Revealed Nvidia GP100 Analysed... full GP100 specs extracted

[Big-G]

We all know that Nvidia revealed some GP100 info based on the P100 Tesla product launch, but two sites actually reviewed the graphics and diagrams and have identified... that GP100... isn't the full blown GP100.

 

Firstly from Guru3D -> http://www.guru3d.com/news-story/nvidia-gp100-gpu-architecture-recap-full-gpu-has-3840-shader-processors.html

Quote

Tesla Products	Tesla K40	Tesla M40	Tesla P100	GP100
GPU	GK110 (Kepler)	GM200 (Maxwell)	GP100 (Pascal)	GP100 (Pascal)
SMs	15	24	56	60
TPCs	15	24	28	30
FP32 CUDA Cores / SM	192	128	64	64
FP32 CUDA Cores / GPU	2880	3072	3584	3840
FP64 CUDA Cores / SM	64	4	32	32
FP64 CUDA Cores / GPU	960	96	1792	1920
Base Clock	745 MHz	948 MHz	1328 MHz	~1328 MHz
GPU Boost Clock	810/875 MHz	1114 MHz	1480 MHz	~1480 MHz
Texture Units	240	192	224	240
Memory Interface	384-bit GDDR5	384-bit GDDR5	4096-bit HBM2	4096-bit HBM2
Memory Size	Up to 12 GB	Up to 24 GB	16 GB	16 GB
L2 Cache Size	1536 KB	3072 KB	4096 KB	4096 KB
Register File Size / SM	256 KB	256 KB	256 KB	256 KB
Register File Size / GPU	3840 KB	6144 KB	14336 KB	14336 KB
TDP	235 Watts	250 Watts	300 Watts	~300 Watts
Transistors	7.1 billion	8 billion	15.3 billion	15.3 billion
GPU Die Size	551 mm²	601 mm²	610 mm²	610 mm²
Manufacturing Process	28-nm	28-nm	16-nm	16-nm

 

As the block diagram now shows, the GP100 features six graphics processing clusters (GPCs). Just look at the diagram and count along with me - each GPC holds 10 streaming multiprocessors (SMs) and then each SM has 64 CUDA cores and four texture units. Do the math and you'll reach 640 shader processors per GPC and 3840 shader cores with 240 texture units in total.

• 6 (GPC) x (10x64) = 3840 Shader processor units in total.

Meaning the GP100 used on the Tesla P100 is not fully enabled. Nvidia is known to out GPU that have disabled segments, it helps them selling different SKUs, the Tesla P100 holds a shader count of 3584 and thus has 56 SMs enabled (from the 60).

 

And corroborating the above, TechPoweUp -> https://www.techpowerup.com/221641/nvidia-pascal-gp100-silicon-detailed

Quote

The GP100 features six graphics processing clusters (GPCs). These are highly independent subdivisions of the GPU, with their own render front and back-ends. With the "Pascal" architecture, at least with the way it's implemented on the GP100, each GPC features 10 streaming multiprocessors (SMs), the basic number crunching machinery of the GPU. Each SM holds 64 CUDA cores. The GPC hence holds a total of 640 CUDA cores, and the entire GP100 chip holds 3,840 CUDA cores. Other vital specs include 240 TMUs. On the Tesla P100, NVIDIA enabled just 56 of the 60 streaming multiprocessors, working out to a CUDA core count of 3,584.

So definitely, the next Gen Titan will most likely be a very serious beast to content with. Although it is not certain if the next Ti will be GP100 or a separate GP104 SKU, but we all hope for the GP100 version

[Deil Grist]

300 watt single-GPUs. I'm okay with it if performance justifies it though. Still waiting for info on Polaris though so I can see if I should get a freesync monitor. 

[patrickjp93]

2 hours ago, Deil Grist said:

300 watt single-GPUs. I'm okay with it if performance justifies it though. Still waiting for info on Polaris though so I can see if I should get a freesync monitor. 

Given the clock speed is a whopping 40% above the Maxwell Titan X, that's a damn efficient GPU.

[Sharkyx1]

I absolutely do not believe that they will release a 610mm GPU on 16nm in 2016, maybe even the first half of 2017

[Stefan1024]

you know yield is a nasty thing. I'm wonering if there is even one chip per waver that has no defect. So stepping down a bit is reasonable.

[Shahnewaz]

I though we knew that already.

It was on their "Inside Pascal" dev blog, where Mark Harris wrote Pascal can have a full configuration of 3840 CUDA cores, but on their Tesla P100 spec sheet, it showed 3584 CUDA cores. It isn't full blown GP100, quite understandable because of early yield defects.

Link: https://devblogs.nvidia.com/parallelforall/inside-pascal/

[Big-G]

1 minute ago, Shahnewaz said:

I though we knew that already.

It was on their "Inside Pascal" dev blog, where Mark Harris wrote Pascal can have a full configuration of 3840 CUDA cores, but on their Tesla P100 spec sheet, it showed 3584 CUDA cores. It isn't full blown GP100, quite understandable because of early yield defects.

Link: https://devblogs.nvidia.com/parallelforall/inside-pascal/

Yeah, this is just for those who don't know about it.

 

Especially the AMD fanboys

[xAcid9]

Thought we already know that full Pascal have 60SM but only 56SM in GP100? They're pulling Titan and Titan Black once again.

[Master Disaster]

1 hour ago, xAcid9 said:

Thought we already know that full Pascal have 60SM but only 56SM in GP100? They're pulling Titan and Titan Black once again.

I don't think they're pulling anything. As other have said, GP100 is on a brand new manufacturers node and the reality of that is manufacturing defects.

 

It will take them time to refine the process to a point where then can get full blown chips out of it. 

[Prysin]

4 hours ago, patrickjp93 said:

Given the clock speed is a whopping 40% above the Maxwell Titan X, that's a damn efficient GPU.

Let's just pray that they use a better reference power phase and VRM design then for the TitanX... 108c for the VRMs under gaming load is not OK for a 300w GPU with 2x8 pin...

Hell, my 295x2 with a 7% OC to the core hits 104c with 550w or so going through 2x8pins...

 

Just appauling that Nvidia skimped on the most important part of a 1000 USD card, in terms of the cards longevity... I know the VRMs used for the Titan X and higher end Radeon cards are rated for 120c max operating temp. But anyone who has any insight into electromechanical engineering knows that copper and insulation will degrade at an accelerated rate if constantly pushed above 80% of its rated max.

[Stefan1024]

6 hours ago, Prysin said:

Let's just pray that they use a better reference power phase and VRM design then for the TitanX... 108c for the VRMs under gaming load is not OK for a 300w GPU with 2x8 pin...

Hell, my 295x2 with a 7% OC to the core hits 104c with 550w or so going through 2x8pins...

 

Just appauling that Nvidia skimped on the most important part of a 1000 USD card, in terms of the cards longevity... I know the VRMs used for the Titan X and higher end Radeon cards are rated for 120c max operating temp. But anyone who has any insight into electromechanical engineering knows that copper and insulation will degrade at an accelerated rate if constantly pushed above 80% of its rated max.

This is true. Also the temperature you measure on the oustde can be significantly lower than the silicon itself. And the 120C is for the silicon.

But if your card dies just after the warrenty exieres, nvidia will be happy to sell you a 1000$ card again...

[patrickjp93]

13 hours ago, Sharkyx1 said:

I absolutely do not believe that they will release a 610mm GPU on 16nm in 2016, maybe even the first half of 2017

Too bad, because you're already wrong. The Tesla P100s have been sampling since December and will be shipping at the end of Q3 as Nvidia has already stated. Now, that is a cut down core, and it is enterprise, but yes, in fact Nvidia will release a 600mm sq. GPU this year.

[patrickjp93]

8 hours ago, Prysin said:

Let's just pray that they use a better reference power phase and VRM design then for the TitanX... 108c for the VRMs under gaming load is not OK for a 300w GPU with 2x8 pin...

Hell, my 295x2 with a 7% OC to the core hits 104c with 550w or so going through 2x8pins...

 

Just appauling that Nvidia skimped on the most important part of a 1000 USD card, in terms of the cards longevity... I know the VRMs used for the Titan X and higher end Radeon cards are rated for 120c max operating temp. But anyone who has any insight into electromechanical engineering knows that copper and insulation will degrade at an accelerated rate if constantly pushed above 80% of its rated max.

The VRMs on the Titan X are fine. Just toss a water lock on it. Nvidia skimps on the cooling solutions, but not really the other components.

[Prysin]

3 hours ago, patrickjp93 said:

The VRMs on the Titan X are fine. Just toss a water lock on it. Nvidia skimps on the cooling solutions, but not really the other components.

so you want to say.

"its fine, just dump another 200-300 USD for a custom loop onto the card to solve Nvidia's fucking shitty job"

Come on. If that is your "solution" you can just fuck off patrick. That is the shitties most fanboyish excuse ive ever heard from you.

 

Seriously patrick, if you think that its "fine" for a company to design a product to physically break right atfer warranty by being used as intended. You need to go back to primary school and learn about ethics and morals from scratch, again.

[79wjd]

29 minutes ago, Prysin said:

so you want to say.

"its fine, just dump another 200-300 USD for a custom loop onto the card to solve Nvidia's fucking shitty job"

Come on. If that is your "solution" you can just fuck off patrick. That is the shitties most fanboyish excuse ive ever heard from you.

 

Seriously patrick, if you think that its "fine" for a company to design a product to physically break right atfer warranty by being used as intended. You need to go back to primary school and learn about ethics and morals from scratch, again.

Not that I agree with what he said, but a reference 290x isn't going to last a particularly long time either. And I'm arguably okay with that. Reference coolers all basically suck. It would be nice if they didn't, but also somewhat unrealistic as they'd likely hurt aftermarket cooler sales too much (although there aren't any aftermarket coolers for the Titans....there should be though).

[Notional]

27 minutes ago, Prysin said:

so you want to say.

"its fine, just dump another 200-300 USD for a custom loop onto the card to solve Nvidia's fucking shitty job"

Come on. If that is your "solution" you can just fuck off patrick. That is the shitties most fanboyish excuse ive ever heard from you.

 

Seriously patrick, if you think that its "fine" for a company to design a product to physically break right atfer warranty by being used as intended. You need to go back to primary school and learn about ethics and morals from scratch, again.

That's the exact same argument he had when Titan Z got destroyed by the 295x2. The card was so hot and the cooler so bad, that the chips ran at a measly 850mhz (or so). His argument was to take this 3k$ card and add a water block and overclock it for it to even be near the half priced 295x2. Sometimes it sounds like NVidia is Scientology, and nothing they do is ever bad.

[Prysin]

3 minutes ago, djdwosk97 said:

Not that I agree with what he said, but a reference 290x isn't going to last a particularly long time either. And I'm arguably okay with that. Reference coolers all basically suck. It would be nice if they didn't, but also somewhat unrealistic as they'd likely hurt aftermarket cooler sales too much (although there aren't any aftermarket coolers for the Titans....there should be though).

EVGA has TitanX ACX2.0+ coolers for sale and has had them for sale since the TitanX came out

[Nineshadow]

Sorry, I just couldn't resist.

[Sharkyx1]

4 hours ago, patrickjp93 said:

Too bad, because you're already wrong. The Tesla P100s have been sampling since December and will be shipping at the end of Q3 as Nvidia has already stated. Now, that is a cut down core, and it is enterprise, but yes, in fact Nvidia will release a 600mm sq. GPU this year.

If I said graphics card would  you accept my post?  on the enterprise side nvidia will just tank the yields and price accordingly, they can't sell titans for 6k usd though. 

 

I don't doubt the chance that nvidia might impress me and release earlier than I expect, but do you think we'll see a shipping 610mm graphics card in 2016?

[niofalpha]

Revealed Nvidia GP100 Anal

What it said on the main page...

 

But in all seriousness the power draw is a bit high for my liking... I wonder how temps are goona look

[SurvivorNVL]

Thank god it should be a beast of a card, but I will still sit on this 970 until Volta, I think.  Volta and OLED, but damn it should be nice.  And hell bring on more watts!

[79wjd]

2 hours ago, Prysin said:

EVGA has TitanX ACX2.0+ coolers for sale and has had them for sale since the TitanX came out

I thought board partners weren't allowed to release modified coolers. Well okay then. 

[awesomeness10120]

25 minutes ago, djdwosk97 said:

I thought board partners weren't allowed to release modified coolers. Well okay then. 

It's actually just a fan and backplate. You tear apart the Titan X and put the ACX cooler on.

 

[patrickjp93]

3 hours ago, Prysin said:

so you want to say.

"its fine, just dump another 200-300 USD for a custom loop onto the card to solve Nvidia's fucking shitty job"

Come on. If that is your "solution" you can just fuck off patrick. That is the shitties most fanboyish excuse ive ever heard from you.

 

Seriously patrick, if you think that its "fine" for a company to design a product to physically break right atfer warranty by being used as intended. You need to go back to primary school and learn about ethics and morals from scratch, again.

An enthusiast card for enthusiast use cases. AMD does exactly the same quality of work but throws a liquid cooler on it. The VRMs between the Titan X and 295x2 are the exact same components.

 

It's perfectly ethical. It's up to businesses to be honest and consumers to be discerning.

[patrickjp93]

29 minutes ago, djdwosk97 said:

I thought board partners weren't allowed to release modified coolers. Well okay then. 

Not modified whole packages*