Nvidia details next-gen GPU platform, announces $3K card

Nvidia's (NVDA +0.3%) promises its Pascal GPU architecture, the successor to its recently-launched Maxwell, will deliver a 5x-12x improvement in GPU-to-CPU and GPU-to-GPU transfer speeds by leveraging the chipmaker's NVLink interconnect. Pascal, due in 2016, will also support 3D memory chips, something it expects will improve densities by 2.5x and memory energy efficiency by 4x.

Pascal will underpin mobile processors and HPC GPUs in addition to PC GPUs. CEO Jen-Hsun Huang goes as far as to say it will enable new machine learning/A.I. apps.

Nvidia, which is hosting both its investor day (webcast) and its annual GPU conference (live blog) today, has also: unveiled Titan Z, a flagship dual-GPU graphics card solution (relies on Nvidia's Kepler architecture) that will retail for a whopping $3K. Nvidia's current flagship (the GTX Titan Black) only goes for $999, and AMD's top-of-the-line R9 290X (previous) has a $549 MSRP.

Nvidia has been growing GPU sales in spite of a weak PC market with the help of strong HPC and PC gaming market demand. Likely helping: The company has been ramping R&D spend (+13% Y/Y in the Jan. quarter) at a time when AMD has been cutting costs and directing more resources towards embedded/console opportunities.

From other sites
Comments (7)
  • User 12115671
    , contributor
    Comments (2735) | Send Message
    NVlink will be interesting for the enterprise and supercomputer worlds that use PowerPC, but this will obviously not be incorporated into x86. Can't wait for the Titan-Z, dual Hawaii showdown....
    25 Mar 2014, 03:06 PM Reply Like
  • e92m3
    , contributor
    Comments (179) | Send Message
    Indeed, though it is still using pci-e for GPU-CPU communication and offloading GPU-GPU to nvlink.


    There is no aggregate bandwidth increase with regards to GPU-CPU interconnect due to nvlink.


    While not using the same interconnect for GPU-GPU and GPU-CPU will free some bandwidth, nvidia is implying an increase in maximum CPU-GPU bandwidth, which clearly isn't the case according to their own diagrams. That's why they imply rather than state ;)
    25 Mar 2014, 05:13 PM Reply Like
  • e92m3
    , contributor
    Comments (179) | Send Message
    Sorry, I should clarify that on some specific POWER configurations they may possibly increase aggregate bandwidth (their diagrams of links moving about with little technical explanation aren't all that helpful), though the increase will not be applicable to x86, nor the majority of nvidia customers.


    Additionally, the next pci-e spec will most likely be released or nearly release by the time nvlink is actually in POWER products, reducing the aggregate bandwidth advantage of the proposed 'nvlink' considerably and possibly exceeding it.
    26 Mar 2014, 04:05 PM Reply Like
  • 3D Investing
    , contributor
    Comments (1552) | Send Message
    Pascal is due in 2016! this is tech development. That's 2 generations worth of time.


    Titan Z would have almost no impact on their financials as only a handful will be sold.
    25 Mar 2014, 03:13 PM Reply Like
  • e92m3
    , contributor
    Comments (179) | Send Message
    Keep in mind that nvidia falsified the quoted mantle results, is still not achieving competitive minimums, used deceptive graph scaling to make 4 fps appear significant (probably learned this trick from Anad), purposely chose the least bottlenecked configuration they could (let's see crossfire vs. sli on the same testbed nvidia...) and have pushed UMA (still not comparable to HUMA in the slightest) to 2016...All this to help save face and explain why they are having difficulty competing with GPUs that use significantly less die space.


    Obviously, some are deluded by their careful (deceitful) marketing.


    I care not for the future promises of nvidia, as they are more often than not empty.


    I love the JHH quote where he claims he only hypes real, existing products instead of future products. It's like he doesn't even remember the rubbish he has spoken just a few years ago.


    Just checked... nvidia is using FALSE mantle AND dx 'results' for the 290x. Unacceptable, all of you can verify that they are indeed lying with just few minutes of research.
    25 Mar 2014, 03:37 PM Reply Like
  • Ashari
    , contributor
    Comments (245) | Send Message
    I'm sure that it's just pure coincidence (cough, cough) but as per their settlement with Nvidia Intel will no longer have to support PCI-Express in 2016. Obviously PCIe isn't going anywhere but Intel will be able to make life very difficult for Nvidia and with the advent of the APU they'll have a good reason to do so.
    26 Mar 2014, 08:47 AM Reply Like
  • flashjordan23
    , contributor
    Comments (46) | Send Message
    Jen hsun is a marketer. I am sure some blogs and lapdog analyst are lapping it up.
    It is part f the job really, we will see in a year or 2 how good Rory read is at the 'art of war'.
    26 Mar 2014, 10:03 AM Reply Like
DJIA (DIA) S&P 500 (SPY)
ETF Screener: Search and filter by asset class, strategy, theme, performance, yield, and much more
ETF Performance: View ETF performance across key asset classes and investing themes
ETF Investing Guide: Learn how to build and manage a well-diversified, low cost ETF portfolio
ETF Selector: An explanation of how to select and use ETFs