-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Nvidia Geforce 'Maxwell' Thread
- Thread starter Boomstick777
- Start date
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
I'm about 90% sure that TSMC isn't offering true 16nm, the top layer can get a bit denser because finfets will be a little smaller but the metal layer remains the same size and the distance between transistors remains the same, 64nm, same on glofo/tsmc/samsung/intel. Glofo 28nm is 80 or 86nm or something, forget which, Intel 22nm is 80nm. it's as low as you can get with single patterning. All companies are engaging double patterning on the new processes, and everyone is achieving 64nm with it.
Metal gate pitch gives a realistic "true" size in terms of die size. You can save marginal amounts with smaller 16nm transistors, but rather than being 20% smaller(20>16nm), it's really 20+64nm >16+64nm, or 84>80, less than 5% smaller. When you factor in that not everything on die is a finfet, the difference will be less than that. With different versions it's usually a set of design rules and some slightly different transistor shapes, sometimes material, where the chip copes with higher switching speeds, or offers less leakage.
low power isn't really anything more than a term that sounds good when 99% of the industry made chips are looking for lower power usage.
10nm in 2015 is simply not going to happen in the way people think, if we see chips before 2017 I'd be very surprised.
Should be a fairly large industry push for 10nm as it seems 450mm wafers were pushed back from this gen to next gen... hopefully next gen. 450mm won't help us get better chips, it will just help offset the cost increases from things like double patterning. IN a given amount of floor space and equipment, every process node that adds extra stages adds time to production, in weeks rather than hours. The way to offset wafers taking say twice as long to make, is to get twice as many chips per wafer using bigger wafers.
450mm should help bring costs back down again. Not sure if they are still planning on EUV for 10nm, it keeps slipping back, or if they are wanting to go triple patterning for 10nm. Haven't really read a lot about 10nm which suggests it's no where near as close as they are saying.
Looking up news reports, more technical ones saying TSMC are saying Q4 2015 for risk production with volume production not expected before Q2 2017. Those are decent enough targets but they'll slip.
Keep in mind the same roadmap that shows 10nm in Q4 2015, showed the same level reached(risk production) for 20nm at the beginning of 2013, and 16nm in Q4 2013(Nov) also.
At this point we are around 18 months on from risk production of 20nm and only one shipped chip I'm aware of, incredibly low volume mining asics. Real 16nm finfet chips not expected till probably mid 2015, again 18 months or so on from initial risk production.
Foundries don't talk about roadmaps in terms of when WE get chips, but when the first non working piece of crap wafer comes out of the oven but with something measured at 10nm on it... then they claim a win
Metal gate pitch gives a realistic "true" size in terms of die size. You can save marginal amounts with smaller 16nm transistors, but rather than being 20% smaller(20>16nm), it's really 20+64nm >16+64nm, or 84>80, less than 5% smaller. When you factor in that not everything on die is a finfet, the difference will be less than that. With different versions it's usually a set of design rules and some slightly different transistor shapes, sometimes material, where the chip copes with higher switching speeds, or offers less leakage.
low power isn't really anything more than a term that sounds good when 99% of the industry made chips are looking for lower power usage.
10nm in 2015 is simply not going to happen in the way people think, if we see chips before 2017 I'd be very surprised.
Should be a fairly large industry push for 10nm as it seems 450mm wafers were pushed back from this gen to next gen... hopefully next gen. 450mm won't help us get better chips, it will just help offset the cost increases from things like double patterning. IN a given amount of floor space and equipment, every process node that adds extra stages adds time to production, in weeks rather than hours. The way to offset wafers taking say twice as long to make, is to get twice as many chips per wafer using bigger wafers.
450mm should help bring costs back down again. Not sure if they are still planning on EUV for 10nm, it keeps slipping back, or if they are wanting to go triple patterning for 10nm. Haven't really read a lot about 10nm which suggests it's no where near as close as they are saying.
Looking up news reports, more technical ones saying TSMC are saying Q4 2015 for risk production with volume production not expected before Q2 2017. Those are decent enough targets but they'll slip.
Keep in mind the same roadmap that shows 10nm in Q4 2015, showed the same level reached(risk production) for 20nm at the beginning of 2013, and 16nm in Q4 2013(Nov) also.
At this point we are around 18 months on from risk production of 20nm and only one shipped chip I'm aware of, incredibly low volume mining asics. Real 16nm finfet chips not expected till probably mid 2015, again 18 months or so on from initial risk production.
Foundries don't talk about roadmaps in terms of when WE get chips, but when the first non working piece of crap wafer comes out of the oven but with something measured at 10nm on it... then they claim a win
Last edited:
Ok so if the rumours are true, and the info we have certainly seem to indicate that they are. This manifest shows the GM204 on APR 10, six months latter would put availability in October.
Then this manifest shows the GM200, six months latter would put availability in January.
Now obviously it might take more than six months, but it is my understanding that, that is the sort of time frame it takes to bring these things to market.
But it definitely looks like GM204 will be with us before Christmas with the GM200 cards shortly after.
Then this manifest shows the GM200, six months latter would put availability in January.
Now obviously it might take more than six months, but it is my understanding that, that is the sort of time frame it takes to bring these things to market.
But it definitely looks like GM204 will be with us before Christmas with the GM200 cards shortly after.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
It depends on yields really, a successful tape out, then 5-6 months. Takes a while to test, bin the first wafers, check how you think they work, be happy with it, put in an order for wafers and stockpile a bunch of chips.
But the 480gtx taped out... then got respun, etc. If GM200 is huge, shipping a few back to Nvidia doesn't mean everything is good with it. Also depending on cost and yields you could be looking at another Gk200 situation. Trickle supply for professional, very very slow stock piling for a higher volume launch later on.
Wouldn't think there would be a problem with GM204, they've made a Maxwell chip already, know the process and it shouldn't be that big.
But the 480gtx taped out... then got respun, etc. If GM200 is huge, shipping a few back to Nvidia doesn't mean everything is good with it. Also depending on cost and yields you could be looking at another Gk200 situation. Trickle supply for professional, very very slow stock piling for a higher volume launch later on.
Wouldn't think there would be a problem with GM204, they've made a Maxwell chip already, know the process and it shouldn't be that big.
Soldato
abit more on it here http://www.techpowerup.com/202938/nvidia-moving-around-the-fabled-geforce-gtx-titan-ii.html
3000+ maxewell shaders 60mm² + and a 512bit bus, would be a utter beast read about. With a price tag to match,
Associate
im just hoping for something faster than a 690.
Thanks.
I might consider that but the stuff I'm into (flight sims) tend to play better on Nvidia and I'm going m-itx so it will need to really tempting.
I'm just hoping for a 770 price drop really.
Speculated GeForce GM200 GPU Details:
Confirmed: GeForce Maxwell architecture
Safe: planned for the enthusiast / professional segment (successor of the GK110 chips)
Pretty sure: 28nm manufacturing by TSMC
Pretty certain tape-out June / July 2014
Speculative: up to ~ 600mm ² die area
Speculative: DirectX 12 in hardware
Speculative: ~ 4000 shader units
Speculative: 512-bit DDR memory interface
Accepted: performance index to ~ 4000 shader units roughly ~ 750% to ~ 800% (level GeForce GTX 780 Ti +40% to +50%)
http://wccftech.com/nvidias-flagship-geforce-maxwell-gm200-gpu-spotted-ready-consumers-year/
Confirmed: GeForce Maxwell architecture
Safe: planned for the enthusiast / professional segment (successor of the GK110 chips)
Pretty sure: 28nm manufacturing by TSMC
Pretty certain tape-out June / July 2014
Speculative: up to ~ 600mm ² die area
Speculative: DirectX 12 in hardware
Speculative: ~ 4000 shader units
Speculative: 512-bit DDR memory interface
Accepted: performance index to ~ 4000 shader units roughly ~ 750% to ~ 800% (level GeForce GTX 780 Ti +40% to +50%)
http://wccftech.com/nvidias-flagship-geforce-maxwell-gm200-gpu-spotted-ready-consumers-year/
Last edited:
R9 390X / GTX Titan II
Im not buying into those specs it is too much. Why would they do it anyways when people have pretty much accepted dual cards now. The big leaps went out the window when people paid £800 for a Titan for minimal improvements over a 780Ti.
If Nvidia learned anything from this it is that single cards now carry a huge premium if they are fast. And that there is no pressure anymore since people will go out in desperation and link two togeather. They might even link 3 up although until recently it was a fake frames per second which at times looked like 30fps.
SLI/CF = The Devil
If Nvidia learned anything from this it is that single cards now carry a huge premium if they are fast. And that there is no pressure anymore since people will go out in desperation and link two togeather. They might even link 3 up although until recently it was a fake frames per second which at times looked like 30fps.
SLI/CF = The Devil
Last edited:
Im not buying into those specs it is too much. Why would they do it anyways when people have pretty much accepted dual cards now. The big leaps went out the window when people paid £800 for a Titan for minimal improvements over a 780Ti.
If Nvidia learned anything from this it is that single cards now carry a huge premium if they are fast. And that there is no pressure anymore since people will go out in desperation and link two togeather. They might even link 3 up although until recently it was a fake frames per second which at times looked like 30fps.
SLI/CF = The Devil
I find it a bit strange when nVidia wasn't happy with 20nm for the GM200 let alone 28nm unless this isn't the original full Maxwell design at all.
SLI (and CF) depend a lot how you use it - I've generally used the 2nd card to top up an already good FPS so I can get i.e. as close to 120fps solid as possible which it works quite well for, but if your using 2-4 cards just to hit 30fps then its a waste of time as its quite ropey at those kind of levels. Its very much hit and miss though IMO but sometimes you get the right setup of upper mid-range or lower-high end where the price and performance ratios make sense to go 2 card SLI.
Im not buying into those specs it is too much. Why would they do it anyways when people have pretty much accepted dual cards now. The big leaps went out the window when people paid £800 for a Titan for minimal improvements over a 780Ti.
If Nvidia learned anything from this it is that single cards now carry a huge premium if they are fast. And that there is no pressure anymore since people will go out in desperation and link two togeather. They might even link 3 up although until recently it was a fake frames per second which at times looked like 30fps.
SLI/CF = The Devil
Whats wrong with sli/CF?
Most new games get profiles for it reasonable quick these days and if its a week or two later then release does it matter?
Hopefully i think both sides have realised selling cards as 4k ready(yes i know you can turn settings down and it does work..) and them not really being ideal single cardwise has given them a kick in the pants and the nextgen high end gpus from both sides will be a good 30-40% jump on what we have atm
Because it can let them slack off on what a single card can do. It is just a lazy excuse to lump two bits of silicon togeather and the power usage is mad for some models. Not to mention it aint as smooth as single cards and it is only recently they bothered to investigate this and do something. For years they just sold them and half the games never worked lol.
Imagine a world where it was never invented, And do you still think they would charge you £800 to play at 4K? Anyone remember having to pay the same back when 1080p came? You cannot say games are moving at a faster pace than back in those days either. The only thing that has changed is everything is slowing down and prices are going up.
Yup.
Imagine a world where it was never invented, And do you still think they would charge you £800 to play at 4K? Anyone remember having to pay the same back when 1080p came? You cannot say games are moving at a faster pace than back in those days either. The only thing that has changed is everything is slowing down and prices are going up.
Yup.
Associate
lol at the robocop meme. Would love to see what a 980 Ti could do if they follow the Kepler branding.
R9 390X / GTX Titan II
4000 GCN cores with the same memory interface/ROPs as Hawaii vs. 4000 maxwell cores with an upgraded memory bus and more ROPs (compared to GK110) wouldn't even be a competition unfortunately.
Caporegime
except Nvidia didn't scale as well as Hawaii in steeping up from previus gen (680 and 7970)
7970 to 290X = 40% more cores = 40% more performance = 100% scaling
680 to 780TI = 90% more cores = 45% more performance = 50% scaling
Edit. you also need to take this into account, while there maybe no truth in this, as there never is with any rumour, according to that Fiji is 20nm and has Memory stacking.
http://videocardz.com/50472/amd-launch-new-flagship-radeon-graphics-card-summer
Last edited:
Advise someone looks at varied benches, a 780TI is sometimes slightly less and sometimes same and sometimes faster than 2x670/680's.
Cannot load 2013 GPU's page with 2014 GPU's page so 690 vs 780Ti is near enough.
http://www.anandtech.com/bench/product/1072?vs=1184
Cannot load 2013 GPU's page with 2014 GPU's page so 690 vs 780Ti is near enough.
http://www.anandtech.com/bench/product/1072?vs=1184
Last edited: