[Kitguru]Nvidia`s big Pascal GP100 have taped out - Q1 2016 release

[happy medium]

They said it supports HBM2

ok gotcha.

 

[RoarTiger]

..
Disregarding that Pascal wont be available in Q1 because of HBM2 like someone here tries to say, is just silly when Eyrines didnt even mention it and that its not even a requirement.
....

3D memory is one of the three defining characteristics of Pascal ( Unified Memory, 3D Memory and NVLink). Every Nvidia slide presented at GTC included "3D Memory" when discussing Pascal. There is a reason every promotional image like the one included in the OP show HBM memory...

http://www.anandtech.com/show/7900/nvidia-updates-gpu-roadmap-unveils-pascal-architecture-for-2016
"What’s new is that NVIDIA has confirmed they will be using JEDEC’s High Bandwidth Memory (HBM) standard, and the test vehicle Pascal card we have seen uses entirely on-package memory, so there isn’t a split memory design. "

 

[Samwell]

Even HBM2 is not impossible for it. You have to look at the quantities Nv is needing for GP100 and the price they'll sell it. This Chip will come in the professional space first and way later for consumers. I could imagine not before 2017 for regular consumers. NV is selling maybe 10k Teslas a month. So 40k HBM Stacks would be needed per month. That's not even mass production scale. They could buy pre mass production HBM2 and pay 100$ dollar per stack. Also if yield is so low, that a die costs 500$ and together they spend 1k$ as production cost of the card, it doesn't matter. They'll sell it for at least 3k $ anyway.

 

[Silverforce11]

"This is a change from Kepler and Maxwell where the smallest chip (GK107 and GM107 respectively) taped out first. Maybe the experience with 20nm was enough for NV to go back to their usual big die first strategy (Fermi)."

Did NV ever sell a big 20nm dGPU?

Or did they only make tiny mobile SOCs on that node..

Experience with 20nm != 16nm ff. The lesson of Fermi was a strong one. For them to ignore that when they dominate the market and their competition isn't putting pressure on marketshare (ie, no need to take risks!)... unlikely!

 

[maddie]

"This is a change from Kepler and Maxwell where the smallest chip (GK107 and GM107 respectively) taped out first. Maybe the experience with 20nm was enough for NV to go back to their usual big die first strategy (Fermi)."

Did NV ever sell a big 20nm dGPU?

Or did they only make tiny mobile SOCs on that node..

Experience with 20nm != 16nm ff. The lesson of Fermi was a strong one. For them to ignore that when they dominate the market and their competition isn't putting pressure on marketshare (ie, no need to take risks!)... unlikely!

Nvidia has always been a 'hungry' company that fights hard. I think they will take risks if the reward is great. We have at least two scenarios.

1) Nvidia realizes AMD is very vulnerable and is trying to bury them forever.
2) AMDs new GPU line has Nvidia very worried that they might lose their gains made recently.

Third option?

 

[96Firebird]

Nvidia has always been a 'hungry' company that fights hard. I think they will take risks if the reward is great. We have at least two scenarios.

1) Nvidia realizes AMD is very vulnerable and is trying to bury them forever.
2) AMDs new GPU line has Nvidia very worried that they might lose their gains made recently.

Third option?

Sure, since we're just speculating, I'll add a third option...

Nvidia wants to come out with a DP enabled GPU to replace it's Tesla lineup and sell more $5000 graphics cards.

 

[maddie]

Sure, since we're just speculating, I'll add a third option...

Nvidia wants to come out with a DP enabled GPU to replace it's Tesla lineup and sell more $5000 graphics cards.

Valid reason. I don't think a small die DP compute product will be a big seller.

You think Q1 is realistic for a big die?

 

[Silverforce11]

Sure, since we're just speculating, I'll add a third option...

Nvidia wants to come out with a DP enabled GPU to replace it's Tesla lineup and sell more $5000 graphics cards.

That's the most sound option. $ talks. $5K Teslas make low yields worth it.

But also lets not forget NV has been milking it hard with small dies, GK104 at flagship prices, then release big Kepler.. they did it again with Maxwell.

That approach works and on a a new node, makes the most sense for consumer dGPUs.

 

[Samwell]

Sure, since we're just speculating, I'll add a third option...

Nvidia wants to come out with a DP enabled GPU to replace it's Tesla lineup and sell more $5000 graphics cards.

Exactly what i' thinking. Even if Fiji is good, they can compete with price in consumer space, but in professional? Intel is coming with a 600mm² monster chip with HMC in 14nm in the next months. No one will buy Tesla after Intels launch till Pascal comes.

 

[videogames101]

Do you think having the microbump connections to the interposer allow better signal routing to various parts of the die? Could this be one significant difference between memory controller designs.

Yes. As far as on-die routing, in general microbumps make for simpler routing than die-edge pads do. As far as signal integrity, an interposer should outperform a PCB. The difference is significant for the physical design and PHY guys, but I don't see how it might affect the logic inside the memory controller. However Nvidia in all likelihood was already using a flip-chip microbump style, so it still shouldn't be a huge change for the controller design.

 

[JoeRambo]

Did NV ever sell a big 20nm dGPU?

Or did they only make tiny mobile SOCs on that node..

Experience with 20nm != 16nm ff. ...

Actually 16ff is 20nm with finfets. Probably all the rules of 20nm design still apply, except resulting silicon is more performant and can be tighter due to finfets increasing transistor performance? So i guess original author meant that 20nm experience with mobile stuff was enough for Nvidia to gather experience to go stright to GP100.

 

[Vesku]

Valid reason. I don't think a small die DP compute product will be a big seller.

You think Q1 is realistic for a big die?

It really comes down to what Nvidia thinks Intel and AMD will release in 2016 for HPC. They've been touting Pascal+HBM2 but if they feel under pressure I think Nvidia is a company that isn't afraid to fund a stopgap solution to cover until Q3-Q4 of 2016.

 

[krumme]

Yes. As far as on-die routing, in general microbumps make for simpler routing than die-edge pads do. As far as signal integrity, an interposer should outperform a PCB. The difference is significant for the physical design and PHY guys, but I don't see how it might affect the logic inside the memory controller. However Nvidia in all likelihood was already using a flip-chip microbump style, so it still shouldn't be a huge change for the controller design.

I dont understand this. Space consuming analogue logic driving long traces to gddr5 is the same as logic as driving hbm???
You make it sound like a walk in the park. Do you actually know anything about it?

 

[jpiniero]

It really comes down to what Nvidia thinks Intel and AMD will release in 2016 for HPC. They've been touting Pascal+HBM2 but if they feel under pressure I think Nvidia is a company that isn't afraid to fund a stopgap solution to cover until Q3-Q4 of 2016.

I thought Knights Landing was coming out at the end of this year.

What I was imaging was a die with a core count similar to the Titan X or maybe a bit more. But with Half DP. There would be Tesla, Quadro and Titan models. It would be nice if it was HBM, but if it has to be GDDR5 then so be it.

I was also thinking... would mobile Titan (complete with Half DP) make sense? You could sell a pretty pricey Mobile Titan and then sell some ultra high end mobile models with the DP cut, but with big performance gains compared to Mobile Maxwell 2 since the core count would be much higher. Then you would fill in the rest of the mobile and desktop lineup with small tweaks or rebrands.

 

[3DVagabond]

I thought Knights Landing was coming out at the end of this year.

What I was imaging was a die with a core count similar to the Titan X or maybe a bit more. But with Half DP. There would be Tesla, Quadro and Titan models. It would be nice if it was HBM, but if it has to be GDDR5 then so be it.

I was also thinking... would mobile Titan (complete with Half DP) make sense? You could sell a pretty pricey Mobile Titan and then sell some ultra high end mobile models with the DP cut, but with big performance gains compared to Mobile Maxwell 2 since the core count would be much higher. Then you would fill in the rest of the mobile and desktop lineup with small tweaks or rebrands.

Does nVidia currently offer anything with 1/2 DP?

 

[JDG1980]

Intel is coming with a 600mm² monster chip with HMC in 14nm in the next months. No one will buy Tesla after Intels launch till Pascal comes.

It's a mistake to assume that Intel will always succeed at whatever they try. Just because they have a process advantage and a massive R&D budget doesn't magically give them the industry expertise in graphics and HPC. Not to mention that a lot of Tesla users specifically want CUDA, which Knights Landing won't have.

Intel has a lot going for them, but they've fallen flat on their face before. Consider Larrabee. Consider the repeated attempts to get into ARM's ultra-mobile market, which has largely failed so far and shows no sign of turning around, despite massive infusions of "contra revenue" (parts giveaways). Consider how long it's taken them to get their best iGPUs to match the performance of something as paltry as an AMD Cape Verde (a low end product from 2012).

 

[Headfoot]

It's very possible, likely even, that they release a Quadro/Tesla/Professional GPU nearly the same time as the midrange chip a-la original GK110 in the Titan Supercomputer. Then eventually release it for desktop. They can get a midrange release, milk the people who for some reason buy the hopped-up midrange chips, then milk them again when they release the high end chip all the while immediately releasing the high end chip for professionals to battle Knights Landing

 

[MeldarthX]

Nvidia has to be careful in the professional market - they dropped down 65% a couple quarters ago with AMD gettiing for the first time 25% of the market and Intel 10% - what it is now; I can't say.

That is bread and butter - AMD goes biggest DP card on the market and looks like its going to double down with Fury - AMD will gain market share back in gamer's market - but its the professional market they keep chipping away that will really hurt Nvidia....

 

[Genx87]

I'd say this will be for the professional markets. With consumer products showing up later in 2016.

 

[Genx87]

Nvidia has to be careful in the professional market - they dropped down 65% a couple quarters ago with AMD gettiing for the first time 25% of the market and Intel 10% - what it is now; I can't say.

That is bread and butter - AMD goes biggest DP card on the market and looks like its going to double down with Fury - AMD will gain market share back in gamer's market - but its the professional market they keep chipping away that will really hurt Nvidia....

http://jonpeddie.com/publications/workstation_report

Q1 2015 Nvidia had ~80% of the professional graphics market.

 

[Enigmoid]

Nvidia has to be careful in the professional market - they dropped down 65% a couple quarters ago with AMD gettiing for the first time 25% of the market and Intel 10% - what it is now; I can't say.

That is bread and butter - AMD goes biggest DP card on the market and looks like its going to double down with Fury - AMD will gain market share back in gamer's market - but its the professional market they keep chipping away that will really hurt Nvidia....

Fiji with 4 GB is largely worthless for a lot of DP calculations (not all but any kind of large system calculation). 64 bit values take up twice the memory space; running a simulation with all doubles vs. floats will take up ~ 2x more vram.

There is also not much room for growth. If you change the simulation its pretty easy to go over 4GB. The W9100 in comparison has 16 GB vram so this is a tremendous step down in terms of perf/vram.

 

[Stuka87]

Fiji with 4 GB is largely worthless for a lot of DP calculations (not all but any kind of large system calculation). 64 bit values take up twice the memory space; running a simulation with all doubles vs. floats will take up ~ 2x more vram.

There is also not much room for growth. If you change the simulation its pretty easy to go over 4GB. The W9100 in comparison has 16 GB vram so this is a tremendous step down in terms of perf/vram.

Performance is not the reason nVidia is ahead though. Last I checked Hawaii has more DP compute than Maxwell. nVidia holds onto their market because of CUDA. As long as companies are locked into a single vendor, nVidia will stay in the lead. But as more companies switch to OpenCL so they then have three choices instead of one, this may change.

 

[nvgpu]

http://www.kitguru.net/components/a...-out-worlds-first-16nm-finfet-system-on-chip/

KitGuru Says: The tape-out of Xilinx confirms that the first wave of chips to be made using TSMC’s CLN16FF+ manufacturing technology has taped out. It is believed that Nvidia Corp.’s code-named GP100 graphics processing unit – the first GPU to be based on the company’s “Pascal” architecture – is also among the first to be made using TSMC’s 16nm FinFET+ fabrication process.

 

[railven]

I turned on my Maxwell 2 timer. If this card goes obsolete when Pascal launches and it seems I'm going to be on the NV boat for a while, I'll just jump over when Pascal launches and hand Maxwell 2 to the GF.

I hope it's at least a year so I can maintain my upgrade schedule and not hear the GF complain

 

[DooKey]

Wow, hope this is true. I'd love to have me some GP100 lovin' within the next year.