DX12 + i7 5775C + Win10 = ?

Discussion in '3D Hardware, Software & Output Devices' started by gongo, May 3, 2015.

  1. homerdog

    homerdog donator of the year
    Legend Veteran Subscriber

    Joined:
    Jul 25, 2008
    Messages:
    6,292
    Likes Received:
    1,074
    Location:
    still camping with a mauler
    They only offer GT2 graphics on the desktop. GT3 and/or the eDRAM only made it to mobile chips. It was never even an option on the socketed chips. I have no doubt they could have sold a lot of i7s with 64 or 128MB of eDRAM on LGA 1150 even if they were $400. I think they will offer eDRAM for Broadwell and Skylake desktop parts.
     
  2. Andrew Lauritzen

    Andrew Lauritzen Moderator
    Moderator Veteran

    Joined:
    May 21, 2004
    Messages:
    2,553
    Likes Received:
    633
    Location:
    British Columbia, Canada
    As long as there's a nice balance of these vs. "there shouldn't be an iGPU at all" comments, it's in a good place right? ;)

    Well this thread is about us giving you a socketed broadwell GT3e option right? :)
     
  3. homerdog

    homerdog donator of the year
    Legend Veteran Subscriber

    Joined:
    Jul 25, 2008
    Messages:
    6,292
    Likes Received:
    1,074
    Location:
    still camping with a mauler
    Heh, well more options are better IMO. But yes I would rather Intel offer a 4c/8t CPU (on LGA1150) with no IGP at all than one with Iris Pro graphics. Even GT2 graphics take up like half of the die on a 4790K.
     
  4. Andrew Lauritzen

    Andrew Lauritzen Moderator
    Moderator Veteran

    Joined:
    May 21, 2004
    Messages:
    2,553
    Likes Received:
    633
    Location:
    British Columbia, Canada
    It doesn't really, but yeah. It's only the 2+3 configs that get into those sorts of ratios, even if you take a fairly liberal accounting of the shared resources (LLC, etc). In any case die space is really a factor for the economics of producing the chip, not an end-user concern per se. The iGPU isn't harming you even if you never use it :)

    For more cores, I'll remind folks that there's HSW-E and it's sweet (I say while compiling stuff on 16 threads :)). If you just want 6 cores + 2 channels of memory, it's not even that much more expensive than the 4790k, and you get lots of other goodness with it too (lots of cache, more PCI-E lanes, etc).

    I get that there may be reasons people may prefer a 4790k or similar, but it's untrue to say that there are no options.
     
  5. DmitryKo

    Regular

    Joined:
    Feb 26, 2002
    Messages:
    918
    Likes Received:
    1,122
    Location:
    55°38′33″ N, 37°28′37″ E
    Because eDRAM primarily improves power efficiency, as capacitor-based DRAM consumes much less compared to transistor-based SRAM.

    eDRAM in GT3e does improve memory bandwidth 5-fold to 100 Gbyte/s thanks to 512 bit memory bus, but it doesn't translate to noticeable performance improvements as shown by benchmarks, as there are arguably not enough processing blocks to saturate this bandwidth...

    That depends on the type of post processing, but I understand it's generally ALU bound - we should ask Max to clarify the details.

    Anyway, shared memory bandwidth and shader processing performance are quite limited on the HD 4600 (GT2), and top-end Iris Pro 5200/6200 (GT3e) only has twice the processing power. So I wish anyone luck making any use of this, but anything lower than Skylake GT4e with 72 computing units IMHO does not even worth the effort...

    I was just making an obvious point that eDRAM won't make the same integrated GPU design perform significantly faster and on the level of similarly priced discrete GPUs.
    HD 4600 is seemingly ~6-8 times slower than GeForce 750 doing the same post-processing task, and I believe Iris Pro 6200 (GT3e) will still be ~3-4 times slower than GeForce 750, even though they have a very similar memory bandwidth, ~90-100 GBytes/s.

    Oh yeah... give me a socketed Skylake GT4e - I'd disable the graphics right away and just use it for L4 cache :) Or maybe I wouldn't...
     
  6. Andrew Lauritzen

    Andrew Lauritzen Moderator
    Moderator Veteran

    Joined:
    May 21, 2004
    Messages:
    2,553
    Likes Received:
    633
    Location:
    British Columbia, Canada
    It absolutely does make a big difference to performance! What performance benchmarks are you alluding to? I believe Intel has even published results with the same SKU with eDRAM off vs. on and it's quite a big difference. In terms of just raw performance from consumer SKUs, there's a large difference between 28W GT3 and 47W GT3e. Obviously an imperfect comparison due to different TDPs and CPU configuration, but I'm not sure where you're getting the notion that eDRAM doesn't help performance from...

    Certainly it may not need all ~100GB/s (50 + 50) in current configurations since the LLC helps, but it definitely needs a chunk of it :)

    Pricing is a finicky issue, but that aside obviously the memory interface is designed to feed the width of the GPU.

    It's ~6.5x by my math whereas it should be ~5x on paper. So indeed there's clearly performance being dropped in this test, but it's early and there's no fundamental reason why it needs to be so far from optimal performance on the iGPU (although there's always going to be some overhead of spinning up/down a GPU vs. running it in a steady state... and that's ignoring potential power/frequency effects of letting the GPU idle which would skew the scaling math). Post-processing is almost always bandwidth/tex bound and since this looks to be ~2x off the perf that it should be, it may actually be quite bandwidth bound here.

    Broadwell increases tex disproportionately to ALU, so cases like post processiong will benefit just from the architecture shift. On paper, Broadwell GT3 is roughly half a 750 at iso clocks (24 vs 40 tex, 384 vs 640 ALU) and there's not really much magic that happens with post processing to skew any architecture too far from optimal.

    That said I would be completely unsurprising if a chunk of perf is currently lost in the software/drivers/runtime at this point. Lots of optimization work yet to come for these new use cases.
     
    #46 Andrew Lauritzen, May 7, 2015
    Last edited: May 7, 2015
Loading...

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...