Nvidia Geforce 'Maxwell' Thread

[Boomstick777]

The GM200 is a different beast altogether. There's no mention of which process the chip will be based on, but it will succeed the GK110, and should offer performance increments worthy of being a successor. For that, it has to be based on the 20 nm process. It will tape-out in June 2014, and products based on it will launch only in or after Q2 2015.

So it looks like the first proper next gen card will be in Q2 2015, this could likely be a Titan successor and be the fastest card until Q4 2015 with the high end 20nm refresh of the earlier Maxwell cards.

Anyone else think this is looking like another repeat of the last gen. I.e mid-tier cards coming first, guess we have to get used to paying big money to get the best tech early in the form of the Titan..

Might go RED this next round of cards.. I hope they have something good coming..

 

[drunkenmaster]

Not really sure what you're getting at. If GM200 is in 20nm then the refresh of earlier 28nm maxwell cards won't be faster.

By name then a GM204 would be a 300mm^2 part at 20nm or maybe 500mm^2 at 28nm, all that would happen is that redone at 20nm it would be shrunk, not noticeably faster.

I would guess GM204/206 would be the equivalent parts at 20nm as the name suggests(a 100 part being around the 500mm^2 mark, a 204 being around the 300mm^2 mark and a 206 being around the 150-200mm^2 mark in size) simply done larger and using more power at 28nm. Done again at 20nm they would merely get smaller not faster. If GM200 is a 500mm^2 part at 20nm then it will, like Gk110, be generally the fastest part they can do on 20nm.


The real question is, will GM200 be 20 or 16nm. 20nm has been in production for some time and no one has rushed to get it out. AMD/Nvidia could likely have 20nm gpu's out now but the wafer cost and lower yield and poor power reduction would make them horribly expensive. When talking about mid 2015, you're talking about a full year after 20nm gpu's could be out, and 16nm is about a year behind 20nm.

I said for quite some time that 20nm could be skipped. for a 100mm^2 mobile part with higher yields and where power is the absolute limit mobile parts can still take large advances with 20nm. GPU's are more limited by reasonable power draw of a single gpu, 20nm giving a small ultimate reduction means twice the transistor count without a large power drop makes a pretty useless device.

Mid 2015 makes 16nm a real possibility, and makes a lot more sense for a high end part and financially. Why spend millions taping out a 20nm part when 16nm is starting production and offers huge advantages?

 

[andybird123]

Boom was just commenting on the content within the article using its own internal logic, if you want to have a go at someone about the 20/16 differentiation then go poke the author of the article

 

[Boomstick777]

Boom was just commenting on the content within the article using its own internal logic, if you want to have a go at someone about the 20/16 differentiation then go poke the author of the article

Ty, I'm not sure how that comment could have been misunderstood tbh?

** edited **

 

[RetrogamerX]

So it looks like the first proper next gen card will be in Q2 2015, this could likely be a Titan successor and be the fastest card until Q4 2015 with the high end 20nm refresh of the earlier Maxwell cards.

GTX880 and GTX880Ti will just be more power efficient on 28nm.

Not worth a upgrade if you have 780 or above

What I get from all this is It'll be late 2015, but what they say is always 4 to 6 months later so early 2016 before we see any true successor GK110.

 

[Sampsy]

It really sounds like Nvidia would be much better off dumping Global Foundries / TSMC and going into manufacturing themselves. This isn't the first time they've been screwed by their supplier's incompetence.

 

[N19h7m4r3]

It really sounds like Nvidia would be much better off dumping Global Foundries / TSMC and going into manufacturing themselves. This isn't the first time they've been screwed by their supplier's incompetence.

It's been happening a lot lately it seems.

Although would they have the ability to get such a large fab up and running quickly? Would the investment pay off for them?

 

[drunkenmaster]

Boom was just commenting on the content within the article using its own internal logic, if you want to have a go at someone about the 20/16 differentiation then go poke the author of the article

** Comment removed ** Boom VERY clearly suggested that the GM200 would be the fastest GPU UNTIL refreshes of the 28nm maxwell's. A mid range GM204 on ANY process is still a midrange card, and will not be faster than a GM200 high end card.

This was not based on the article or it's own internal logic, it's wrong, simply put.

I THEN spoke about the potential for them being 16nm rather than 20nm, so go do one. I didn't remotely bring up Boom's differentiation of 16nm, nor the article's author doing the same.


*** comment removed ***

 

[drunkenmaster]

It really sounds like Nvidia would be much better off dumping Global Foundries / TSMC and going into manufacturing themselves. This isn't the first time they've been screwed by their supplier's incompetence.

It's been happening a lot lately it seems.

Although would they have the ability to get such a large fab up and running quickly? Would the investment pay off for them?

nvidia don't have the money, engineering staff, or capability to get into manufacturing. They also don't use Global Foundries in any way and never have, likely never will but they might.

They have never been screwed by their suppliers incompetence, ever. At every stage when Nvidia has blamed TSMC ALL of their competitors have been able to make working chips without problems. Nvidia pass the buck for blame, citing yield problems at TSMC when AMD had working chips out months before and multiple other customers. When all customers bar one has no problems with the process and that one customer who has had trouble for 3 generations in a row blames TSMC..... lol.

For the record, the touted price of upcoming(delayed but still coming) 450mm wafer fabs is in the 14 billion range, and one small customer can absolutely not come close to making that financially viable. That is JUST the production and equipment, R&D will be a minimum of another several billion. Nvidia have exactly zero experience in an industry in which extremely experienced people are selling off their manufacturing capabilities because of the exponentially increasing costs and complexity. IBM have more knowledge than most, bigger volume than Nvidia, but small enough volume that they want to sell off their fabs, AMD did so, again with larger volume than Nvidia and several other manufacturers have also.

If they started R&D for a future process and then planned a new fab and built it, you're talking about them not having the slightest chance of being up and running within 6 years. GloFo trying to get into the business have spent pushing 20billion now and it will take a LOT more investment two more fabs planned at probably 15billion + a piece for New York and a potentially a new fab in the mid east. With billions in investment to the fabs in Singapore and Germany. It's a business where the small producer(several times larger than Nvidia's volume) can't compete and can't turn anything but huge losses.

Intel is struggling has cancelled one fab, closing another and has low utilisation and they have many many times the volume that AMD/Nvidia do, they are struggling with the costs/production difficulties of having too few customers. Silicon production is already being dominated and will likely be completely dominated by those who can spread the cost of R&D and equipment amongst hundreds of thousands of customers over a period of 5-10 years per process.

nvidia, AMD, Intel, they need the latest and greatest process and have no business for old process/equipments. TSMC has a huge number of fabs. They'll spend 15billion on a 10nm 450mm fab but 10 years later that equipement will be churning out calculator chips by the millions at tiny costs, but effectively extending the life of a process by 4-5 times longer than someone like Nvidia/AMD/Intel would use that process. It's the only way to really succeed in the industry.

 

[Boomstick777]

What I get from all this is It'll be late 2015

^^ This.

We could see a big Titan GK110/GMXXX like card and price in early 2015, with the rest of the line getting the big GMXXX core at the end of the year Q4 2015. A repeat of the GTX 680 > Titan in early part of the year and then later 780 > 780 Ti in Q4. Nvidia made a ton of money this way, so very likely to want to repeat it lol.

 

[andybird123]

.... Boom VERY clearly suggested that the GM200 would be the fastest GPU UNTIL refreshes of the 28nm maxwell's. A mid range GM204 on ANY process is still a midrange card, and will not be faster than a GM200 high end card.

This was not based on the article or it's own internal logic, it's wrong, simply put.

what? no he didn't

the article says "The GM200 is a different beast altogether. it will succeed the GK110. For that, it has to be based on the 20 nm process."

HE then said "So it looks like the first proper next gen card will be in Q2 2015, this could likely be a Titan successor and be the fastest card until Q4 2015 with the high end 20nm refresh of the earlier Maxwell cards."

Meaning that mid range 28nm cards will come first, then GM200, THEN 20nm mid range cards AND more high end cards, he never suggested that a 20nm GM204 would be faster than a 20nm GM200
earlier Maxwell cards, meaning all of them, not JUST the GM204

as you well know, as a process improves it allows improvements to the chips, Boom was saying that the first GM200 could be Titan branded and then in Q4-2015 we could be looking at 980ti type of thing, similar to how we had 680 > Titan > 780ti in terms of top tier performance cards

Insert 16nm anywhere it says 20nm, as the article only mentioned 20nm, i think it's fairly safe to assume that people are talikng about 20nm and/or its derivatives

 

[Boomstick777]

GeForce GTX 880 ES Intercepted En Route Testing Lab, Features 8 GB Memory?

GeForce GTX 880 ES Intercepted En Route Testing Lab, Features 8 GB Memory?

An engineering sample (ES) of the GeForce GTX 880 was intercepted on its way from a factory in China, to NVIDIA's development center in India, where it will probably undergo testing and further development. The shipping manifest of the courier ferrying NVIDIA's precious package was sniffed out by the Chinese press. NVIDIA was rather descriptive about the ES, in its shipping declaration. Buzzwords include "GM204" and "8 GB GDDR5," hinting at what could two of the most important items on its specs sheet.

http://www.techpowerup.com/200505/g...n-route-testing-lab-features-8-gb-memory.html

 

[Yaayuh!]

On way to test lab?
Shows they're testing an 8GB card at least I suppose.

 

[PhysicsMan94]

Interesting, 256-bit bus though? Hmmm :/

 

[Silent_Scone]

Interesting, 256-bit bus though? Hmmm :/

That part I can almost assure you is ********

 

[PhysicsMan94]

That part I can almost assure you is ********

Indeed, that's what I'm thinking, it wouldn't make much sense for Nvidia to release the 880 with such a small memory bus.

 

[drunkenmaster]

Yes it would, it's a GM204 It's got to be cost and power effective. Higher memory bus = more power, higher memory bus = bigger die.

Memory speeds can increase(somewhat counter productive, or can be). The 680gtx had 256bit mem bus. Nvidia could only fit a 384bit bus on a core that was over 500mm^2, where do you suppose the magic die space comes from to fit >256bit bus on a core that will likely be <400mm^2?

The answer is, it won't, midrange is midrange. I seem to recall a 580gtx had a 384bit bus... yet a 680gtx had a.... 256bit bus. One was >500mm^2 one was near enough 300mm^2.

We can pretty much see that 880gtx = upper midrange sold at high end for a while before being replaced by the real high end. So 980gtx = 20nm 780gtx replacement. 880gtx = 680/770gtx replacement. GM204 will either be rebranded to 970gtx later on or possibly respun to 20nm and called a 970gtx.

Real question is maybe die size, if it's 400-500mm^2 then it's probably intended to be shrunk into a 250mm^2 20nm part, if it's 300-400mm^2 range then I wouldn't expect to see it at 20nm, it would frankly be too small, profitable and Nvidia could do that, but we might see a new part slotted between them.

I wouldn't bet against, GM204 @28nm, GM200 @20nm 500mm^2 512bit bus, then later on 20nm a GM202(or a new 204) that is 384bit bus, more cores and to better sit around the 300mm^2 mark.

 

[Rofflay]

If it is midrange why put on 8GB? 8GB is not needed by even 4K and would offer no benefit so it would be an expensive card no doubt. The 290X 8GB is £599 so Nvidia prices on that would be £700?


I would like to think they will at least do something to offset the delay and offer lower power for 2014 around the £500 mark. I would take 20% more than a stock 780 Ti with much lower power and 4GB of memory no problem as long as we get HDMI 2.0 and HDCP 2.2 and a few displayports would be nice as well. Hell it would be nice to see Maxwell offering something like the all in wonder ATI cards with six displayports on one card or something nuts.


Nvidia have a few options they could use i just wish they would announce something soon and my other wish would be that people stop buying 780 cards to force thier hand.

 

[drunkenmaster]

memory amount doesn't matter, got to go bigger with each new gen, marketing buzzwords. Don't forget that Nvidia and AMD(though Nvidia started it) rebrand the SAME cards frequently just so Dell and the rest can sell "new" computers with better graphics in them.

In terms of compute the more memory the better. In terms of stacked mem, I'm trying to find out what the minimum amount of memory you can get in a stack actually is because currently the HMB is a stack of memory on top of a organic substrate with a 128bit bandwidth(the memory can be MUCH more than that internally but the external connection is currently limited as they haven't got to interposers yet, interposer = silicon = 10 times thinner traces for massively more bandwidth, organic = much thicker, far lower bandwidth possible). Meaning if you can get a minimum of 2gb of mem a stack then to get 512bit bandwidth for a gpu you would need 4 stacks and 8gb minimum and that is a 28nm gpu that needs 512bit bus. If we get roughly speaking two times or maybe closer to 2.5x the cores on 14nm gpu's, 512bit might well not be enough meaning requiring even more memory.

Nvidia doesn't have stacked mem for 28nm and basically no chance till Pascal. But by Pascal they'll want 8-16gb being the range gpu's are using. AMD will be in the same boat. So it's likely we'll see 8gb being pushed at 20nm and 16gb on 14nm potentially.

Memory prices are coming down and density is increasing, on top of that volume for gddr5 will have increased with PS4 production, 8gb gddr5 per console means production volume is up which means price per unit will be down. One of the biggest reasons for not having 8-16gb of memory recently is really the number of chips and number of traces on a pcbm stacked mem(interposer or not) will effectively remove that limitation completely while also reducing power significantly. It's also worth remembering that all bandwidth is not the same, memory controller, the way the memory's own logic accesses it, everything can be improved. HMB over Gddr5 in current speeds is supposed to both be 35% faster and use 40% less power.

Anyway, in marketing terms it's no different to some low end POS card getting the same memory amount as a high end card. For compute, being able to store a lot more information locally means a lot less deleting and fetching across pci-e and ability to use larger data sets which for compute is becoming critical.

Like system memory where people were saying better off with 4 or 8gb of okay memory than 2gb of uber fast memory. Not having enough was significantly worse in every way than losing 5-10% in speed.

 

[Boomstick777]

Second generation Maxwell to support H.265

20nm parts later this year

We heard some rather interesting information for anyone who thinks that we need more than Full HD at 1920x1080 video. As you can imagine Nvidia is always thinking how to help the decoding process of the next generation codecs and since 4K video plays an important role in company's roadmap, it is not a big surprise to learn that the next generation Maxwell supports the H.265 codec.

Fudzilla so usual pinch of salt.. http://www.fudzilla.com/home/item/34697-second-generation-maxwell-to-support-h265