[KitGuru] Pascal's code names revealed?

[RussianSensation]

May 9, 2015

"The upcoming PK100 and PK104 graphics processing units will not only feature major architectural innovations, but they will also be made using either 14nm or 16nm manufacturing technology with fin-shaped field-effect transistors (FinFETs).

“Pascal” GPUs will be the first graphics chips from Nvidia to be made using an all-new process technology since “Kepler” in 2012. Finer fabrication process is something that should permit Nvidia engineers to considerably increase the number of stream processors and other units inside the company’s future GPUs, dramatically increasing their performance.

In addition, next-generation graphics processing units from Nvidia will support second-generation stacked high-bandwidth memory (HBM2). The HBM2 will let Nvidia and its partners build graphics cards with up to 32GB of onboard memory and 820GB/s – 1TB/s bandwidth. Performance of such graphics adapters in ultra-high-definition (UHD) resolutions like 4K (3840*2160, 4096*2160) and 5K (5120*2880) should be extremely high.

For supercomputers, the “Big Pascal” chip will integrate NVLink interconnection tech with 80GB/s or higher bandwidth, which will significantly increase performance of “Pascal”-based Tesla accelerators in high-performance computing (HPC) applications. Moreover, NVLink could bring major improvements to multi-GPU technologies thanks to increased bandwidth for inter-GPU communications.

“We have found over the years to be able to focus on just one thing, which is visual computing, and be able to leverage that one thing across PC, cloud, and mobile, and be able to address four very large markets with that one thing: gaming, enterprise, cloud, and automotive,” said Mr. Huang. “We can do this one thing and now be able to enjoy all and deliver the capabilities to the market in all three major computing platforms, and gain four vertical markets that are quite frankly very exciting.”

Kit-Guru

Sounds phenomenal!

Bonus:

At a press conference in Tokyo, Japan, Nvidia revealed that it would release its graphics processing units powered by the “Volta” architecture in 2017, one year after the first GPUs featuring the “Pascal” architecture will see the light of the day, 4Gamer.net reports. Nvidia did not reveal the difference between “Pascal” and “Volta”, but one of the features the latter is expected to support is second-generation NVLink with 80GB/s – 200GB/s bandwidth.

Kit-Guru

 

[hawtdawg]

They must be scared of what AMD is about to drop if they're already starting the hype train for a product that's a year or more away.

 

[RussianSensation]

They must be scared of what AMD is about to drop if they're already starting the hype train for a product that's a year or more away.

I think it had more to do with NV'st 1st quarter earnings and setting the stage for what's next for investors/shareholders. They also had rather poor guidance for Q2 2016 (i.e, calendar Q2 2015), which means the current lull in PC gaming leading up to Windows 10 and currency headwinds (NV reports in USD but since USD went up relative to a lot of world's currencies, that makes it more expensive to buy NV products for countries that don't use $USD).

Forward looking expectations
"Although NVIDIA came in largely in-line during the most recent quarter, it missed its revenue guidance for the second fiscal quarter by quite a lot. Analyst consensus sat at $1.18 billion going into the report, but NVIDIA is calling for just $1.01 billion at the midpoint of its guidance. Even at the high end of its guidance range, or about $1.03 billion, NVIDIA's guidance significantly misses the mark.

What's going on?

According to CFO Colette Kress, two major factors that led to NVIDIA guiding as it did -- aside from typical seasonal trends -- were the following. Firstly, she said that thanks to "European currency weakness," demand is lower than where it likely would have been without said currency headwinds.

Next, Kress said that a "lull" in the overall PC market ahead of the launch of Windows 10 is impacting the company's revenue outlook for the coming quarter. This appears to be consistent with PC chipmaker Intel's view of the market."

Source

------------------

Anyway, I am very excited for Post-GCN and Pascal architectures as 28nm and GDDR5 is getting long in the tooth.

 

[96Firebird]

Weird that they would switch up the chip names with Pascal. Wonder what the PK stands for...

For that matter, what do the G's in the current chip names stand for? Graphics? GFxxx, GKxxx, GMxxx... Now PKxxx?

 

[n0x1ous]

Weird that they would switch up the chip names with Pascal. Wonder what the PK stands for...

For that matter, what do the G's in the current chip names stand for? Graphics? GFxxx, GKxxx, GMxxx... Now PKxxx?

Got to be either GeForce or Graphics

 

[96Firebird]

Got to be either GeForce or Graphics

I always figured it was GeForce, but the more I thought about it today after reading this post, the more I think it is Graphics. These chips are used in more than just the GeForce line.

 

[Silverforce11]

We're what, 1 month from AMD's launch and still nothing major is known.

To talk about something potentially 18 months or more away (bless JH if he thinks its a smooth ride on a new node!)... bit premature, no?

 

[raghu78]

We're what, 1 month from AMD's launch and still nothing major is known.

To talk about something potentially 18 months or more away (bless JH if he thinks its a smooth ride on a new node!)... bit premature, no?

well said. The foundries are struggling to ramp <100 sq mm mobile SoC chips with <=5w TDP with high yields and in high volume. Good luck with 250 - 300 sq mm high performance GPUs using a completely new manufacturing methodology (2.5D stacking on silicon interposer) using a state of the memory technology HBM2 which will just start ramping in Q2 2016. If Nvidia wants us to believe this will be smooth sorry I am going to call their bluff. The last time they tried something as risky was Fermi GF100 and we all know how that turned out.

I would be surprised if Nvidia got a 16/14nm FINFET GP104 with HBM2 before Q4 2016 given their complete lack of experience with 2.5D manufacturing. As for the question of yields and ramping in high volume well lets say I am mildly pessimistic. :whiste:

TSMC in one of their earnings calls last year said 2017 would be the year when FINFET would be widely adopted and would overtake 28nm in wafer volume. So I think people should tone down their expectations for 2016. The best people should expect is a late Q3 2016 launch with very high volume in 2017.

 

[Serandur]

I've said elsewhere that they never even stated where they got the supposed code names from. I personally think it's just an incorrect guess they derived from melding Pascal and Kepler's names, not anything official. What could the "K" possibly stand for?

This looks more official:

And this:

 

[DownTheSky]

Nothing new except volta stuff.

 

[RussianSensation]

I've said elsewhere that they never even stated where they got the supposed code names from. I personally think it's just an incorrect guess they derived from melding Pascal and Kepler's names, not anything official. What could the "K" possibly stand for?

Ya, I was wondering about that too. I was always under the impression NV would just continue with GP200 and GP204 codenames. Thanks for your post!

 

[ShintaiDK]

They must be scared of what AMD is about to drop if they're already starting the hype train for a product that's a year or more away.

Its no different than with Kepler and Maxwell.

 

[n0x1ous]

I've said elsewhere that they never even stated where they got the supposed code names from. I personally think it's just an incorrect guess they derived from melding Pascal and Kepler's names, not anything official. What could the "K" possibly stand for?

This looks more official:

And this:

Thank you! that makes more sense. Theres no reason for them to change the codename to a different cadence.

 

[Qwertilot]

HBM2 for 'big' Pascal is a given of course but it presumably has to be in question if small/medium Pascal/4xx will get it?

Hardly obvious if they even really need it of course. 4GB is plenty of VRAM for a midrange card and HBM1 is hardly slow

I don't think something like this would be too implausible if NV are following their strategy from Maxwell:
A 2GB 950/ti, end of this year/early 2016 maybe even without HBM;
Some medium sized range with 4GB of HBM1 in ~the middle of 2016;
The bigger stuff following on towards the end on the year/into 2017.

Or they might of course do something rather different this time!

 

[exar333]

Its no different than with Kepler and Maxwell.

NV will really want this ASAP to refresh their Tesla line-up. Those coming from Fermi or Kepler to this should see some MASSIVE performance increases for GPGPU stuff....

Exciting.

 

[imaheadcase]

We're what, 1 month from AMD's launch and still nothing major is known.

To talk about something potentially 18 months or more away (bless JH if he thinks its a smooth ride on a new node!)... bit premature, no?

They are not competing with amd anymore, so its a nonissue.

 

[Silverforce11]

They are not competing with amd anymore, so its a nonissue.

Its not a reference to competition, its a reference to the facts we know so little prior to tech launches today, its irrelevant to speculate on something that may or may not come in 18 months time.

I mean what would you even say? That its gonna be the fastest GPU ever, the most efficient? That is assured on 14/16nm.

 

[ShintaiDK]

NV will really want this ASAP to refresh their Tesla line-up. Those coming from Fermi or Kepler to this should see some MASSIVE performance increases for GPGPU stuff....

Exciting.

Depends, Pascal focuses on half precision that seems to be a new market. And I assume KNL is the reason. It seems more to me that nVidia is focusing on gaming again.

 

[PPB]

Depends, Pascal focuses on half precision that seems to be a new market. And I assume KNL is the reason. It seems more to me that nVidia is focusing on gaming again.

Assuming Nvidia's results are a product of their intentions and not of what they are left with by the circunstances.

The truth is that to improve perf/watt they threw first FP64 under the bus, and now for marketing sake they will hype a deformity such as half precision/FP16 (like it is something new), while in reality it is at least single precision/FP32 what is used for GPGPU (obviously FP64 is used too).

Considering the worsening GPGPU dev relations, the lack of any meaningful FP64 Maxwell product and the threathening competition it will face from both AMD and Intel in this area, seems like Nvidia is rapidly doing an 180° and shifting to being a full gaming company. Saying Pascal will focus on FP16 like earlier products couldnt do 1/2 FP32 to FP16 ratio too is just falling to corporate slides and NV hype machine.

 

[Cookie Monster]

Considering the worsening GPGPU dev relations, the lack of any meaningful FP64 Maxwell product and the threathening competition it will face from both AMD and Intel in this area, seems like Nvidia is rapidly doing an 180° and shifting to being a full gaming company. Saying Pascal will focus on FP16 like earlier products couldnt do 1/2 FP32 to FP16 ratio too is just falling to corporate slides and NV hype machine.

Hmm..? What is a GPGPU dev relations?

 

[Cookie Monster]

Depends, Pascal focuses on half precision that seems to be a new market. And I assume KNL is the reason. It seems more to me that nVidia is focusing on gaming again.

Given the 14nm jump, theyll surely refocus on FP64 (well for the one that will succeed GK210 as it will be getting abit long in the tooth). Hence the existence of GK210 instead in the era of Maxwell uarch.

 

[ShintaiDK]

Assuming Nvidia's results are a product of their intentions and not of what they are left with by the circunstances.

The truth is that to improve perf/watt they threw first FP64 under the bus, and now for marketing sake they will hype a deformity such as half precision/FP16 (like it is something new), while in reality it is at least single precision/FP32 what is used for GPGPU (obviously FP64 is used too).

Considering the worsening GPGPU dev relations, the lack of any meaningful FP64 Maxwell product and the threathening competition it will face from both AMD and Intel in this area, seems like Nvidia is rapidly doing an 180° and shifting to being a full gaming company. Saying Pascal will focus on FP16 like earlier products couldnt do 1/2 FP32 to FP16 ratio too is just falling to corporate slides and NV hype machine.

AMD in its Q1 result suffered greatly in the compute part as well. I just think the GPGPU compete as we know it (HPC) is slowly over.

Its really really hard to see nVidia or AMD having any sort of competition to KNL.

 

[exar333]

AMD in its Q1 result suffered greatly in the compute part as well. I just think the GPGPU compete as we know it (HPC) is slowly over.

Its really really hard to see nVidia or AMD having any sort of competition to KNL.

It would be surprising (to me) if NV totally abandoned this market. Its a key revenue area for them today. Yes Maxwell somewhat abandons this, but what else can they do when they need to keep-up on the gaming side and still are dealing with 28nm constraints?

Now, if the word on FP64 is totally absent leading-up to the Pascal launch, then maybe we can put a nail in the coffin for that going forward. BUT the 14/16nm node should provide them some ability to allocate some space back to more compute functions.

Just thinking out loud, but maybe NV will follow a more tick-tock type release where the initial node offering provides more GPGPU space, while subsequent gens on the same node decrease this. This allows them to refresh their Tesla line every-other product line, while still having breathing room to allocate more space to gaming, if need be. This might be an insurance policy if they hit another node like 28nm, where they are stuck there for multiple generations.

Either way, NVLink points to a focus on GPGPU functionality, but NV could just abandon the FP64 market and focus just on FP32 and below? I guess we will find out closer to launch.

 

[ShintaiDK]

By comeptition I should have written DP in case anyone didnt get that from the context..

nVidia talks a lot about half precision with Pascal. So thats obviously their future bet. But we really only know for sure when they spill the beans on Volta.

 

[exar333]

By comeptition I should have written DP in case anyone didnt get that from the context..

nVidia talks a lot about half precision with Pascal. So thats obviously their future bet. But we really only know for sure when they spill the beans on Volta.

That makes sense. I do mostly agree with you on that point...

Compute is starting to be a fragmented market and I agree that we will likely see different companies start targeting slices of that market with hardware designed to be efficient in that space. That differs from most existing compute options today that are more 'jack of all trades'.