NVIDIA preparing four Maxwell GM204 SKUs (VideocardZ via S/A)

[ShintaiDK]

There're 60hz SST 4K monitors available right now on dp 1.2

But a single stream with DP1.2 doesnt have the bandwidth to drive a 4K screen at 60hz. The DP controller is only capable at 1920x2160 60hz on 1.2 per stream.

 

[3DVagabond]

SST means either 30hz on DP1.2 or DP1.3 support for 60hz. And since DP1.3 is nonexistant...

That was an nVidia problem. It was supposed to be fixed with the 337.88 drivers though.

 

[Leadbox]

But a single stream with DP1.2 doesnt have the bandwidth to drive a 4K screen at 60hz.

Pretty sure there's a samsung monitor and asus one using the same scaler that offers 60Hz at 4k.It is not a bandwidth problem that caused the MST approach to early 4K monitors, dp1.2 has enough bandwidth

 

[ShintaiDK]

DP1.2 doesnt have enough bandwidth in a single stream. 1920x2160 60hz is the maximum per stream or 4K@30Hz. Hence 2 streams is needed for 4K@60Hz.

 

[3DVagabond]

DP1.2 doesnt have enough bandwidth in a single stream. 1920x2160 60hz is the maximum per stream or 4K@30Hz. Hence 2 streams is needed for 4K@60Hz.

DP 1.2 has bandwidth for 4k @60 Hz with 10bit color.

 

[Leadbox]

DP1.2 doesnt have enough bandwidth in a single stream. 1920x2160 60hz is the maximum per stream or 4K@30Hz. Hence 2 streams is needed for 4K@60Hz.

This and this are 4K SST 60Hz monitors.

 

[mla]

GM200 would be the first processor with cuda 4k
http://videocardz.com/51032/nvidia-maxwell-gm200-spotted-new-gtx-titan-ii-core

 

[SirPauly]

Splitting the gen into 2 halves could very well be the new trend in GPUs as a result of more expensive die shrinks and longer time to market for lower nodes.

Or it can be a recipe for disaster -- if one plans on offering incremental and evolutionary increases and your competitors do not -- recipe for disaster!

 

[BrightCandle]

If AMD led with a big core and Nvidia leads with its smaller one then the performance difference will be pretty dramatic. Nvidia at that point could be a year away from releasing its big cores on a process that is well booked up. It could really swing the market, but its also risky. AMD's strategy so far has always been to appeal to the mid market and sell competitively there. The 290(X) is actually a marked difference in strategy from them. Could be interesting to see what they both do this year.

 

[Pinstripe]

Do you guys think there's any realistic chance for GM206 being sold for ~250€, has 3-4GB VRAM and performance levels around 15-20% faster than GTX 770?

It would be the perfect 1080p gamer card that is humanly affordable.

 

[RussianSensation]

Or it can be a recipe for disaster -- if one plans on offering incremental and evolutionary increases and your competitors do not -- recipe for disaster!

It wouldn't be a disaster for a lot of reasons:

1) At just 438mm2 and only 1Ghz clocks, Hawaii XT already uses more power than 780Ti. Unless AMD releases a 550-600mm2 die with standard liquid cooler, they themselves are limited by 28nm.

Even if AMD releases some miraculous 580mm2 die, Nv could raise clocks to 1.2Ghz on GM204, raise TDP to 225-250W and more than double 770's performance assuming 2X the performance/watt increase of Maxwell.

Finally, since they will have a smaller die, they could price it much lower than AMD with a 550-600mm2 die and liquid cooler.

2) We already have history of Fermi vs. Cypress that shows Nv can be late by 6-9 months and it hardly affects them. People even bought 285 instead of 5850 or waited for 6+ months to get 480. It would take 2-3 consecutive generations of NV being late by 6 months and to have AMD beat them in efficiency each of those 2-3 gens for this to really affect them. Also, it is not as if AMD went out of business when 780 beat 290 to launch by 5 months.

3) Even if somehow 390X beats 880Ti, Nv will continue to dominate mobile dGPU which is where most of the dGPu sales and growth is coming from.

In other words, Nv can easily launch a midrange GM204 first with little risk. It would be different if AMD had access to stackable VRAM and 16/20nm node much earlier than NV. But we have not evidence to suggest that. Also, we have real world proof that 750Ti offers 2X the performance/ watt validating the efficiency claims for Maxwell and nothing at all from AMD to suggest GCN 2.0 can somehow match this increase in efficiency. AMD barely improved the efficiency with 290X over 7970 compared to what NV did with 780Ti vs. 680.

Maybe I sound very pessimistic for AMD next round but I don't see them winning next gen, especially not in the mobile dGPu space. I think until AMD can get stackable VRAM and a lower node, they will compete on price/performance and this way even if GM200 wins by 20%, AMD could have 2 second best cards offer 80% of the performance for much less (I.e, 290s for $750-800 vs. 780Tis for $1,200-1,400).

 

[SirPauly]

I

2) We already have history of Fermi vs. Cypress that shows Nv can be late by 6-9 months and it hardly affects them. People even bought 285 instead of 5850 or waited for 6+ months to get 480. It would take 2-3 consecutive generations of NV being late by 6 months and to have AMD beat them in efficiency each of those 2-3 gens for this to really affect them. Also, it is not as if AMD went out of business when 780 beat 290 to launch by 5 months.

But it does effect -- do you think nVidia enjoyed losing over-all discrete leadership? Or having AMD so close on desktop share considering the investments for their nVidia ecosystem?

Do you know that there was one quarter were over-all revenue dropped off a cliff -- desktop revenue dropped 34 percent for nVidia?

 

[TrulyUncouth]

2) We already have history of Fermi vs. Cypress that shows Nv can be late by 6-9 months and it hardly affects them. People even bought 285 instead of 5850 or waited for 6+ months to get 480. It would take 2-3 consecutive generations of NV being late by 6 months and to have AMD beat them in efficiency each of those 2-3 gens for this to really affect them. Also, it is not as if AMD went out of business when 780 beat 290 to launch by 5 months.

Just my personal anecdote to give a different opinion on this. I used nothing but Nvidia graphics cards for the better part of a decade because the conventional wisdom was that AMD had poor drivers and support. When the 7970 came out I was excited for the performance but waited for the Nvidia release. Since they had supply issues and I was overdue for an upgrade I picked up a nice aftermarket 7950.

I now know from personal use that I have run into zero issues with AMD and so in the future I will be buying whoever has better perf/$. The only thing that even made me consider going NV next round if they don't win on perf/$ was the shield- but Steam has added streaming that works shockingly well, so that reason is gone and I am back to 100% about perf/$.

 

[raghu78]

It wouldn't be a disaster for a lot of reasons:

1) At just 438mm2 and only 1Ghz clocks, Hawaii XT already uses more power than 780Ti. Unless AMD releases a 550-600mm2 die with standard liquid cooler, they themselves are limited by 28nm.

Even if AMD releases some miraculous 580mm2 die, Nv could raise clocks to 1.2Ghz on GM204, raise TDP to 225-250W and more than double 770's performance assuming 2X the performance/watt increase of Maxwell.

Finally, since they will have a smaller die, they could price it much lower than AMD with a 550-600mm2 die and liquid cooler.

2) We already have history of Fermi vs. Cypress that shows Nv can be late by 6-9 months and it hardly affects them. People even bought 285 instead of 5850 or waited for 6+ months to get 480. It would take 2-3 consecutive generations of NV being late by 6 months and to have AMD beat them in efficiency each of those 2-3 gens for this to really affect them. Also, it is not as if AMD went out of business when 780 beat 290 to launch by 5 months.

3) Even if somehow 390X beats 880Ti, Nv will continue to dominate mobile dGPU which is where most of the dGPu sales and growth is coming from.

In other words, Nv can easily launch a midrange GM204 first with little risk. It would be different if AMD had access to stackable VRAM and 16/20nm node much earlier than NV. But we have not evidence to suggest that. Also, we have real world proof that 750Ti offers 2X the performance/ watt validating the efficiency claims for Maxwell and nothing at all from AMD to suggest GCN 2.0 can somehow match this increase in efficiency. AMD barely improved the efficiency with 290X over 7970 compared to what NV did with 780Ti vs. 680.

Maybe I sound very pessimistic for AMD next round but I don't see them winning next gen, especially not in the mobile dGPu space. I think until AMD can get stackable VRAM and a lower node, they will compete on price/performance and this way even if GM200 wins by 20%, AMD could have 2 second best cards offer 80% of the performance for much less (I.e, 290s for $750-800 vs. 780Tis for $1,200-1,400).

Its very likely that AMD will have HBM with their high end dGPUs and APUs. Nvidia will be atleast 12 - 18 month behind on HBM. It would be just like when AMD got to GDDR5 first with HD 4870 .

http://www.microarch.org/micro46/files/keynote1.pdf

http://wccftech.com/amd-feature-gen...0-graphics-cards-allegedly-launching-2h-2014/

http://pc.watch.impress.co.jp/docs/column/kaigai/20140428_646233.html

http://wccftech.com/amd-carrizo-apu-28nm-stacked-dram-alleges-italian-leak/

http://seekingalpha.com/article/230...hics-and-total-compute-performance-leadership

 

[tviceman]

Its very likely that AMD will have HBM with their high end dGPUs and APUs. Nvidia will be atleast 12 - 18 month behind on HBM. It would be just like when AMD got to GDDR5 first with HD 4870 .

http://www.microarch.org/micro46/files/keynote1.pdf

http://wccftech.com/amd-feature-gen...0-graphics-cards-allegedly-launching-2h-2014/

http://pc.watch.impress.co.jp/docs/column/kaigai/20140428_646233.html

http://wccftech.com/amd-carrizo-apu-28nm-stacked-dram-alleges-italian-leak/

http://seekingalpha.com/article/230...hics-and-total-compute-performance-leadership

What exactly do we know about HBM? How much will it help high-end discrete GPU's? Does HBM add bandwidth ontop of external GDDR ram or does it act as the first-access "buffer" before having to resort to GDDR5? Is HBM actually confirmed to be on high end Radeon GPU's this year or next year? I know Tonga rumors have been in and out, but is it confirmed to be on Tonga? What about higher end GPU's? To me, it sounds like an aggressive R&D roadmap to introduce a noticeable architecture update AND HBM at the same time. AMD laid off engineers during the past 24-36 months and R&D expenditure has dropped along with it. If AMD didn't make any significant changes to GCN's current iteration, will HBM be enough to offset Maxwell's current advantages?

Amplifying these questions, we already see that Maxwell is a very capable architecture, destroying GCN 1.1 at perf/watt and handily winning at perf/mm^2 all while doing so with less transistors per mm^2. We literally know nothing of what GCN 2.0 is going to bring. Nothing.

There are significantly question marks in the air regarding AMD's next generation architecture and there isn't anything other than rabid fan-made rumors to fan the flames.

 

[raghu78]

What exactly do we know about HBM? How much will it help high-end discrete GPU's? Does HBM add bandwidth ontop of external GDDR ram or does it act as the first-access "buffer" before having to resort to GDDR5? Is HBM actually confirmed to be on high end Radeon GPU's this year or next year? I know Tonga rumors have been in and out, but is it confirmed to be on Tonga? What about higher end GPU's? To me, it sounds like an aggressive R&D roadmap to introduce a noticeable architecture update AND HBM at the same time. AMD laid off engineers during the past 24-36 months and R&D expenditure has dropped along with it. If AMD didn't make any significant changes to GCN's current iteration, will HBM be enough to offset Maxwell's current advantages?

Amplifying these questions, we already see that Maxwell is a very capable architecture, destroying GCN 1.1 at perf/watt and handily winning at perf/mm^2 all while doing so with less transistors per mm^2. We literally know nothing of what GCN 2.0 is going to bring. Nothing.

There are significantly question marks in the air regarding AMD's next generation architecture and there isn't anything other than rabid fan-made rumors to fan the flames.

How did you conclude that AMD has not made significant improvements to GCN for improving efficiency on the R9 390X ? Just because you don't know yet does not mean AMD has not done anything. Wait till the final product arrives. Obviously AMD would not want to reveal information before actual product launch in early Q1 2015.

a 4 chip stack with 4 Hi for each stack means 512 GB/s bandwidth for the ultra high end R9 390X. Power consumption is cut significantly and that could go towards adding more SP or improving efficiency. Then come architectural improvements.

http://www.skhynix.com/inc/pdfDownl.../Databook/Databook_3Q'2014_GraphicsMemory.pdf

Hynix is already in production of HBM and they are AMD's partner for HBM production. As early as last Dec AMD and Hynix have started talking about next gen APUs and dGPUs using HBM. The HBM DRAM is being built now for integration into 2015 products.

Did you even read the links I posted. the seekingalpha article by me is about AMD's initial approach to HBM on APUs. It calls for a high capacity / high bandwidth L3 cache. The capacity is 1GB with a bandwidth of 128 GB/s for a single stack with 4 Hi config. APUs would find that more than sufficient. for high end dGPUs like R9 390X a 4 chip stack 4 Hi config brings 4GB capacity at 512 GB/s . More than enough bandwidth to power the AMD next gen flagship dGPU.

 

[tviceman]

How did you conclude that AMD has not made significant improvements to GCN for improving efficiency on the R9 390X ?

I don't know whether they have or have not, that's why I posed the question. As I said, AMD's engineering staff has declined since prior to GCN and their R&D expenditure has been in steady decline. Just over 4 years ago, AMD was spending over $420 million per quarter in R&D. Now it's spending under $280 million. Even just 2 years ago, R&D was still above $350 mil per quarter. That was during Hawaii and Bonaire's development. Going from $420 million/quarter to $280 million/quarter is a huge dip. In fact, about a year ago Nvidia surpassed AMD on R&D expenditure for the first time ever and has maintained a lead since. Those loses don't have zero impact. It's sounds aggressive to for me, given these conditions, that significant changes to GCN are inbound relatively simultaneously with merging their discrete GPU's with HBM.

If AMD engineers are doing more with less, then that is truly awesome and great for them. But rarely (ever???) has a tech company ever cut R&D and engineering staff while simultaneously improving their output. HBM isn't just slapped onto a chip and voila. The R&D to make HBM a reality, the R&D of tuning memory controllers, the possible unknown expense of potential yield issues, and combined with a potential architectural overhaul sounds like a pretty huge undertaking.

If it sounds too good to be true............................

 

[raghu78]

I don't know whether they have or have not, that's why I posed the question. As I said, AMD's engineering staff has declined since prior to GCN and their R&D expenditure has been in steady decline. Just over 4 years ago, AMD was spending over $420 million per quarter in R&D. Now it's spending under $280 million. Even just 2 years ago, R&D was still above $350 mil per quarter. That was during Hawaii and Bonaire's development. Going from $420 million/quarter to $280 million/quarter is a huge dip. In fact, about a year ago Nvidia surpassed AMD on R&D expenditure for the first time ever and has maintained a lead since. Those loses don't have zero impact. It's sounds aggressive to for me, given these conditions, that significant changes to GCN are inbound relatively simultaneously with merging their discrete GPU's with HBM.

If AMD engineers are doing more with less, then that is truly awesome and great for them. But rarely (ever???) has a tech company ever cut R&D and engineering staff while simultaneously improving their output. HBM isn't just slapped onto a chip and voila. The R&D to make HBM a reality, the R&D of tuning memory controllers, the possible unknown expense of potential yield issues, and combined with a potential architectural overhaul sounds like a pretty huge undertaking.

If it sounds too good to be true............................

The R&D for die stacking on a 2.5d silicon interposer has been done way back in 2010 - 2012. In fact AMD along with their parthers Amkor / Hynix showed a 2.5 silicon on interposer implementation at TFE 2011. This has been mentioned in my seekingalpha article. The HBM based L3 cache controller patent for APUs was filed on Jun 25,2012 and published on Dec 26,2013. AMD has been waiting for HBM DRAM to enter volume production and for die stacking to mature from a yields and cost point of view. And that time is now. Products using HBM and 2.5D stacked DRAM on silicon interposer will be out in 2015. AMD's high end APUs and dGPUs will use it. APUs for high capacity / high bandwidth L3 cache and dGPUs for graphics memory.

Also AMD's SOC design methodology focusses on IP re-use across various products. The development focusses on IP blocks or modules and not individual products. the product teams just assemble the various IP blocks into the final product they need. There is a lot of emphasis on being efficient in use of development resources.

 

[NostaSeronx]

Products using HBM and 2.5D stacked DRAM on silicon interposer will be out in 2015.

To go into detail of this;
2H 2014 for 20-nm LP, refresh GPUs.
Mid-2015 for 20-nm LP, APUs that are more efficient than 28-nm SHP APUs.
Late 2015 for 20-nm LP or 14-nm SOI FinFET, CPUs that are MASSIVE MM squares.

---
Nvidia is going 20-nm SOC for Maxwell and 14-nm FinFET Turbo for Pascal with 3 to 6 Hybrid Memory Cubes.

Expectation for Pascal is 960/1440 GB/s to 1920/2880 GB/s. <== Much better than Pirate Islands Peak of 1024 GB/s with HBM2.

That is right GIGABYTES per second, pretty close to TERABYTES per second.

 

[raghu78]

To go into detail of this;
2H 2014 for 20-nm LP, refresh GPUs.
Mid-2015 for 20-nm LP, APUs that are more efficient than 28-nm SHP APUs.
Late 2015 for 20-nm LP or 14-nm SOI FinFET, CPUs that are MASSIVE MM squares.

---
Nvidia is going 20-nm SOC for Maxwell and 14-nm FinFET Turbo for Pascal with 3 to 6 Hybrid Memory Cubes.

Expectation for Pascal is 960/1440 GB/s to 1920/2880 GB/s. <== Much better than Pirate Islands Peak of 1024 GB/s with HBM2.

That is right GIGABYTES per second, pretty close to TERABYTES per second.

http://www.anandtech.com/show/7900/nvidia-updates-gpu-roadmap-unveils-pascal-architecture-for-2016

"What&#8217;s new is that NVIDIA has confirmed they will be using JEDEC&#8217;s High Bandwidth Memory (HBM) standard, and the test vehicle Pascal card we have seen uses entirely on-package memory, so there isn&#8217;t a split memory design"

So both AMD and Nvidia will have the same 1024 GB/s or 1TB/s bandwidth for the 2016 flagship GPUs using HBM2. The VRAM capacity on the 2016 flagship GPUs will be a humongous 8GB /16 GB using a 4 stack config using 8 gigabit 4/8 Hi stacks at 256 GB/s per stack.

 

[BrightCandle]

That is an awesome increase, at least 2-3x the current high end if not a lot more. Presumably there will also be a massive decrease in latency as well which may allow them to reduce the amount of cache on board the GPU and instead put those transistors into compute cores. Stacked memory is really going to bring a lot of performance improvements, and for GPUs at least where we don't get to add more RAM its ideal.

 

[Cloudfire777]

So I`m guessing Nvidia have finished the engineering phase since TMSC shipped 35 GM204 die`s to Nvidia June 26th.

Thats the most they have ever shipped and Im guessing thats the finished design.
Also they haven`t shipped any more since June 26th.

 

[Cookie Monster]

http://www.guru3d.com/news-story/nvidia-geforce-gtx-870-and-gtx-880-available-in-november.html

I can write with pretty good indication that in the November 2014 we will see Nvidia launch their new GeForce GTX 870 and GTX 880 graphics cards. Last week at an Nvidia event a little birdie already whispered in my ears 'we'll see you again soon Hilbert'. More info from an external source however indicates that 28nm Maxwell based GPUs (indeed not 20nm) will be released within roughly three months time.

 

[RussianSensation]

Well if Maxwell is only 28nm sounds like NV is doing a repeat of GK104/110 split or some kind of a "hybrid" split. Perhaps we will see 28nm GM204 gaming flagship 880, 28nm GM200 as Titan 2 in 3-4 months for the HPC market, then in late 2015/early 2016 GM210 consumer gaming flagship #2 on 16/20nm.

I will be very surprised if a 2560 core, 256-bit Maxwell card will provide a 50% increase in performance over 780Ti and essentially be near the top for 2 years of Maxwell generation.

Have to command NV on aiming for a November launch. Sounds like a good time for GPU purchases.

 

[tviceman]

Unless 20nm is no good for anything other than SoC's, or unless gm200/210 will be stripped of some it's graphics rendering transistors (to save on die space and make it exclusive to the HPC market) I still think GM200 may debut on 20nm.

Either that, or Nvidia went with 20nm HPM. But who knows.