facebook rss twitter

Nvidia launches the Pascal GPU-based TITAN X

by Mark Tyson on 22 July 2016, 09:34

Tags: NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qac4uq

Add to My Vault: x

Nvidia has just launched what it refers to as "The Ultimate" graphics card. The new Nvidia TITAN X was unveiled by Nvidia CEO, Jen-Hsun Huang, at a Stanford University event last night, to an audience of over 500 academics, researchers, and students. According to the Nvidia blog, this graphics card is the result of a bet between Nvidia's CEO and its top hardware engineer Brian Kelleher, over whether the green team could squeeze 10 teraflops of computing performance from a single chip. It sounds like the Nvidia CEO was doubtful whether it could be achieved, as he now owes Kelleher a dollar…

Diving into the hardware spec, Nvidia says the new TITAN X is based upon the largest GPU ever built, with a record breaking 3,584 CUDA cores to do the owner's bidding. The GPU cores run at 1.5GHz, providing "11 TFLOPS of brute force". Backing up the powerful GP102 GPU, Nvidia has installed 12GB of GDDR5X memory, "one of the fastest memory technologies in the world," rather than HBM2 memory as installed on its Tesla P100 and GP100 products.

The full bullet pointed new TITAN X spec list, as provided by Nvidia is as follows:

  • 11 TFLOPS FP32
  • 44 TOPS INT8 (new deep learning inferencing instruction)
  • 12B transistors
  • 3,584 CUDA cores at 1.53GHz (versus 3,072 cores at 1.08GHz in previous TITAN X)
  • Up to 60% faster performance than previous TITAN X
  • High performance engineering for maximum overclocking
  • 12GB of GDDR5X memory (480GB/s)

Nvidia says that the new TITAN X will be available from 2nd August for $1,200 direct from nvidia.com in North America and Europe, or via select system builders. Customers in Asia will also be able to acquire a card "soon".



HEXUS Forums :: 40 Comments

Login with Forum Account

Don't have an account? Register today!
Just what I need for my new 8K TV.
So thats £1200 then, or worse if the pound was to continue falling.
i can't afford this…oh!well ! note7 will be my choice…
i was thinking the GP102 GPU would be the Nvidia 1080 Ti & come out around October for Battlefield 1 as a show case and still no HBM2 ?! it would have been better for nvidia to delay this until there was more HBM2 and then this would have been a real killer of a card. Titan X FLOP XD
To say the claimed performance is disappointing is an understatement. Especially given the price…

They claim up to 60% over the old Titan X (the naming is also daft!), and given the 1080 is already 37% over the old Titan X according to TPU, that makes the new Titan X just 16% faster than the 1080? Seriously??

Even assuming perfect scaling and comparing the TFLOPS ratings you only get about 22% for shader-bound applications.

I do wonder what is going on with the GP102 though - it has a smaller transistor count than the GP100 so it does look like it's actually a different die but the partitioning is weird, we don't have 6 full GPCs here as you'd expect so I wonder if they've rearranged the GPCs vs the smaller dies or if this isn't fully enabled - not all that unlikely given it's such a large die on a new node.

I wonder where that leaves the 1080Ti though? There's not much performance room between the 1080 and the TX as it is. Maybe we'll see a 780Ti-esque release with a full die Ti some time down the line when yields improve?

WRT yields and stuff though - I wonder how this works out financially? Having a separate huge die for such a low volume market is weird given how expensive 16nm development is!

Edit: Not sure where I got 37% for OTX>1080, must have pressed a wrong button somewhere along the line, it should be ~32%. I'm getting more like 21% for 1080>NTX which is in-line with the quoted shader performance.