(WCCF) Nvidia ‘Big Daddy’ Maxwell GM200 GPU Alleged Specs Leaked

[ShintaiDK]

How likely is it that this lands before The Witcher 3?

I think its very likely. Witcher 3 is feb 24th after all.

 

[badb0y]

I think its very likely. Witcher 3 is feb 24th after all.

I can wait then. Seems like this is the card to get and considering SLI has been worthless for me I'll just grab a 980 Ti.

 

[Enigmoid]

I don't think GM200 can have as big a leap over GM204 as GK110 over GK104 because there isn't enough die space left moving up from 398mm^2 to accommodate all the compute-specific transistors on top of more memory controllers, ROP's, and SMM's. GK110 had 87.5% more cores than GK104 and a 250mm^2 larger die size. In my mind (before this thread was started), I was thinking GM200 could maybe squeeze out 50% more cores than GM204, adding up to 24 SMM's and 3072 cores. I still think that is a possible outcome. 22 SMM's make no sense because nothing adequately divides into 22 for balance purposes, and since 20 is a rather small (25%) increase over GM204's 16, 24 makes the most sense to me.

But 50% more cores, ROP's, and bandwidth != 50% more performance. I think 35%, maybe 40% max more performance than GM204 with a real-world power usage slightly lower than GK110 is both feasible and reasonable for a fully functional GM204 chip. On top of that, I think GM200 is going to come before GCN 2.0, so I think Nvidia will slow roll GM200 just like they did with GM204. If it is to have 24 total SMM's, I think Nvidia would release a gtx 980 TI with 20 SMM's active and DP castrated. Titan II would have 24 SMM's and full functional DP operation. Then later on Nvidia will release a full functional, DP castrated GTX 985 (or whatever) to combat a new high end AMD chip, or if no chip exists, just to get more sales at a higher margin than 980 TI.

TL;DR I think GM204 on 20nm will have a slightly lower TDP than GK110, but will not be quite as much a jump over GM204 as GK110 is over GK104. I also think GM200 will have some serious OC headroom just like GM204 has, and I think Nvidia will slow roll GM200 (introducing harvested dies first) like they did with GK110.

I agree but don't forget that Maxwell reuses DP shaders as single precision unlike Kepler. So their will be much less die bloat. This is why Maxwell moves from 1/24 mode to 1/32 mode; the double precision shaders are operating in single precision mode.

 

[Ajay]

TL;DR I think GM204 on 20nm will have a slightly lower TDP than GK110, but will not be quite as much a jump over GM204 as GK110 is over GK104. I also think GM200 will have some serious OC headroom just like GM204 has, and I think Nvidia will slow roll GM200 (introducing harvested dies first) like they did with GK110.

I guess we can hope it's on 20nm. If that's the case, then the boost over GM204 would much more than 50% (20nm is behaving somewhere between a full node and 1/2 node shrink - hence the reason Intel went with FinFet @ 22nm). If we were talking 6 months out, then there might be a reason to hope, 20nm has been a b*tch for TSMC. They have been running R&D on triple shifts to accelerate yields.

If it's 28nm part, then were are just talking about verification, inventory build up and market timing - and an ~40% performance jump depending on the game/resolution. Not chump change, but the the AIBs built with such a large GPU will still be expensive (despite being on 28nm) - NV has to get their R&D money back.

 

[CrazyElf]

I don't think GM200 can have as big a leap over GM204 as GK110 over GK104 because there isn't enough die space left moving up from 398mm^2 to accommodate all the compute-specific transistors on top of more memory controllers, ROP's, and SMM's. GK110 had 87.5% more cores than GK104 and a 250mm^2 larger die size. In my mind (before this thread was started), I was thinking GM200 could maybe squeeze out 50% more cores than GM204, adding up to 24 SMM's and 3072 cores. I still think that is a possible outcome. 22 SMM's make no sense because nothing adequately divides into 22 for balance purposes, and since 20 is a rather small (25%) increase over GM204's 16, 24 makes the most sense to me.

Realistically, the GM200 may be perhaps as little as 30% faster than GM204, perhaps even less. Two exceptions.

First, it`s technically possible for a >600mm^2 die to come out. Anyone know what the reticule size is? Perhaps as a process matures, maybe a larger die will be practical?

Second, there's the possibility I suppose (slim) that there have been some minor architecture improvements in the months in between.

On top of that, I think GM200 is going to come before GCN 2.0, so I think Nvidia will slow roll GM200 just like they did with GM204. If it is to have 24 total SMM's, I think Nvidia would release a gtx 980 TI with 20 SMM's active and DP castrated. Titan II would have 24 SMM's and full functional DP operation. Then later on Nvidia will release a full functional, DP castrated GTX 985 (or whatever) to combat a new high end AMD chip, or if no chip exists, just to get more sales at a higher margin than 980 TI.

Possibly. I think the Titan was only released with 14/15 SMX because 28nm yields were not good enough at the time for such a big chip. 28nm is mature now so, there's probably less technical reason for this to happen. Of course, it possible that Nvidia might do it for milking purposes.

I highly doubt that we are going to get a 20nm chip though - TSMC is only offering 1 process for 20nm. Reportedly the 20nm process was only 10% faster at high power for a given level of power consumption than 28nm. So I expect high power to jump straight to 16nm FinFET, unless they change their plans.

 

[Ajay]

I highly doubt that we are going to get a 20nm chip though - TSMC is only offering 1 process for 20nm. Reportedly the 20nm process was only 10% faster at high power for a given level of power consumption than 28nm. So I expect high power to jump straight to 16nm FinFET, unless they change their plans.

Yeah, but 16FF+, the next most desirable node won't be ready for a large die GPU for at least 2 years or maybe longer!

20nm doesn't need more speed (faster clocks), just more xtors. GPUs are essentially highly parallel stream processors - adding more CCs gives the most bang for the buck.

 

[CrazyElf]

I know. We could be stuck on 28nm for a while it is looking like.

The other issue is that price per transistor is no longer going down. It's going up. 28nm is looking like the cheapest process. The problem is of course, to add more cores for a bigger die needs more transistors ... something that is no longer getting cheaper with each new node.

My guess is even Intel is feeling the pain at 16nm.

 

[Grooveriding]

I can wait then. Seems like this is the card to get and considering SLI has been worthless for me I'll just grab a 980 Ti.

I don't think it will be too long, maybe even before the year's end. GK110 has been available as their pro card for almost as long as gtx680 has been available, two years ? Probably time for a refresh. I also doubt we'll see the bifurcation we saw in gk110 with titan, 780 and 780ti. I don't think there will be as much room performance wise between gm204 and gm200 to fit three cards. My thought is a titan as the full gm200 and a cut down gm200 as 980ti.

Full gm200 should be enough to hopefully give 40% over a 780ti and constitute a large enough performance increase to make max settings 4K gaming viable with multi gpu setups.

 

[Kippa]

Tech wise if they do release a Titan II, how much video ram could they realistically put on it? The Titan I had 6gb, do you think they might up it to 8gb or even 10gb?

 

[Grooveriding]

Tech wise if they do release a Titan II, how much video ram could they realistically put on it? The Titan I had 6gb, do you think they might up it to 8gb or even 10gb?

I think we'll see another 384bit bus, so it will have either 6gb or 12gb. Going on my other hypothetical post my expectation is that a 980ti using GM200 will have 6GB and a Titan 2 with GM200 will have 12GB.

We may not even see a Titan repeat, perhaps it will just be two standard geforce cards with a ti version using the full chip and a second one as a cut down version. Both shipping with 6GB of memory. It's tough to see how two such cards would be priced with a $550 gtx980 there as well. 28nm is mature now, perhaps just one full chip GM200 card for $700-$800 could be what gets released.

 

[nvgpu]

http://www.gputechconf.com/

MARCH 17-20, 2015

GM200 will probably debut first as a Tesla card, Nvidia biggest compute customers have been waiting for a long time for a more power efficient architecture to power their largest DP workloads and reduce power consumption.

 

[Mondozei]

We'll see if the GM200 is priced like Titan or like the 780 Ti. If it's another Titan, who cares? But if it will be priced like a 780 Ti, then the ones who bought a 980.. well, feel sorry for you.

I'm guessing the pricing will depend on AMD. I was surprised not to see anything of consequence from them, after their "watch this space on 25th sept" tweet.

 

[jpiniero]

Yeah, I suspect for the consumer models there will only be that one model - the Titan, and $999.

 

[therealnickdanger]

GM200 will come out as a response to AMD's next move, whenever that is. If AMD unveils the 390X for $500-600 with better performance than 980, then Nvidia will reduce the $549 price of the 980 (built-in extra margin), then unveil the GM200 at an even higher price point (600-800) with higher performance. Everybody wins... except people who buy the 980 now.

 

[ShintaiDK]

Everybody wins... except people who buy the 980 now.

They already won by buying now.

Else you can postpone upgrades forever because something new is always coming in the future.

 

[CrazyElf]

We'll see if the GM200 is priced like Titan or like the 780 Ti. If it's another Titan, who cares? But if it will be priced like a 780 Ti, then the ones who bought a 980.. well, feel sorry for you.

I'm guessing the pricing will depend on AMD. I was surprised not to see anything of consequence from them, after their "watch this space on 25th sept" tweet.

I get the feeling we are going to see anything of consequence until at least next year.

Hopefully they have something decent or the Nvidia GPUs are going to cost a lot.

 

[PhIlLy ChEeSe]

Tic Took, Nvidia raises the cheese again........................:awe: Same old story all ova again.

 

[therealnickdanger]

They already won by buying now.

Else you can postpone upgrades forever because something new is always coming in the future.

970 = incredible perf/w/price
980 = incredible perf/w but is $100 too expensive

With OC, you can get the 970 up to stock 980 performance for $220 less... almost enough to buy a second 970 for SLI (if you like that sort of thing). The 980 is being marketed as a flagship card even though we all know that the real Maxwell flagship will be here in a couple months. Seems like if you're really anxious to upgrade, get the 970 now, because when you go to sell it in three months for the 980Ti, it won't lose as much value compared to the 980.

 

[x3sphere]

970 = incredible perf/w/price
980 = incredible perf/w but is $100 too expensive

With OC, you can get the 970 up to stock 980 performance for $220 less... almost enough to buy a second 970 for SLI (if you like that sort of thing). The 980 is being marketed as a flagship card even though we all know that the real Maxwell flagship will be here in a couple months. Seems like if you're really anxious to upgrade, get the 970 now, because when you go to sell it in three months for the 980Ti, it won't lose as much value compared to the 980.

You may have a point but I couldn't care less if the 980 loses more value. I bought it as a long term upgrade. Won't upgrade until the 980 "replacement" comes, which probably will be 1.5-2 years from now, and hopefully we'll be on something other than 28nm by then.

 

[Cerb]

There's always newer and better around the corner. It's just how much better, for how much money. Am I going to get my money's worth out of it, for the 2 years I intend to keep it?

 

[ShintaiDK]

There's always newer and better around the corner. It's just how much better, for how much money. Am I going to get my money's worth out of it, for the 2 years I intend to keep it?

It was 2½ years since I got my GTX680 that the GTX980 replaced. So its certainly possible.

 

[Cerb]

It was 2½ years since I got my GTX680 that the GTX980 replaced. So its certainly possible.

I just meant that as a way to make a judgment call.

I'm now 4 years into my GTX 460. Spending Nearly $400 for a GTX 760, or $450+ that for a GTX 770, didn't sit well, but I got close a few times, along with some of the good 280X and 290 deals that have come up. While idle noise has improved, it hasn't improved enough for me to use stock cooling, based on building PCs for others with much newer HW than mine, and I like heavily modded sandbox games, so my perf/$ measures are different than most . I could blow through 2GB VRAM in builds I made, just doing some playing for stress testing before giving it up. The semi-passive stock coolers on a few of the 970s, plus 4GB VRAM for a decent price, are both answers to my prayers.

Even better Maxwells could come out next year, but that should only be a problem for someone if (a) they wrap their self-esteem up in their e-peen, (b) can't make good use of what they bought (Pentium Anniversary + GTX 970 = :|), (c) made a poor performance upgrade choice (R9 290X -> GTX 980).