Hi everyone
 
After seeing all them Radeon R9 threads I'm like it's time for a Maxwell topic, post any type of rumor you can find about the new Maxwell line of GPGPU's here, pics of Engineering Samples are also greatly welcome! And of course news updates too, klets hope that the GeForce GTX 800 series will also feature a Titan II or a new GX2 also GTX 890 or so?
 
Anyways I do have some news well it may be seen as a rumor as well, that NVIDIA plans to release Maxwell in Q1 of 2014, I am speculating this would be week 9 to 12 of 2014
As most Engineering Samples I have had lead to that the final ones were often from week 7 of a certain year and that the first production models appeared around week 09 to week 12 of that same year.
 
So here that news I found so far:
http://videocardz.com/45403/nvidia-to-launch-more-cards-this-year-maxwell-in-q1-2014
 
Here some people are trying to compare that the performance of a GTX 880 would be similar to that of fourHD 7970's in Quad CFX, but that is kind of a strange way of comparing I myself think it would runs games probably better since it's one big chip that does not have to rely on other chips working along side it. A bad comparison though:
http://www.rage3d.com/board/showthread.php?t=33994629
As for an i7 3930K being underpowered? I don't think so just OC it towards the 4.5 a 4.8Ghz and you should be okay, but hey that is just my speculations hehe
 
Happy posting and please do keep it civil here, so it's a pleasant read for everyone
 

Xavier Zepherious
And rumors are abound with GM100 - Maxwell
 
http://pastebin.com/jm93g3YG
 

got around some maxwell plans from nvidia. quite interesting stuff. there will be two lines, maybe the second one aligns with the finfet stuff, don't know. anyway, that's what i gathered:

the smx structure changes slightly. nv did some optimization that they can use now the dp alu also for sp, it supports now all sp instructions and can be used in parallel. it means an smx looks now to have 256 alu. technically, that reduces maxwell's dp rate to 1:4, but in reality it just boosts the sp performance in comparison to kepler. nv found out how to gate of the unused parts of the dp alu to keep the power down when doing sp stuff.

but the real changes are in the cache area. that will boost the efficiency big time.

 

first off, the registers are doubled per smx. more threads using a lot of registers can now run in parallel and better hide the latencies. and the caches got increased as well. the L1 cache also used as shared memory is now 128kb (doubled) and can be split between cache and shared memory in steps of 32/96, 64/64, or 96/32. maxwell keeps the 16 tmus per smx.

the gpcs consist of usually 3 smx, but got changed quite a bit. there is still that geometry engine and stuff, but each gpc now includes 768kb of l2 cache, backing the r/w-L1 as well as the read only texture L1 in the smxs and also serve as instruction cache for the smx. all this gets topped off with a much larger l3 cache than in kepler. now to some numbers for the first line.

gm100:

 

8 gpc (8 triangles per clock), 24 smx, 384 tmus, 6144 alu, 8mb l3 (and there are also 8 l2s in the gpcs!), 64 rops, 512 bit interface, up to 8 gb @ 6+ ghz

 

target frequency for gf 930mhz, boost 1GHz

 

target frequency for tesla 850mhz, gives 2.61 dp tflops, double that of kepler, comes with 16gb

gm104:

 

5 gpc, 15 smx, 240 tmu, 3840 alu, 4mb l3, 40 rops, 320 bit interface (7 ghz), 2.5gb for cheap models, probably a lot of asymmetric 3gb or (symmetric again) 5gb models, target 1+ ghz, can do dp only with 1:16 rate

gm106:

 

3 gpc, 9 smx, 144 tmu, 2304 alu, 4mb l3, 24 rops, 192 bit interface, 7ghz, 3gb ram

gm108:

 

2 gpc, 4 smx, 64 tmu, 1024 alu, 2mb l3, 16 rops, 128bit interface, 2 gb ram

but really interesting gets the refresh, probably waiting for tsmc's finfets. then 64 bit arm cores developed by nv gets integrated on the same die. they can coherently access the common l3 cache. the big thing is that they will be used by the graphics driver to offload some heavy lifting from the system cpu. basically most part of the driver will be running on the gpu itself! nvidia expects this will give them at least the same speed up as amd will get from mantle, but without using a new api with straight dx11 or opengl code! and it will also help with the new cuda version for maxwell, where one can access both gpu as well as cpu cores seamlessly.

the specs are planned to stay almost the same for gm110/114/116, just the 110 gets full 8 ARM v8 cores and a doubled l3 (16mb!) compared to the gm100. the finfets may also allow a further speed boost. the 8 arm core version is actually called gm110soc, so maybe nv will start to market them as standalone processors for hpc. the consumer version is likely cut down to 4 arm cores, the same as gm114 will get (which also gets a doubled l3 to 8mb). the gm116 will only get 2 cpu cores on die, i have not seen that a gm118 got mentioned.

 

 
Hmm Pastebin isn't that reliable as it can always be edited like Wikipedia can, correct me if I am mistaken here though :)
As for that speculation, well it's pretty decent but still very overdone, this is what I think Maxwell would be like:
 
GM100:
4096 Maxwell CUDA Cores
64 ROPS
256 TMU's
4096MB/8192MB 512Bit GDDR5
SM 5.2
DX 11.2
OGL 5.xx
CUDA 6.xx
NV Glide x64? [Possible Competitive API versus AMD Mantle]
 
GM104:
2560 Maxwell CUDA Cores
48 ROPS
192 TMU's
2048MB/4096MB 256Bit GDDR5
SM 5.2
DX 11.2
OGL 5.xx
CUDA 6.xx
NV Glide x64? [Possible Competitive API versus AMD Mantle]
 
GM106:
1536 Maxwell CUDA Cores
32 ROPS
128 TMU's
2048MB/4096MB 256Bit GDDR5
SM 5.2
DX 11.2
OGL 5.xx
CUDA 6.xx
NV Glide x64? [Possible Competitive API versus AMD Mantle]
 
GM108:
1024 Maxwell CUDA Cores
16 ROPS
64TMU's
1024MB/2048MB 128Bit GDDR5
SM 5.2
DX 11.2
OGL 5.xx
CUDA 6.xx
NV Glide x64? [Possible Competitive API versus AMD Mantle]
 
That is what I am thinking, I could be way off course, but I could be close too :)
All worth speculating I'm like!

Anything over $600 for a top (over 256bit) end GPU is a insult.
 
 

Early on, I found it difficult to accept that nVidia actually changes their architectures since they increase number of CUDA cores with every shrink and give it a new name, and almost like pacifiers, everyone plays cheerleader.  Therefore I concluded Maxwell = 20nm Kepler thinking they'll just do that again.  But if the rumors are true that Maxwell will be brought up before 20nm lithography comes full swing, well, I'm confused.  That doesn't mean nVidia won't 'do what works' and just add CUDA cores, I'm sure they can and will.  
 
I highly doubt they have been sitting on their hands waiting for AMD, as if they have been caught surprised, that is just stupid.  These companies know more about each other than any rumor we find online.  I'm sure nVidia has found ways of refining the same dies such that GK180 will be rebranded as GM104, and GM110 will arrive when the 20nm shrink comes.
 
As for price guessing, I don't care; either I can afford it or can't.  I'll just gleefully smirk at those that complain
 

Bruno747
GM100 $1200 in geforce form....maxwell titan
GM104 $750
GM106 $450
GM108 $325
Given the huge upswing in gpu prices over the last 3 years, this should be right in line.

It isn't that your estimates are wild, it looks like you didn't put any thought into it.
 

• GK104 debuted as GTX680 at $499 MSRP, GTX670 at $399 MSRP, and GTX660 Ti at $299 MSRP; 
• GK106 debuted as GTX660 at $229 MSRP, GTX650 Ti Boost at $169, GTX650 Ti at $149; 
• GK107 debuted as GTX650 at $109 and GT640 at $99.

 
Kepler refresh hasn't touched the bottom two which actually carry the bulk of nVidia's graphics revenue.  Don't just wildly-estimate those prices since people that buy those GPUs are penny-pinchers, they avoid spending unless they have to and tend to keep for many years, I can say most still own Fermi or older.  You can't win trying to overcharge them, they aren't like enthusiasts that will eventually settle if they want to keep up-- they don't think like that.
 
That said, now that AMD is back in the game with their R series, we have less reason to feel Maxwell prices will be higher than average.  If AMD still hasn't come around, I'd suspect Maxwell would only come to succeed high-end while GK106/107 remain unchallenged as they have thus far.

hehe Well I speculated the GM 100 for the GTX 880 as for a Titan II I kind of left that one out and that the GM110 would be used for the Titan II, this chip would be the GM100 done right and in a refreshed format, the GM110 would be used for the Titan II & GTX 990 if that would ever take place, as for GM114, GM116 & GM118 they too will play their role as GM104, 106 & 108 would be doing as mentioned in my first post
 
Sorry for leaving out the Titan II as I didn't not have a good speculation of what this could be so due to that I left it out.
Titan II Could get a layout like this:
 
Full Spec GM110 GPGPU
4096 Maxwell CUDA Cores
128 ROPS
512 TMU's
8192MB 512Bit GGDR5 or XDR2
SM 5.2
DX 11.2
OGL 5.xx
CUDA 6.xx
NV Glide x64? [Possible Competitive API versus AMD Mantle]
 
As for it's CUDA cores, they are the same amount because all what the GM110 is is a refresh of the larger  in Die Size GM100. To get this speculation, I took a look at the GF100 & GF110 in perspectives and in basics they have both the same amount of Fermi CUDA Cores each and show very similar similarities, even that production models of the GF100 were cut to 480 Fermi CUDA Cores, they were designed to have 512 none the less.
 
As for Kepler this is a different subject since there was no GK100, but there was a GK110, so kind of an odd setup there, It could be that GK104 was once named as GK100, but as that goes I could rely on their specs to how NV would do their GM110 compared to the GM100.
 
Now as for that GK180 the GK110-400-A1 refresh that was indeed a very eye catching and interesting news topic hmm, but would it be useful to make a Titan 1.5 or 2.0 with that Chip but then with the full 2880 Kepler CUDA Cores & 8192MB 512Bit GDDR5? I dunno how much effect this VGA could have if it were made and put up against the 580 Euro USD R9 290X which is about 10% faster than the regular Titan with beta drivers.. NV would be forced to sell this new Titan at the same price range as the R9 290X, otherwise it would be wasted money people would spending on something that performance slightly faster than a 580 euro R9 290X... Crazy right?

lehpron
 

Bruno747
GM100 $1200 in geforce form....maxwell titan
GM104 $750
GM106 $450
GM108 $325
Given the huge upswing in gpu prices over the last 3 years, this should be right in line.

 
It isn't that your estimates are wild, it looks like you didn't put any thought into it.
 

• GK104 debuted as GTX680 at $499 MSRP, GTX670 at $399 MSRP, and GTX660 Ti at $299 MSRP; 
• GK106 debuted as GTX660 at $229 MSRP, GTX650 Ti Boost at $169, GTX650 Ti at $149; 
• GK107 debuted as GTX650 at $109 and GT640 at $99.

 
Kepler refresh hasn't touched the bottom two which actually carry the bulk of nVidia's graphics revenue.  Don't just wildly-estimate those prices since people that buy those GPUs are penny-pinchers, they avoid spending unless they have to and tend to keep for many years, I can say most still own Fermi or older.  You can't win trying to overcharge them, they aren't like enthusiasts that will eventually settle if they want to keep up-- they don't think like that.
 
That said, now that AMD is back in the game with their R series, we have less reason to feel Maxwell prices will be higher than average.  If AMD still hasn't come around, I'd suspect Maxwell would only come to succeed high-end while GK106/107 remain unchallenged as they have thus far.

 
I may have put the wrong number of years in the earlier post....at any rate, the part that you didn't figure in was the jump from GF to GK chips, lets take your numbers and roll back another architecture.
 
GK110/GK100 VIA gtxtitan @ $1000 vs GF110/GF100 Via gtx480/580 @ $499   Increase amount-- 100%
 
GK104 via gtx680 (actual gtx660 until last minute sticker and price upgrade) @ $499 vs GF104 via gtx460 @ $229  Increase amount-- 118%
 
GK106 via gtx660 @ 229 vs GF106 GTS450 @ $129  Increase amount-- 77%
 
So given the jump from fermi to kepler, provided nvidia plans on taking the bend over here it comes pricing strategy again (which is very likely), I would say those guesses on Maxwell would be if anything just slightly on the high side for the lowest end chips. Given the number of people that fell all over them selves to buy a titan or 690 or two or three, $1200 doesn't seem to be too far out of nvidia's capable money grabbing hands. Combine that with amd's recent release of Hawaii, this is shaping up to be another 7970/680 fiasco all over again.
 
Nvidia will either stick with the pricing they have established and fans so graciously accepted or go higher until they alienate enough to notice a decent dip in sales.

What still baffles my mind most, is why NV never made a GK100 and went right on with GK104 and then later on with GK110 but never made a GK114 :X Talk about oddities there  You'd expect GK100 was probably made as an Engineering Sample right?
Makes me wonder if NV might drop the GM100 and start off with GM104 like they did with Kepler's GTX 680 the GK104.
 
Just thought there

Wait, a $1K plus GPU isnt gold plated?

Bruno747
GK110/GK100 VIA gtxtitan @ $1000 vs GF110/GF100 Via gtx480/580 @ $499   Increase amount-- 100%
GK104 via gtx680 (actual gtx660 until last minute sticker and price upgrade) @ $499 vs GF104 via gtx460 @ $229  Increase amount-- 118% 
GK106 via gtx660 @ 229 vs GF106 GTS450 @ $129  Increase amount-- 77%

Each GPU codename doesn't directly relate to another GPU codename, as shown in my previous reply, the span is huge such that it might not seem obvious which product is the direct succession.  It generally goes by price, i.e. GTS450/GTX550 Ti was not succeeded by GTX660, rather GTX650 Ti with no price percentage increase since like I said, customers in this price range are picky and nVidia already knows that:

Anandtech
The GTX 650 Ti is the aptly named successor to 2011’s GeForce GTX 550 Ti, and will occupy the same $150 price point that the GTX 550 Ti launched into.

Same deal with GK104 versus GF104, they aren't directly relational and cater to entirely different customers; we can speculate forever on the meaning of codenames and what could have been, but in the end it is about price.  No one with a GF104 based GTX460 was meant to upgrade to a GK104-based GTX680, the price pretty much set those customers apart from each other forcing GTX460 folks to wait for GTX660 which is GK106-based as the price change made more sense.  GTX460 1GB debuted at $229-- same as GTX660-- they were aimed at the same types of customers with zero-percentage increase.
  
Like I said before, it is like you haven't put much thought into this.  Furthermore, I get the impression you're straight up ignoring the influence or impact of AMD for you to see the 100% increase from GTX680 to Titan as some kind of trend.  AMD wasn't there to keep them honest; GTX580 on back wasn't outrageous, because AMD was around, don't ignore that.
 
I think the question now is what will happen to the current Titan once R9-290X finally arrives, I don't think it will dip in price; I think nVidia will EOL it and replace it something better, i.e. GeForce GTX Altas or something.

How accurate do people think the "Q1 2014" release timeframe is?

Gold Leader
What ever happens to Titan there will be that day that I will get one for my collection, this idea and design is something too great top miss. Even that trying to obtain a GK100 Engineering Sample or that of the GK1190 would be an even greater challenge, as I did succeed with them GF100 Engineering Samples even it took me three years and seven months to find them, just never give up the hope of finding one or more
 
Although did GK100 also have 2880 Kepler CUDA Cores like GK110 has? Was there a similarity as with GF100 & GF110 as CUDA Cores go and Die Shrink? That being Same amount of CUDA Cores ut a an improved silicon & Die shrink?
 

 
I would have to do some research again, but if I remember right, the general consensus seemed to be that GK100 was planned to be 2880 and 512bit but I think they toned it down in gk110 to 384bit to eliminate some of the defects. A working gk100 with 512bit even if not all cores were functional would be a sweet collectors item.

Although not a rumor, I was wondering what you guys think would be the design choice for a GTX 890? Dual GPU and dual ARM chips? Or Dual GPU and a single ARM chip? Your thoughts are appreciated :D

Zephmeister
Although not a rumor, I was wondering what you guys think would be the design choice for a GTX 890? Dual GPU and dual ARM chips? Or Dual GPU and a single ARM chip? Your thoughts are appreciated :D

I'm pretty sure the secondary processor is to help facilitate the information to all the CUDA cores, like a co-processor, thus maybe only needing one per card regardless of how many GPUs are onboard.  
 
 
I'd make the analogy of the change from CPU-physics processing to the introduction of the PPU early on, but as technology progressed, the GPU was powerful enough to replicate the PPU.  So in all probability, the use of ARM processors in Maxwell is temporary until a future GPU architecture is powerful enough to take care of itself, whereas now Kepler is having a hard time, hence Maxwell to the rescue.

 
This whole performance per watt comparison chart NVidia has produced is worrisome to me.  If you're more efficient per-unit, but give me less cores, it's a net wash.  Give me a ridiculous chip at cheap prices.  Don't give me a bumped up cell phone chip. 
 
The focus on mobile feels like holding everyone back.  The idea that the mobile chips need to scale up to desktop parts - what is this, Windows 8 hardware design??

Everything I'm seeing says TSMC is ramping up for mass production on 20nm early next year, so I'm not convinced any Q1 Maxwell would be on 28.

starsmine

Drazhar
Everything I'm seeing says TSMC is ramping up for mass production on 20nm early next year, so I'm not convinced any Q1 Maxwell would be on 28.

Ramping, not mass producing.
The very earliest we will see 20nm cards would be in q2. 

I said ramping up for mass production.
 
http://www.xbitlabs.com/news/other/display/20131022230815_TSMC_Shares_More_Details_Regarding_16nm_FinFET_and_20nm_Progress.html
http://www.taipeitimes.com/News/biz/archives/2013/12/13/2003578882
http://www.digitimes.com/news/a20130926PD205.html
 
 
Now admittedly I'm no guru, but given the number of things that say "volume production" and short time frames, I'm not seeing anything to really indicate they couldn't produce 20nm until Q2, all I'm really seeing is that NV isn't happy that 20nm won't provide any cost decrease over 28nm.
 
and apparently we have this news regarding mobile Maxwell
 
http://wccftech.com/nvidia-maxwell-geforce-800m-mobility-lineup-launches-february-2014-geforce-gtx-880m-8-gb-lead-pack/