Tesla P100 — 5.4 DP TeraFLOPS — Pascal

[Mark Rose]

http://www.nvidia.com/object/tesla-p100.html

It's a beast!

• 5.3 TeraFLOPS double-precision performance
• 10.6 TeraFLOPS single-precision performance
• 21.2 TeraFLOPS half-precision performance
• 160 GB/s interconnect bandwidth with NVIDIA NVLink
• 720 GB/s memory bandwidth with CoWoS HBM2 Stacked Memory
• 16 GB of CoWoS HBM2 Stacked Memory

[firejuggler]

ok, what is the speed of a titan, to compare?

[mackerel]

From wikipedia:
Titan X (Maxwell): SP 6.1, DP 0.19
Titan Black (Kepler): SP 5.1, DP 1.7

[firejuggler]

Digging thru the web, I found that Nvidia now offer a rack with 8 Tesla 100, named DGX-1 for 129k$
http://www.engadget.com/2016/04/05/n...supercomputer/

NVIDIA has moved its GPU business far beyond just powering gaming computers. For several years now, the company has been focused on using its technology to power advanced machine learning technology, and today the company announced the positively wild $129,000 DGX-1 supercomputer. NVIDIA CEO Jen-Hsun Huang called the DGX-1 the equivalent of 250 servers in a box -- it has 7TB of SSD storage, eight of the just-announced Tesla P100 GPUs and two Xeon processors.

[tServo]

Originally Posted by Mark Rose

It's a beast!]

Indeed it is, Mark.
Also worth noting:
This is a Tesla board.
It has 3584 cuda cores.
Its speed is 1328 Mhz, boost is 1480 Mhz.
It only (!) draws 300 watts.
The memory bus width is 4096 bits !!!!! ( Previous highest was 384 bits ).
This thing from another universe has over 15 Billion transistors ( more than double Kepler ).

Zounds!

[ldesnogu]

This beast has the potential to crush Intel upcoming Knight Landing!

AnandTech has a chart comparing Pascal to Kepler and Maxwell.

[ixfd64]

The consumer cards will be much more cost-effective unless you're planning to do LL testing or P-1 factoring.

[ATH]

nevermind, I see it is not a normal card but an "accelerator"

[Madpoo]

Originally Posted by firejuggler

Digging thru the web, I found that Nvidia now offer a rack with 8 Tesla 100, named DGX-1 for 129k$
http://www.engadget.com/2016/04/05/n...supercomputer/

That seems... high. Averaging $16,250 USD per Tesla 100 card? Are they really expected to sell for that much individually? Probably not, meaning the advantage is having it in one dual CPU server with all the gee-whiz cooling and what-nots built in, but that seems like a hefty premium.

Whatever... you know that no matter the price, some government agency is going to buy a bunch and send taxpayers the bill.

[TheJudger]

Hi,

https://devblogs.nvidia.com/parallel...ures-revealed/

CUDA 8 will be available in August 2016 and there will be a release candidate available around June.

IIRC correctly CUDA 7.5 can't create code for CC 6.x.


-----

Originally Posted by tServo

The memory bus width is 4096 bits !!!!! ( Previous highest was 384 bits ).

Doesn't really matter but that is not true... GT200(b) has 8x64b.

Oliver

[Mark Rose]

This talk goes over the new architecture in detail:

http://on-demand.gputechconf.com/gtc...deo/S6176.html