I use ALUs

Article: Impressions of Kepler
By: anon (no.delete@this.thanks.com), March 23, 2012 2:07 pm
Room: Moderated Discussions
Mark Roulo (nothanks@xxx.com) on 3/23/12 wrote:
---------------------------
>anon (no@thanks.com) on 3/23/12 wrote:
>---------------------------
>>David Kanter (dkanter@realworldtech.com) on 3/23/12 wrote:
>>---------------------------
>>>Please do not call execution units cores. That's Nvidia's rather confusing (and wrong) marketing-speak.
>>>
>>>Kepler has 1.5K execution units, but only 8 cores. A core needs to be able to fetch
>>>instructions, decode instructions, schedule them (unless it's statically scheduled),
>>>execute instructions, access a data cache and write results back to a register file.
>>
>>I definitely agree that NVidia's marketing speak is wrong. But I'm not convinced
>>that an "SMX" couldn't be argued to consist of multiple cores. For instance, the
>>techreport review claims that there are 4 schedulers inside an SMX, each tied to
>>a distinct set of 48 execution units. These 4 apparently share an instruction and
>>L1 data cache, but it doesn't seem impossible that they could be issuing instructions
>>from different kernels. Note - I have no idea whether that's the case or not - there
>>are conflicting descriptions of SMX internals in other reviews.
>>
>>Any chance of shedding some more light on that?
>
>I try to talk about ALUs.

Makes sense.

>A Kepler has 1.5K for single precision floating point loads.
>
>An 8-core SandyBridge has 8x8= 64 for single precision floating point loads.
>
>We *still* have to adjust for things like
>*) clock speed (2.5x - 3x advantage for SandyBridge
>*) Instructions/clock (edge to SandyBridge)

I think I agree with all your factors except this one. Each of the 8 SMXs has 4 schedulers, each capable of issuing up to 2 instructions from a single warp per clock. I may very well be making a mistake but doesn't that mean 64 instructions per clock peak on Kepler vs. 32 on SandyBridge (4 wide issue)? For task level parallelism, SandyBridge supports 16 independent programs running in parallel. I have no idea what the corresponding number is on Kepler.

>*) Cache (edge to SandyBridge)
>*) and Memory Bandwidth (edge to GPUs)

Let me add:
*) Memory Capacity (edge to SandyBridge)

I guess one can claim that GPUs can make use of system memory. But then you are throwing out the memory bandwidth advantage of a GPU. Is there also a memory latency advantage due to use of higher clocked memory?
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
First impressions of Nvidia's KeplerDavid Kanter2012/03/22 06:00 PM
  First impressions of Nvidia's Keplerfellix2012/03/23 01:25 AM
  First impressions of Nvidia's KeplerMike2012/03/23 08:24 AM
    First impressions of Nvidia's KeplerDavid Kanter2012/03/23 09:02 AM
      First impressions of Nvidia's KeplerMike2012/03/23 09:34 AM
        First impressions of Nvidia's KeplerDavid Kanter2012/03/23 12:15 PM
      First impressions of Nvidia's Kepleranon2012/03/23 11:37 AM
        I use ALUsMark Roulo2012/03/23 12:59 PM
          I use ALUsanon2012/03/23 02:07 PM
            I use ALUsMark Roulo2012/03/23 03:12 PM
              I use ALUsanon2012/03/23 04:08 PM
        Makes no sense...EduardoS2012/03/23 05:30 PM
          Makes no sense...anon2012/03/23 06:14 PM
            Makes no sense...David Kanter2012/03/25 10:45 AM
          Makes no sense...fellix2012/03/24 05:41 AM
  Comparing against the 560Cat2012/03/26 08:51 AM
    Comparing against the 560David Kanter2012/03/26 09:24 AM
  Shuffle InstructionMartin2012/03/27 06:17 AM
    Shuffle InstructionDavid Kanter2012/03/27 08:47 AM
      Shuffle InstructionMartin2012/03/27 10:52 AM
        .msi unarchiver?hobold2012/03/28 10:20 AM
          .msi unarchiver?Joe2012/03/28 12:55 PM
          .msi unarchiver?Martin2012/03/29 12:53 AM
      Shuffle InstructionRohit2012/03/27 12:04 PM
  Workgroups vs. warps/wavefrontsAndrew McDonald2012/03/28 02:31 PM
    Workgroups vs. warps/wavefrontsDavid Kanter2012/03/28 03:14 PM
      Workgroups vs. warps/wavefrontsRohit2012/03/28 08:53 PM
        Workgroups vs. warps/wavefrontsLee Howes2012/03/29 06:38 AM
          ThreadsDavid Kanter2012/04/09 11:36 AM
    Fixed (NT)David Kanter2012/04/09 11:37 AM
  Heterogeneous GPUsOscar Eddington2012/03/28 07:41 PM
    Heterogeneous GPUsGary M.2012/04/06 04:35 PM
      Different shader coresDavid Kanter2012/04/09 11:29 AM
        Different shader coresTom2012/04/11 02:36 PM
          Nope...David Kanter2012/04/12 01:10 AM
            Nope...Tom2012/04/13 03:58 PM
              Nope...David Kanter2012/04/14 12:24 PM
                Load balancing between Tesla and graphics boardsTom2012/04/15 06:11 PM
                  Load balancing between Tesla and graphics boardsDavid Kanter2012/04/15 10:11 PM
                    Load balancing between Tesla and graphics boardsAnon2012/04/16 05:05 PM
  Why isn't AMD hardware used for GPU computing?Richard G.2012/04/03 07:39 PM
    Why isn't AMD hardware used for GPU computing?anon2012/04/04 03:11 AM
      Why isn't AMD hardware used for GPU computing?Soupdragon2012/04/04 05:24 AM
        Why isn't AMD hardware used for GPU computing?Groo2012/04/04 08:41 AM
      Why isn't AMD hardware used for GPU computing?Michael S2012/04/04 06:24 AM
      Why isn't AMD hardware used for GPU computing?Alexko2012/04/04 08:43 AM
        Why isn't AMD hardware used for GPU computing?EduardoS2012/04/04 03:37 PM
          Why isn't AMD hardware used for GPU computing?David Kanter2012/04/09 02:51 PM
    Why isn't AMD hardware used for GPU computing?Ricardo B2012/04/04 12:57 PM
    Why isn't AMD hardware used for GPU computing?Tom2012/04/04 05:36 PM
      Why isn't AMD hardware used for GPU computing?Brett2012/04/04 06:55 PM
      Why isn't AMD hardware used for GPU computing?David Kanter2012/04/09 02:55 PM
  Predictions about KeplerRussell Baker2012/04/18 02:09 PM
    Predictions about Kepler01000102012/04/18 03:14 PM
      Predictions about KeplerEduardoS2012/04/18 03:38 PM
        Predictions about KeplerAnon2012/04/18 08:48 PM
          Predictions about KeplerEduardoS2012/04/19 03:03 PM
            Predictions about KeplerMeeps2012/04/19 03:39 PM
      Predictions about KeplerJohn P.2012/04/18 06:13 PM
      Predictions about KeplerFoo_2012/04/19 12:15 PM
        Predictions about KeplerEduardoS2012/04/19 03:07 PM
    Predictions about KeplerGroo2012/04/19 09:13 AM
      Predictions about Kepleranon2012/04/19 03:26 PM
        Predictions about KeplerGroo2012/04/20 08:01 AM
    Predictions about KeplerAlex L.2012/04/20 03:41 PM
      Predictions about KeplerAnon2012/04/21 09:34 AM
        Predictions about Keplermpx2012/04/21 11:23 PM
          Predictions about Keplerac2012/04/22 01:49 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell avocado?