Results 1 to 7 of 7

Thread: CPU to GPU memory bottleneck

  1. #1
    Member
    Join Date
    May 2014
    Location
    Canada
    Posts
    141
    Thanks
    68
    Thanked 21 Times in 12 Posts

    CPU to GPU memory bottleneck

    GPU codec performance gets ruined by slow memory transfer between CPU and GPU over PCIe.

    Three new chip architectures on the horizon promise to reduce the CPU to GPU memory bottleneck:

    1) Intel Broadwell R chip with integrated GPU and 128 MB L4 cache: this cache can be accessed by the GPU - release scheduled for Christmas 2014

    2) AMD Kaveri APU: Heterogenous System Architecture and hUMA hetergenous unified memory architecture: both CPU and GPU have access to entire memory space (up to 32 GB) - available now

    3) Intel Knights Corner CPU - 16GB of DRAM 3D stacked on-package, providing up to 500GB/sec of memory bandwidth (along with up to 384GB of DDR4-2400 mainboard memory) - available in 2015

    # 3 will be pretty pricey, but #2 looks very promising and should be much cheaper.


  2. #2
    The Founder encode's Avatar
    Join Date
    May 2006
    Location
    Moscow, Russia
    Posts
    4,009
    Thanks
    399
    Thanked 397 Times in 152 Posts
    Most likely, the Broadwell release will be delayed... (Q2 2015?)

  3. #3
    Member
    Join Date
    Jun 2009
    Location
    Kraków, Poland
    Posts
    1,488
    Thanks
    26
    Thanked 130 Times in 100 Posts
    The solution would be to keep data on the GPU as long as possible to reduce transfer.

    But hUMA from AMD also works and is e.g. present in XB1 and PS4 consoles, so that's pretty large userbase already established.

  4. #4
    Member
    Join Date
    May 2014
    Location
    Canada
    Posts
    141
    Thanks
    68
    Thanked 21 Times in 12 Posts
    Quote Originally Posted by Piotr Tarsa View Post
    But hUMA from AMD also works and is e.g. present in XB1 and PS4 consoles, so that's pretty large userbase already established.
    Yes, I think this is the way to go. I just worry about AMD staying in business: they have brutal competition from nVidia and Intel.
    Bitcoin mining is not going to carry them forever.

  5. #5
    Member
    Join Date
    Feb 2013
    Location
    San Diego
    Posts
    1,057
    Thanks
    54
    Thanked 72 Times in 56 Posts
    That's always been the story about AMD.

    Intel has a history of cancelling ambitious projects (Larrabee?).

  6. #6
    The Founder encode's Avatar
    Join Date
    May 2006
    Location
    Moscow, Russia
    Posts
    4,009
    Thanks
    399
    Thanked 397 Times in 152 Posts
    BTW, Intel Core i7-4770R with 128 MB L4 and Iris Pro 5200 graphics now available! It is hard to find thing, however Gigabyte BRIX Pro has it!

  7. #7
    Member
    Join Date
    May 2014
    Location
    Canada
    Posts
    141
    Thanks
    68
    Thanked 21 Times in 12 Posts
    Quote Originally Posted by encode View Post
    BTW, Intel Core i7-4770R with 128 MB L4 and Iris Pro 5200 graphics now available! It is hard to find thing, however Gigabyte BRIX Pro has it!
    Sounds nice!

Similar Threads

  1. Fastest in memory compression c++
    By JangoFatXL in forum Data Compression
    Replies: 29
    Last Post: 10th April 2014, 22:03
  2. Memory-mapped files
    By Bulat Ziganshin in forum The Off-Topic Lounge
    Replies: 0
    Last Post: 22nd March 2012, 19:54
  3. CLI memory teste needed
    By SvenBent in forum The Off-Topic Lounge
    Replies: 7
    Last Post: 21st April 2010, 08:06
  4. Memory Limit?
    By Earl Colby Pottinger in forum Data Compression
    Replies: 28
    Last Post: 11th April 2010, 03:59
  5. 2G+ memory blocks
    By Shelwien in forum Data Compression
    Replies: 0
    Last Post: 6th March 2009, 02:13

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •