NVIDIA Pascal Thread

[Arachnotronic]

Hmm, curious, I would think Nvidia would want to amortize GP100 costs over both consumer and professional sales channels.

Nope, NVIDIA management was quite clear at analyst day that they will be building products targeted/optimized for each segment. Tesla/Quadro revenues are large enough that they can support the incremental costs to do separate GPU dies.

Remember, R&D to build a product like GP100 has already been sunk. It's not like a wafer fab where you spend X billions of dollars on a wafer fab and you are getting hit with depreciation costs that you need to assign to every wafer, thus making amortization over a large number of wafers imperative. This is why Intel was able to basically pull the plug on Broadwell so quickly when Skylake was around the corner...it's built on the same process, it was ready to go, and there is zero economic benefit to milking Broadwell chips when you could sell customers better Skylake chips.

 

[3DVagabond]

Listen to NVIDIA's analyst day comments. GP100 was built specifically for HPC. Those FP64 units are totally wasteful on a GPU intended to be used in gaming.

So you think if AMD releases Vega and it's the fastest solution available that nVidia will just cede the performance crown? Most top cards are made for FP64 but they make consumer cards out of them as well. They don't typically make a GPU that's used only in the pro market.

 

[Arachnotronic]

So you think if AMD releases Vega and it's the fastest solution available that nVidia will just cede the performance crown?

No, I am sure they have a specialized, gaming-focused part that uses HBM2. NVIDIA probably makes enough money selling Titan + cut-down variants (i.e. 980 Ti) that it makes sense to do a separate, higher-end product. They literally did this with GM200.

Most top cards are made for FP64 but they make consumer cards out of them as well. They don't typically make a GPU that's used only in the pro market.

This has historically been done to save on R&D spending, but Tesla is a large enough business for NVIDIA to support a separate die/product. Remember that Intel with a much, much smaller share of HPC market is doing Xeon Phi which is basically used only for HPC. NVIDIA is taking its fundamental Pascal architecture and building a special chip with HPC-friendly bits (i.e. FP64) to try to maintain its much larger share of that market.

AMD has minimal HPC accelerator share, ditto with professional workstation cards. This is why they can't really justify shifting to the strategy that NVIDIA now has the luxury of pursuing.

 

[Silverforce11]

@Arachnotronic

If GP100 is made only for HPC and FP64, why does it need TMUs and ROPs?

Frankly, these architectures have always been dual purpose. When low volume or yields are key, they will be in Teslas first. When the situation improves, a GTX SKU can be made from the left overs.

Here's something for you guys who think there's a GP102 for gaming to ponder. NV's recent domination of the market resulted in record revenue, but what's their profits? Not much. It has been said that each unique chip costs ~$300M in R&D, certainly for NV's previous gen. If NV wants to spend a few hundred mil on a GP102, that close to HALF their yearly profits.

Much more profit selling GP100 as GTX once volume/yields are good and Tesla demands die down. We'll see GP100 as a GTX/Titan sometime in 2017.

 

[Arachnotronic]

@Arachnotronic

If GP100 is made only for HPC and FP64, why does it need TMUs and ROPs?

GK210 has TMUs and ROPs, too.

Frankly, these architectures have always been dual purpose. When low volume or yields are key, they will be in Teslas first. When the situation improves, a GTX SKU can be made from the left overs.

We'll see.

 

[Arachnotronic]

Here's something for you guys who think there's a GP102 for gaming to ponder. NV's recent domination of the market resulted in record revenue, but what's their profits? Not much. It has been said that each unique chip costs ~$300M in R&D, certainly for NV's previous gen.

There is no way that a GP102 is going to cost an additional $300 million in R&D. That is a nonsense figure and should not be used for any serious analysis.

 

[Ajay]

This isn't happening. A Titan with HBM2 is what GP102 is for and should be enough to compete with Vega. I would not be surprised if Vega 10 is like GP100 in that it won't be sold as a consumer product, only for HPC and workstations.

So based on what Nvidia has already publicly said, that makes sense. Still, what is this GP102 you speak of? I find scant references to it on the web.

 

[Arachnotronic]

So based on what Nvidia has already publicly said, that makes sense. Still, what is this GP102 you speak of? I find scant references to it on the web.

http://www.3dcenter.org/news/reihen...cht-gp100-gp102-gp104-gp106-gp107-gp10b-gv100

References to it found in NVIDIA drivers.

 

[Ajay]

There is no way that a GP102 is going to cost an additional $300 million in R&D. That is a nonsense figure and should not be used for any serious analysis.

No, the design costs should be lower. But everything else is still costly, from computer verification of the desing to masks to test runs and qualification - so hardly cheap either.

 

[Ajay]

http://www.3dcenter.org/news/reihen...cht-gp100-gp102-gp104-gp106-gp107-gp10b-gv100

References to it found in NVIDIA drivers.

Yes, I saw that and the description - "completely unsure of what nVidia has herewith" (after translation).

 

[Arachnotronic]

No, the design costs should be lower. But everything else is still costly, from computer verification of the desing to masks to test runs and qualification - so hardly cheap either.

Hardly cheap, but remember...NVIDIA was able to justify the costs of building GM200, a part that probably saw the vast majority of its sales volume in the form of Titan X/GTX 980 Ti.

That segment of the dGPU market is likely growing and it is probably very lucrative; NVIDIA can certainly justify the added spending required to make it happen. Remember, they are going to do a non-FP64 Pascal for GP104, so I would expect GP102 to essentially be a scaled up version of the same GP104 fundamental building blocks + the HBM2 implementation that was already paid for with GP100 development.

tl;dr -- the uber-high end dGPU gaming market is worth pursuing with targeted products, especially since NVIDIA wants to defend its share against AMD. NVIDIA would be absolutely stupid to not always be on its toes about a potential AMD comeback.

 

[Silverforce11]

Hardly cheap, but remember...NVIDIA was able to justify the costs of building GM200, a part that probably saw the vast majority of its sales volume in the form of Titan X/GTX 980 Ti.

On a VERY mature 28nm, at a time when mobile SOC companies were moving onto 16nm TSMC, freeing up volume at 28nm.

Now NV is on the cutting edge 16nm and have to compete with all the major players for wafers.

How much sense is there for NV to make a big gaming focused chip in this scenario?

 

[JDG1980]

I would not be surprised if Vega 10 is like GP100 in that it won't be sold as a consumer product, only for HPC and workstations.

This won't happen. First of all, while AMD does have some HPC sales, they don't have nearly as many as Nvidia, not enough to justify a dedicated chip. Secondly, AMD has a more efficient implementation of FP64 that doesn't require special-purpose SIMD units. Look at Hawaii: support for DP at a 1/2 rate (same as GP100), but it still has one of AMD's best showings in gaming perf/mm^2 of any 28nm GPU.

 

[JDG1980]

Hardly cheap, but remember...NVIDIA was able to justify the costs of building GM200, a part that probably saw the vast majority of its sales volume in the form of Titan X/GTX 980 Ti.

Sales volume, sure. That said, I'd be interested to see a breakdown on profits. The GM200-based Quadro M6000 is a $5,000 card; the profit margins on that have to be immense. I wouldn't be surprised if Nvidia made more profit on each M6000 sale than on 25 GTX 980 Ti chips.

 

[antihelten]

Maybe in the past. Nowadays, the next best thing is 980Ti at 6GB. Thats a massive difference.

My point is though that the reason the 980 Ti only exists in a 6GB is because it would cannibalize the Titan X otherwise. With the old Titan this wasn't really an issue since you still had the DP performance which wasn't available on 780/780 Ti, and here we did actually see cards with double the memory (although I'm not sure if there was ever a 6GB 780 Ti, but I think it was rumoured).

So basically the memory split between 980 Ti and Titan X is an artificial one that purely exists to validate the existence of Titan X. So whereas in the past one might pay an extra $50-100 for a card with double the memory, now you have to pay $350 for it. Hence it being a "meh" option in my opinion (to be fair, the kind of compute I do doesn't really need that much memory anyway)

 

[airfathaaaaa]

On a VERY mature 28nm, at a time when mobile SOC companies were moving onto 16nm TSMC, freeing up volume at 28nm.

Now NV is on the cutting edge 16nm and have to compete with all the major players for wafers.

How much sense is there for NV to make a big gaming focused chip in this scenario?

nah since mid 2015 most big one are passing already on 14 and 10nm and some are still on the "simple" 16nm aka 20nm without ff tech

 

[Timmah!]

My point is though that the reason the 980 Ti only exists in a 6GB is because it would cannibalize the Titan X otherwise. With the old Titan this wasn't really an issue since you still had the DP performance which wasn't available on 780/780 Ti, and here we did actually see cards with double the memory (although I'm not sure if there was ever a 6GB 780 Ti, but I think it was rumoured).

So basically the memory split between 980 Ti and Titan X is an artificial one that purely exists to validate the existence of Titan X. So whereas in the past one might pay an extra $50-100 for a card with double the memory, now you have to pay $350 for it. Hence it being a "meh" option in my opinion (to be fair, the kind of compute I do doesn't really need that much memory anyway)

So if i understand you correctly, its "meh" for you, cause in the past you could get equal VRAM "upgrade" with regular Geforces way cheaper. Now thats indeed relevant reason and i totally understand it, but its sort of the same, as when i consider GTX680 or possibly upcoming 1080 (we shall see, may still surprise positively) to be "meh", as i could get similar performance relatively to the rest of its generation in the past way cheaper too.

Anyway, another slow day without any relevant news or leaks so far... disappointing.

 

[FatherMurphy]

This won't happen. First of all, while AMD does have some HPC sales, they don't have nearly as many as Nvidia, not enough to justify a dedicated chip. Secondly, AMD has a more efficient implementation of FP64 that doesn't require special-purpose SIMD units. Look at Hawaii: support for DP at a 1/2 rate (same as GP100), but it still has one of AMD's best showings in gaming perf/mm^2 of any 28nm GPU.

I agree with you that AMD does not have the HPC volume to justify a split SP/DP architecture.

But, I'm not sure that AMD has a "more efficient" implementation of FP64. Because AMD reuses its FP32 cores for FP64, AMD's implementation may use less die space than Nvidia's implementation (though I can't confirm that). But my understanding is that having FP32 cores double as FP64 units means a more complex data path, scheduling hardware, register composition... and it uses more power. Both Hawaii and Fermi, where Nvidia had 1/2 rate DP by having its Cuda Cores do both FP32 and FP64, are (I think, without question) less "efficient" chips power wise.

I'd also note that in AMD's drive to increase performance/watt since Tahiti, AMD has constantly lowered its DP rate in gaming cards -- Tahiti (1/4), Hawaii (1/8), Fiji (1/16). Of course, DP rate in gaming cards is next-to-pointless (unless you use the gaming card for DP compute work as well).

My point is that AMD's DP implementation may cost AMD some performance/watt and die space, but, without funds to create a more modular DP setup like Nvidia, AMD may not have have another option. On the other hand, if Nvidia can simply scale up GP104 (assuming it has far fewer DP units) by 50% and not clutter the die with DP units (vis-a-vis GM204 and GM200), then a GP102 may not be so far-fetched. Plus, GP102 can be marketed as Nvidia's "SP Performance Card" for Quadro (similar to GM200).

 

[Sweepr]

NVIDIA Deep Learning Day Spring 2016

DGX-1 is a super computer with a Tesla P100 × 8, Xeon processor × 2,7TB SSD, Tesla P100 are connected by NVLink, with 40TFLOPS, the arithmetic processing performance of 170TFLOPS at half-precision in double precision. Power consumption is a 3,200W.

　Mr. Lin will appeal "This is the performance of more than (a few years ago) Earth Simulator", the learning of AlexNet take 150 hours of Xeon × 2 server that finish in 2 hours. To achieve the same performance in the Xeon server requires a 250 node, DGX-1 was to achieve the same performance as 250 servers in a 3U rack size.

　Official sales start in Japan also announced, in Japan, GDEP Solutions Co., Ltd., HPC Systems, Inc., from the three companies of Hitachi, Ltd., will be sold from 26 days. There was no announcement about the price, 129,000 dollars (1,419 million and $ 1 = 110 yen equivalent) in the United States to be sold in the announced at GTC 2016 's.

http://pc.watch.impress.co.jp/docs/news/20160427_755418.html

 

[Arachnotronic]

NVIDIA Deep Learning Day Spring 2016

http://pc.watch.impress.co.jp/docs/news/20160427_755418.html

But Sweepr, I was told by the ADF that GP100 won't be available until 1Q 2017? Are you suggesting I was lied to?

 

[Sweepr]

But Sweepr, I was told by the ADF that GP100 won't be available until 1Q 2017? Are you suggesting I was lied to?

I must have missed the woodscrews, let me double check with SA's experts. Also no Pascal Geforce till Q4, obviously the leaked die photos were a stunt by NVIDIA, must be the first dies they got.

 

[Arachnotronic]

I must have missed the woodscrews, let me double check with SA's experts. Also no Pascal Geforce till Q4, obviously the leaked die photos were a stunt by NVIDIA, must be the first dies they got.

Yes, according to SemiAccurate, NVIDIA abandoned the high-end and mid-range dGPU markets.

http://semiaccurate.com/2009/10/06/...x275-gtx260-abandons-mid-and-high-end-market/

 

[DooKey]

Yes, according to SemiAccurate, NVIDIA abandoned the high-end and mid-range dGPU markets.

http://semiaccurate.com/2009/10/06/...x275-gtx260-abandons-mid-and-high-end-market/

Semiaccurate is the Jim Jones to the faithful ADF. Always lots of kookaid available.

 

[USER8000]

But Sweepr, I was told by the ADF that GP100 won't be available until 1Q 2017? Are you suggesting I was lied to?

Who is ADF?? The Nvidia CEO said the following during the GP100 unveil:

http://www.tomshardware.co.uk/nvidia-pascal-tesla-p100-gpu,news-52735.html

The Tesla P100 is in volume production today and will ship “soon." Jen-Hsun Huang said that really means it will show up in the cloud first, and ship via OEMs in Q1 2017.

You did actually watch the event right?? It was said during that event the GP100 would ship to HPC first and be available to OEMs by Q1 2017. Even the price of the system,ie,$129000 was mentioned during that event.

It was the same with the GK110 - it first shipped to ORNL months before other customers got it.

Is ADF another name for Jen-Hsun Huang?? That would JHH right??


Edit to post!

Here is arstechnica reporting on it:

http://arstechnica.co.uk/gadgets/2016/04/nvidia-tesla-p100-pascal-details/

Launching alongside the Tesla P100 is Nvidia's DGX-1 Deep Learning System. The DGX-1 features eight Tesla GP100 GPU cards providing 170 teraflops of half-precision performance from its 28,672 CUDA cores. The DGX-100 also features two 16-core Intel Xeon E5-2698 v3 2.3GHz CPUs, 512GB of DDR4 RAM, 4x 1.92TB SSD RAID, dual 10GbE, 4x InfiniBand EDR network ports, and requires a maximum of 3200W of power. It'll cost a mere $129,000 (~£103,000) when it launches in June. Those after a P100 in other types of servers will have to wait a little longer, with cards expected to reach systems in 2017, likely due to binning and HBM2 manufacturing constraints.

http://www.enterprisetech.com/2016/04/06/nvidia-unleashes-monster-pascal-gpu-card-gtc16/

The P100 is the flagship Pascal architecture offering, and it’s also the first product to implement the heralded HBM2 and NVLink technologies. During his keynote address, NVIDIA CEO Jen-Hsun Huang said the chip had entered volume production and would ship by first quarter 2017. Partners Dell, HPE, Cray and IBM are expected to come out with Pascal-equipped servers by the end of 2016, with production availability in early 2017.

http://techreport.com/news/29946/pascal-makes-its-debut-on-nvidia-tesla-p100-hpc-card

Nvidia says it's producing Tesla P100s in volume today. The company says it's devoting its entire production of Tesla P100 cards (and presumably the GP100 GPU) to its DGX-1 high-density HPC node systems and HPC servers from IBM, Dell, and Cray. DGX-1 nodes will be available in June for $129,000, while servers from other manufacturers are expected to become available in the first quarter of 2017.

It was stated by the Nvidia CEO himself the GP100 will only be available as part of DGX-1 system. People wanting GP100 based cards for anything else will be required to wait until next year it appears. That is what I got from watching it and from what the technical press have seemed to have reported.

 

[Kenmitch]

Who is ADF?? The Nvidia CEO said the following during the GP100 unveil:

http://www.tomshardware.co.uk/nvidia-pascal-tesla-p100-gpu,news-52735.html



You did actually watch the event right?? It was said during that event the GP100 would ship to HPC first and be available to OEMs by Q1 2017. Even the price of the system,ie,$129000 was mentioned during that event.

It was the same with the GK110 - it first shipped to ORNL months before other customers got it.

Is ADF another name for Jen-Hsun Huang?? That would JHH right??


Edit to post!

Here is arstechnica reporting on it:

http://arstechnica.co.uk/gadgets/2016/04/nvidia-tesla-p100-pascal-details/



http://www.enterprisetech.com/2016/04/06/nvidia-unleashes-monster-pascal-gpu-card-gtc16/



http://techreport.com/news/29946/pascal-makes-its-debut-on-nvidia-tesla-p100-hpc-card



It was stated by the Nvidia CEO himself the GP100 will only be available as part of DGX-1 system. People wanting GP100 based cards for anything else will be required to wait until next year it appears. That is what I got from watching it and from what the technical press have seemed to have reported.

Calm down before you have a stroke!

Those who use the term ADF love JHH....They'd probably have his babies if they were capable.

ADF is the term they use to make fun of those who support AMD. It's the best they could come up with I guess.