality of the temporal reprojection (aka temporal synthesis filter) that NVIDIA employs. So while MSAA is clearly still here to stay, it will be interesting to see how effective MFAA is in the future across a variety of different games, and the ~ 30% performance improvement over MSAA may make it a good tradeoff in some scenarios too.

 

[Ailuros]

Yes i know that, and i can't understand why they do not expect to have ready GM200 (is said to have made ​​the tape out shortly after gm204 right?) and give each gpu its rightful place (and price?)

204 taped out in early April and 200 in June; considering the latter is quite a bit more complicated I wouldn't expect it within this year.

 

[xDxD]

204 taped out in early April and 200 in June; considering the latter is quite a bit more complicated I wouldn't expect it within this year.

Well, wait for it in january /february perhaps, hope not on a 1000$ Titan 2

 

[Ailuros]

Well, wait for it in january /february perhaps, hope not on a 1000$ Titan 2

I'm not waiting for it one bit; if I'd buy a GPU today a 970 would be perfectly sufficient for my needs. As for $1000 Titans, as long as they can actually sell them at insane prices they obviously will.

 

[A1xLLcqAgt0qc2RyMz0y]

204 taped out in early April and 200 in June; considering the latter is quite a bit more complicated I wouldn't expect it within this year.

Considering that 28nm at TSMC is now very mature and understood shouldn't that imply that Nvidia could release the GM200 this year if they wanted to?

 

[Alexko]

Considering that 28nm at TSMC is now very mature and understood shouldn't that imply that Nvidia could release the GM200 this year if they wanted to?

Process issues aside, if GM200 follows the usual pattern it probably includes a few new features and should require a slightly longer validation period.

Plus, they might want to launch it in Quadros/Teslas first, and those require more testing still.

 

[dnavas]

Process issues aside, if GM200 follows the usual pattern it probably includes a few new features and should require a slightly longer validation period.

I'm wondering if NV is moving to more of an Intel-style tick/tock release style, given that they're stuck on 28nm. So, Maxwell on 28nm, then a shrink, then Pascal on the same process.... Under that assumption, either they'll have process issues, or they'll have new features, but not both. [And, fwiw, I agree that it's likely to take longer than GM204.]


Oh, and Blazkowicz, looks like you might be getting adaptive sync after all....

Edit: Although, now that I re-read, I suppose you were saying "[Lack of] process issues aside...", and not that you were presupposing an imminent release on 20nm. If that's so, please ignore :>

 

[mczak]

Process issues aside, if GM200 follows the usual pattern it probably includes a few new features and should require a slightly longer validation period.

What do you base this on? Purely gk110? I don't think that's a valid comparison, since gk110 appeared much later (and, it was named gk110 after all, not gk100).
Wouldn't surprise me at all if it's just a beefier Maxwell GM2xx (that's already 2nd generation Maxwell, after all). Plus the usual compute stuff of course (higher DP rate, ECC).

 

[Blazkowicz]

If Ailuros is right GM200 taped out right after 204, and they have the same name mostly.. I agree that they might have the same features. As well as DP and ECC maybe it has e.g. a 4MB L2 and a 512bit memory bus and so you can work on more code and more data.

The rest of the features are locked away in GRID or Quadro. Some of them do affect gamers - for both xx0 and xx4 chips, and lower : running more than 4 displays or using VGA pass-through with IOMMU (run a Windows gaming VM on a bare metal hypervisor). These are fringes cases, but it's like Intel playing games with little features on i5, i7, Pentium, Xeon etc.

 

[Ailuros]

If am right then GM200 could have a <10% higher transistor density than GM204. Higher complexity and theoretically higher density don't sound to me like a project that could get from tape out to mass production in just 5 months; and no there's not any headroom for a wider than 384bit bus IMHO.

In order to come close to ~12 DP FLOPs/W as they've projected, they'd need at least =/>2.6 TFLOPs DP. Do some backwards math with a theoretical frequency of ~0.85GHz which still sounds reasonable to me under 28nm and they'd need somewhere over 1500 FP64 SPs to get there. GM204 has only 64 FP64 SPs after all, so the overhead for FP64 units in GM200 is quite sizeable.

 

[gl33k]

9xx is fresh.
why is my crossfire is fucked ,technically its bullshit for peasant, "i am sorry for myself" is the lesson

 

[xDxD]

According to our source chip should have the following key features:

Chip size from 10 to 15 percent over the GK110 (551 mm2)
Number of CUDA in 2816 (20-22 SMM blocks), apparently the first version will be active all chip in SMM
Memory bus 384 or 512-bit (rather I-384 bit)
Gaming and professional card should come out at the same time, a game like Titan X
Launch of Q4 / 2014
Performance Titan X - 40-50 percent over Titan Black

https://translate.googleusercontent.com/translate_c?depth=1&hl=cs&ie=UTF8&nv=1&prev=_t&rurl=translate.google.cz&sl=cs&tl=en&u=http://pctuning.tyden.cz/component/content/article/1-aktualni-zpravy/30978-geforce-gtx-990-a-nejspis-i-titan-x-na-ceste&usg=ALkJrhgbBJ_ROe91goIDizN7elePDWLskA

 

[iMacmatician]

22 SMMs seems like a weird number unless it will have GPCs with differing numbers of SMMs.

 

[xpea]

According to our source chip should have the following key features:

Chip size from 10 to 15 percent over the GK110 (551 mm2)
Number of CUDA in 2816 (20-22 SMM blocks), apparently the first version will be active all chip in SMM
Memory bus 384 or 512-bit (rather I-384 bit)
Gaming and professional card should come out at the same time, a game like Titan X
Launch of Q4 / 2014
Performance Titan X - 40-50 percent over Titan Black

https://translate.googleusercontent.com/translate_c?depth=1&hl=cs&ie=UTF8&nv=1&prev=_t&rurl=translate.google.cz&sl=cs&tl=en&u=http://pctuning.tyden.cz/component/content/article/1-aktualni-zpravy/30978-geforce-gtx-990-a-nejspis-i-titan-x-na-ceste&usg=ALkJrhgbBJ_ROe91goIDizN7elePDWLskA

when you are not even sure about mem bus width, it's merely a guess game than anything serious...

 

[dnavas]

22 SMMs seems like a weird number unless it will have GPCs with differing numbers of SMMs.

Seems like some intelligent guessing. 50% higher bandwidth gets you a 384bit interface, 96 ROPs (assuming GM204 ratios carry over), and 24 SMMs (similarly). Could be you guess that an SMM or two get switched off. The only problem is that if GK110/GK104 ratios carry over, the 40-50% increases in math/bandwidth translates into (at least) an 87% increase in die size. That would be 744mm^2. Which is pretty big, really.... If ~620mm^2 is accurate, it would imply that the items added into GK110 over GK104 are already part of GM204. That'd be interesting, if true.

The more likely possibility is that 620 is accurate, and the growth in die-size is ~90%. Do the reverse math and you wind up with ~19SMMs, so 20 seems like a pretty reasonable guess. That puts you at 25% faster than the 980, unless you're bandwidth/pixel limited.

Sounds like they're guessing to me, but, it's not a crazy guess at least.