We're gradually importing our scientific contents in French. Please forgive the visual glitches that may result from the operation.
Is the brand new NVIDIA TitanX the new reference for deep learning applications?
By   |  July 25, 2016

NVIDIA has announced the release of its latest ultra-high-end graphics card, the Titan X. The legend about this card says it all started with a bet. Brian Kelleher, Senior Vice President of GPU Engineering launched the challenge with the company CEO, Jen-Hsun Huang: “can we get more than 10 teraflops of computing performance from a single GPU?” Jen-Hsun thought it was crazy.

Well, they did. And since this morning, it is said that Jen-Hsun owes Brian a dollar. The new Titan X already reaches 11 TFLOPS and it might go even further as drivers get optimized in the upcoming weeks. As expected, NVIDIA Titan X is based on the new Pascal architecture. Its GPU, the GP102 is the largest ever built and it has the record of 3,584 CUDA cores but, actually, most of the specifications displayed by this card are staggering:

  • 12 billion transistors
  • 11 TFLOPS FP32 (32-bit floating point)
  • 44 TOPS int8 (new instruction for inference dedicated to deep learning)
  • 3584 CUDA cores at 1.53GHz (compared with 3072 CUDA cores at 1.08GHz in the previous TITAN X)
  • High performance engineering for maximum overclocking
  • Memory 12GB GDDR5X @1250 MHz
  • Memory Bandwith 480GB/s

And more comes with the price. Announced by NVIDIA at 1200$ per unit, this new Titan X might quickly become the new reference for Deep Learning applications. And that’s where we are a little confused about NVIDIA’s strategy as the company has long been promoting the DGX-1 for this specific area. From a sheer specs standpoint, the DGX-1 system provides 85 TFLOPS (FP32) for a $129000 price tag, which gives us a little more than $1500 per TFLOP where the Titan X is around $100. Even considering the price of the setup to house the cards – around $2500-$3000 for a system able to run 4 to 8 cards – a Titan cluster could be up to 90% cheaper for the same performance. That being said, Titan clusters are somewhat less scalable than DGX-1 systems, but for small and medium organizations focused on Deep learning, this new graphics card available August 2nd will probably become a must have.
 
We are not sure yet of the objectives pursued by NVIDIA with this new offering. Some think they may be starting a new product family as this is the first of his graphics card not labelled “GTX” all the more as the product will be sold directly by the company. What remains to be seen now is how the channel, especially TPPs (Tesla Preferred Partners), will welcome the product.

© HPC Today 2017 - All rights reserved.

Thank you for reading HPC Today.

Express poll

Do you use multi-screen
visualization technologies?

Industry news

S01-170208

Brands / Products index