Zero copy matters

Article: AMD Fusion Architecture and Llano
By: Groo (, August 25, 2011 2:11 pm
Room: Moderated Discussions
David Kanter ( on 8/25/11 wrote:
>Gionatan Danti ( on 8/24/11 wrote:
>>Hi David, thank you for this great article.
>>I would pose you a question about zero copy: from my understanding, a discrete
>>graphic card can use texture directly from system memory through, without the need
>>to copy them to local memory. This is achieved using the >GART mapping table found on AGP and PCI-E specifications.
>I am not familiar with this, but it sounds reasonable.
>>Sure this kind of texture access is way lower than accessing local memory, but
>>(to me) closely resemble the zero-copy concept.
>As you described it, yes, this is very similar to the zero copy that Intel and AMD are implementing.
>>So, when AMD presents zero-copy as a Llano's innovations, what exactly means? It
>>regard a zero-copy capability between CPU (system) memory and GPU-reserved memory?
>>It regard both graphic mode and compute mode?
>There is a fairly important distinction from a physical perspective (i.e. what
>the electrons and bits are doing), but they seem to be very similar from a logical (i.e. programmers) perspective.
>The right way to understand the differences is to look at the actual data flow
>for a read operation. Here I use --> to indicate on-die data flow and ==> to indicate off-die data flow.
>Zero copy discrete:
>GPU read-->PCI-E ==> CPU memory controller ==> DRAM
>The important part here is that you are using PCI-E as an external interface.
>The data must flow from the CPU die to the GPU, using PCI-E which costs latency, power and bandwidth.
>Llano zero copy:
>GPU read-->CPU/GPU memory controller==>DRAM
>Sandy Bridge zero copy:
>GPU read-->L3 cache OR
>GPU read-->CPU memory controller==>DRAM
>If you look at Llano, their memory controller can theoretically read 30GB/s (2
>* 8B * 1.866GT/s). The fastest PCI-E interface is theoretically 8GB/s in a single
>direction (2B * 4GT/s). Intel's L3 cache bandwidth to the GPU is ~100GB/s (32B * ~3GT/s).
>So the key to zero copy for AMD or Intel GPUs is that they eliminate an off-die data transfer and achieve:
>1. Vastly higher bandwidth (3-4X for memory, 12X for cache)
>2. Lower latency
>3. Lower power consumption
>So they are similar, but the difference is that using zero copy on a discrete GPU
>tends to lower performance by reducing bandwidth. Zero copy on an integrated GPU
>raises performance, power efficiency, etc.
I am pretty sure that zero copy doesn't copy anything, just tweaks the pointer to memory with the MMU. If so, it seems way more efficient than Intel's scheme. That said, I might be thinking of something that hasn't been released yet. :)

< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
AMD Fusion and Llano article onlineDavid Kanter2011/06/28 08:46 AM
  AMD Fusion and Llano article onlinebakaneko2011/06/28 10:39 AM
    AMD Fusion and Llano article onlineGabriele Svelto2011/06/28 10:52 AM
      AMD Fusion and Llano article onlinebakaneko2011/06/28 11:35 AM
    CostDavid Kanter2011/06/29 01:02 AM
      Costbakaneko2011/06/29 02:52 AM
      A GDDR5 DIMMs future?WillNKate2011/06/29 09:23 AM
        A GDDR5 DIMMs future?Dan Fay2011/06/29 09:28 AM
          A GDDR5 DIMMs future?Gabriele Svelto2011/06/29 09:41 AM
            A GDDR5 DIMMs future?Dan Fay2011/06/29 10:20 AM
          A GDDR5 DIMMs future?Robert Williams2011/06/29 12:51 PM
            A GDDR5 DIMMs future?David Hess2011/06/29 05:47 PM
          A GDDR5 DIMMs future?Megol2011/06/30 01:53 AM
        A GDDR5 DIMMs future?David Kanter2011/06/29 10:28 AM
        A GDDR5 DIMMs future?Shankhadeep Shome2011/07/01 06:08 PM
          A GDDR5 DIMMs future?EduardoS2011/07/01 08:05 PM
            A GDDR5 DIMMs future?Shankhadeep Shome2011/07/02 05:18 AM
            A GDDR5 DIMMs future?Mark Hahn2011/07/02 09:56 AM
              A GDDR5 DIMMs future?David Hess2011/07/02 12:15 PM
          A GDDR5 DIMMs future?Dan Fay2011/07/02 12:24 PM
            A GDDR5 DIMMs future?EduardoS2011/07/02 03:02 PM
            Zero copy mattersDavid Kanter2011/07/02 03:46 PM
              Zero copy mattersDan Fay2011/07/02 03:59 PM
                Zero copy mattersDavid Kanter2011/07/02 08:46 PM
                  Zero copy mattersGionatan Danti2011/08/24 06:31 AM
                    Zero copy mattersDavid Kanter2011/08/25 11:48 AM
                      Zero copy mattersGroo2011/08/25 02:11 PM
                        Zero copy mattersanon2011/08/25 06:04 PM
                        Zero copy mattersDavid Kanter2011/08/25 10:29 PM
                          Zero copy mattersgallier22011/08/26 12:19 AM
                            Zero copy mattersDavid Kanter2011/08/26 03:09 AM
                              Zero copy mattersGroo2011/08/26 08:11 AM
                                Zero copy mattersgallier22011/08/26 08:16 AM
                                Zero copy mattersanon2011/08/26 08:24 PM
                                Zero copy mattersDavid Kanter2011/08/28 10:38 PM
                                  Zero copy mattersgallier22011/08/29 12:40 AM
                                    Sharing a cacheDavid Kanter2011/08/29 09:35 AM
                                      Sharing a cacheAnon2011/08/30 02:08 AM
                                        nVidia Fermi has ~192KB on-chip memory per SMMark Roulo2011/08/30 08:32 AM
                                          nVidia Fermi has ~192KB on-chip memory per SMEduardoS2011/08/30 02:19 PM
                                            GPU and CPU/GPUDavid Kanter2011/08/31 10:07 AM
                                              FB data in tile based renderersRohit2011/08/31 07:10 PM
                                              GPU and CPU/GPUAnon2011/09/02 03:05 AM
                                            nVidia Fermi has ~192KB on-chip memory per SMGionatan Danti2011/09/01 02:30 AM
                                              nVidia Fermi has ~192KB on-chip memory per SMAnon2011/09/02 02:45 AM
                                                nVidia Fermi has ~192KB on-chip memory per SMGionatan Danti2011/09/02 07:06 AM
                                                  nVidia Fermi has ~192KB on-chip memory per SMGabriele Svelto2011/09/02 09:00 AM
                                                    nVidia Fermi has ~192KB on-chip memory per SMGionatan Danti2011/09/02 09:18 AM
                                                    nVidia Fermi has ~192KB on-chip memory per SMEduardoS2011/09/02 01:57 PM
                                                nVidia Fermi has ~192KB on-chip memory per SMRohit2011/09/02 01:18 PM
                                        Sharing a cacheDavid Kanter2011/08/30 04:49 PM
                                          Sharing a cacheGabriele Svelto2011/08/31 07:50 AM
                                        Stream processing is deadBrett Davis2011/08/30 06:53 PM
                                          Stream processing is deadWill Smith2011/08/31 05:00 AM
                                            GPUs and cachingDavid Kanter2011/08/31 02:28 PM
                                              GPUs and cachingBrett Davis2011/08/31 07:15 PM
                            Zero copy mattersGroo2011/08/26 08:15 AM
                        "Zero copy" just means "One copy less than before" (NT)iz2011/08/26 01:40 AM
                      Zero copy mattersGionatan Danti2011/08/25 11:52 PM
  AMD Fusion and Llano article onlineNiceone2011/06/28 02:10 PM
    AMD Fusion and Llano article onlineDavid Kanter2011/06/29 12:55 AM
  I believe I found a typo.I.S.T.2011/06/28 06:47 PM
    Good catchDavid Kanter2011/06/29 12:53 AM
      Good catchI.S.T.2011/06/29 10:43 AM
        Thanks!David Kanter2011/06/29 10:53 AM
  AMD Fusion and Llano article onlineR_Mehta2011/06/28 11:47 PM
  AMD Fusion and Llano article onlineHoward Chu2011/06/29 05:23 AM
    AMD Fusion and Llano article onlinean2011/06/29 07:51 AM
    AMD Fusion and Llano article onlineNiceone2011/06/29 08:57 AM
      AMD Fusion and Llano article onlineDavid Kanter2011/06/29 12:29 PM
    AMD Fusion and Llano article onlineDavid Kanter2011/06/29 10:26 AM
  AMD Fusion and Llano article onlineGeertB2011/06/29 07:29 AM
  AMD Fusion and Llano article onlineDan Fay2011/06/29 09:30 AM
  TrinityAzazel2011/07/03 10:15 PM
    TrinityAntti-Ville Tuunainen2011/07/04 03:18 AM
      TrinityAzazel2011/07/04 06:56 AM
        TrinityDavid Hess2011/07/04 08:07 AM
          Trinitysomeone2011/07/04 10:44 AM
            TrinityDavid Hess2011/07/04 02:06 PM
              Trinitysomeone2011/07/05 07:25 AM
    TrinityMatt Waldhauer2011/07/04 11:22 AM
      TrinityAzazel2011/07/04 09:32 PM
        Trinity - clock/power gating and carsMatt Waldhauer2011/07/05 01:03 AM
          Trinity - clock/power gating and carsDan Fay2011/07/05 07:07 PM
    TrinityDavid Kanter2011/07/05 02:05 AM
Reply to this Topic
Body: No Text
How do you spell green?