'

Nvidia's Tesla P100 does not use the full Pascal GPU core

Nvidia's Tesla P100 does not use the full Pascal GPU core

Nvidia Announces Pascal P100 Volume Production

Nvidia's Tesla P100 does not use the full Pascal GPU core

 

Nvidia's Tesla P100 does not use the full Pascal GPU core, with 256 GPU cores being left out of the design. This many mean that "Big Pascal" may be able to offer a full 11.3TFLOPs of single precision performance. 

Right now Nvidia has been very clear that the  Pascal-based Tesla P100 uses only 56 SM units, but when looking at Nvidia's pascal GP100 core design we can see that there is a total of 60 SMs available, meaning that there are a full 256 Pascal GPU cores that will be unavailable in the Nvidia P100. 

 

Nvidia's Tesla P100 does not use the full Pascal GPU core  

 

Nvidia may not be using the full GP100 pascal GPU core in the Tesla P100 for many reasons, but it is very likely that it is due to yield issues with the new GPU design. The Nvidia GP100 core is the largest that I have ever seen from Nvidia, with a die size that is 610mm squared, making it larger than even the Titan X's 601mm squared core design. This GPU is also the first Nvidia GPU that has been shown that was made using the TSMC 16nm FinFET processing node and the first to use HBM 2.0 memory, making this new GPU core design much more complex to manufacture.    

This GPU is also the first Nvidia GPU that has been shown that was made using the TSMC 16nm FinFET processing node, which is a new processing node from TSMC, and is the first Nvidia GPU to use HBM 2.0 memory, making this new GPU core design much more complex to manufacture.    

With HBM 2 being a larger size than HBM 1 per chip and Nvidia's GP100 GPU core being so large it looks like it will be a very difficult task to fit all of this hardware on a single interposer, likely stretching the current tolerances of this technology to their limits. Just look at the picture below to see how close the four HBM 2.0 chips are to the GPU core if you want to see how tightly packed this GPU design is. 

 

Nvidia's Tesla P100 does not use the full Pascal GPU core  

Below we have looked at the specifications of Nvidia's Tesla P100 and Nvidia's GP100 core design to list the specifications of Nvidia's full GP100 GPU core though many of the specifications remain unknown at this point. 

Right now we do not know if Nvidia will use larger capacity HBM 2 chips on GP100 in the future, as we do know that 8GB HBM chips are due to be manufactured this summer allowing a potential 32GB GP100 based GPU to be released in the future. 

We also cannot comment on what clock speeds will be possible in future GPUs using the GP100 GPU core, as we do not know if Nvidia's stated 1328MHz/1480MHz base and boost clocks are conservative on Nvidia's part or not. For all we know later Pascal chips could run at 1600MHz+ boost in the future if the chip overclocks well. 

 

 Tesla K40Tesla M40Tesla P100Full GP100
GPU CoreGK110 (Kepler)GM200 (Maxwell)GP100 (Pascal)GP100 (Pascal)
SM Units15245660
Cores Per SM Unit1921286464
FP32 CUDA Cores2880307235843840
Base Clock Speeds745MHz948MHz1328MHz-
Boost Clock Speeds810/875Mhz1114MHz1480MHz-
Texture Units240192224240
VRAM CappacityUp to 12GBUp to 24GB16GB-
Memory Interface384-bit384-bit4096-bit4096-bit
Transistors7.1 billion8 billion15.3 billion15.3 billion
Die Size551mm^2601mm^2610mm^2610mm^2
Manufacturing Process28nm28nm16nm16nm

 

Right now Nvidia states that the P100 is in volume production and that the chip will be selling first to the supercomputing market in June 2016 and will be arriving with OEMs in Q1 2017.  

Nvidia has stated that they are not currently planning on releasing the GP100 GPU core in a Geforce GTX part but given the fact that the GK110 and GM200 GPU cores eventually became GTX Titan series GPUs and later GTX X80 Ti series GPUs it seems likely that this GPU will eventually hit the market as a future flagship GPU when production yields and volume make it affordable. 

 

You can join the discussion on Nvidia's full GP100 GPU core specifications on the OC3D Forums.  

 

«Prev 1 Next»

Most Recent Comments

07-04-2016, 11:01:42

NeverBackDown
You can expect the consumer version to be much more toned down. Fully enabled cores but no HBM2 as well as dropping NVlink. They more than likely only went all out because the servers/cloud/supercomputer market for them is much more profitable. No consumer could afford a GPU with this many features. Dropping HBM2/NVlink and I think we have our TX but with some more cores.Quote

07-04-2016, 14:42:16

Kaapstad
Probably down to yields on early production.

There could be a full fat version later with everything enabled.

HBM2 and NVlink will probably stay as it would mean redesigning the chips to get rid of them.

Titan card will probably have all the cores enabled and the Ti version will possibly have specs similar to the OP but with higher clockspeeds.

Don't think gamers will see any of this until 2017.Quote

07-04-2016, 17:36:58

NeverBackDown
Quote:
Originally Posted by Kaapstad View Post
Probably down to yields on early production.

There could be a full fat version later with everything enabled.

HBM2 and NVlink will probably stay as it would mean redesigning the chips to get rid of them.

Titan card will probably have all the cores enabled and the Ti version will possibly have specs similar to the OP but with higher clockspeeds.

Don't think gamers will see any of this until 2017.
Well they could drop it. Don't think NVlink is part of the core. It's on a separate pcb that they can drop the die in. If HBM2 stays, late 2016/early 2017 is when it will launch. If they decide to have a separate consumer build then expect GDDR5(or 5X depending on launch date) and probably costing around $1000-1500)

I think this time they will make a bigger gap between Titan and Ti. They should at least, people will pay for it and make a bigger profit so business wise it would work.Quote
Reply
x

Register for the OC3D Newsletter

Subscribing to the OC3D newsletter will keep you up-to-date on the latest technology reviews, competitions and goings-on at Overclock3D. We won't share your email address with ANYONE, and we will only email you with updates on site news, reviews, and competitions and you can unsubscribe easily at any time.

Simply enter your name and email address into the box below and be sure to click on the links in the confirmation emails that will arrive in your e-mail shortly after to complete the registration.

If you run into any problems, just drop us a message on the forums.