DX12 + i7 5775C + Win10 = ?

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

  1. gongo

    Regular

    Joined:
    Jan 26, 2008
    Messages:
    604
    Likes Received:
    25
    How about physics? Where did the promised physics gameplay.... gone to?

    Here is a new leak on Broadwell/Skylake.... http://diy.pconline.com.cn/637/6375701.html

    Intel now ranks socketed Broadwell 5775C even above Skylake 6700K(!)...it was different in the previous leak..5775C was ranked below 4790K(!)....have they uncovered some interesting results between now and then..?

    And, 5775C comes with 128mb L4 eDram now! ?
    The UE4 multiadaptor demo ran with a 4790K iGPU....can it compare to 5775C with >twice the EUs, improved with DX12...and 128mb of super-fast eDram(?)...

    I feel we are in for a Broadwell/Iris Pro surprise!
     
  2. Davros

    Legend

    Joined:
    Jun 7, 2004
    Messages:
    16,866
    Likes Received:
    4,190
    Explain plz...
     
  3. DmitryKo

    Regular

    Joined:
    Feb 26, 2002
    Messages:
    918
    Likes Received:
    1,122
    Location:
    55°38′33″ N, 37°28′37″ E
    It's not a rank, it's a time scale. 5775C was introduced on top of 4970K, which is being replaced by 6700K, and 5775C continues further on.
    Arguably they could as well have placed the "+" tiers below the "K" tiers, but then you would probably come with a different grand сonspiracy theory :razz:

    As for Iris Pro 6200, eDRAM won't make weak processing cores perform much faster - it's just a way to improve CPU performance and power consumption by unloading the memory bus and SRAM caches from graphics accesses.
     
    #23 DmitryKo, May 6, 2015
    Last edited: May 6, 2015
  4. DmitryKo

    Regular

    Joined:
    Feb 26, 2002
    Messages:
    918
    Likes Received:
    1,122
    Location:
    55°38′33″ N, 37°28′37″ E
     
  5. Blazkowicz

    Legend Veteran

    Joined:
    Dec 24, 2004
    Messages:
    5,607
    Likes Received:
    256
    You're right - well, partly right because if you build a gaming PC around a Bay Trail Atom or AM1 or an Athlon X2 340 (which has 16x lanes at least) it will suck balls unless you plan to play old games at high res.

    But my deranged, ranty reasoning is that if you've fallen behind enough (or use a laptop, old or cheap and new) the GPU in the 5675C will be a major upgrade anyway in performance and especially features/compatibility (same deal as with a quad core AMD APU, some people buy them as an upgrade). Now you could get an AMD but it feels uneasy to get all new hardware (mobo, CPU, RAM) to get CPU performance around Phenom II or overclocked Q6600.
    So, people have to buy a vanilla i5 (i.e. 4460, 4590) to hedge against the CPU requirements of the most CPU demanding games (current, future and even past games) or sometimes an i3. (or FX6350, Athlon 860K at worst)

    But it suck to then having to blow yet more cash on GPU, PSU etc. . i5 5675C is rather cheaper than I expected, and for about that price it would be a better buy than i5 + a geforce GT730 or R7 240/250. yes, i3 + gtx 750 is technically better but anyway you can keep that i5 for at least 5 years (CPU are stagnating) and add a GPU at any other time (GPU don't stagnate, they'll eventually go 14nm + HBM on the mid range).
     
  6. Infinisearch

    Veteran Regular

    Joined:
    Jul 22, 2004
    Messages:
    779
    Likes Received:
    146
    Location:
    USA
    I think he was alluding to compute accelerated physics on the igpu and graphics on dgpu with no cpu intervention enabled by D3D12 multi-adapter. Maybe.

    In the case of post processing as in the above demo isn't memory bandwidth rather important for performance? And if you're ping ponging between buffers having them stay in the L4 cache may help out quite a bit. In addition didn't they say they didn't offload all post processing to the igpu potentially implying that they only offloaded what they thought the igpu could handle. Perhaps the metric was ALU power.
     
  7. Picao84

    Veteran Regular

    Joined:
    Feb 15, 2010
    Messages:
    2,000
    Likes Received:
    1,051
    That does not make any sense. The objective of Optimus and Enduro was not to give better performance, but to get better battery life by switching the discrete GPU completely off when not needed. DX12 might eventually make that easier to do, but that is not what is being discussed on that article.
     
  8. Andrew Lauritzen

    Andrew Lauritzen Moderator
    Moderator Veteran

    Joined:
    May 21, 2004
    Messages:
    2,553
    Likes Received:
    633
    Location:
    British Columbia, Canada
    Uhh... what are you assuming is the hardware in that demo exactly?
     
  9. ToTTenTranz

    Legend Veteran

    Joined:
    Jul 7, 2008
    Messages:
    12,236
    Likes Received:
    7,192
    I assumed it was at least a GT3 either from Haswell or from Broadwell, otherwise the performance scaling would be >100%.
     
  10. Andrew Lauritzen

    Andrew Lauritzen Moderator
    Moderator Veteran

    Joined:
    May 21, 2004
    Messages:
    2,553
    Likes Received:
    633
    Location:
    British Columbia, Canada
    What do you mean by "performance scaling would be >100%"? But yeah I would not make assumptions on the hardware in the demo unless/until Microsoft says what it was.
     
  11. ToTTenTranz

    Legend Veteran

    Joined:
    Jul 7, 2008
    Messages:
    12,236
    Likes Received:
    7,192
    Whatever is the iGPU being used, it's causing a 10% performance increase over a Titan X capable of >6TFLOPs.
    A GT3 from Haswell is capable of up to 830GFLOPs, meaning it's giving less than 100% of its "potential" to the system performance (830*10 = 8300 > 6000 TFLOPs) - in compute performance, of course.
    If it was using a GT2 GPU, then the compute performance effectively halves, and for a 400GFLOPs GT2 GPU to provide a 10% performance bump on a 6TFLOPs discrete GPU, the scaling would go over the total theoretical "compute contribution". Hence the ">100%".
     
  12. Andrew Lauritzen

    Andrew Lauritzen Moderator
    Moderator Veteran

    Joined:
    May 21, 2004
    Messages:
    2,553
    Likes Received:
    633
    Location:
    British Columbia, Canada
    You're still making a lot of assumptions about the hardware. Some of them are actually impossible right now... how would you pair a GT3 with a Titan X? Also do those performance numbers for Elemental even make sense for a Titan X?
     
  13. ToTTenTranz

    Legend Veteran

    Joined:
    Jul 7, 2008
    Messages:
    12,236
    Likes Received:
    7,192
    Did you read the first post in this thread?
    Or any of the posts from the first page even?
     
  14. Andrew Lauritzen

    Andrew Lauritzen Moderator
    Moderator Veteran

    Joined:
    May 21, 2004
    Messages:
    2,553
    Likes Received:
    633
    Location:
    British Columbia, Canada
    Did you see a source link from the wfcctech article or any indication that they aren't just making up what hardware was used for the multiadaptor demo? The primary source is Max's talk and the DX blog. Anything beyond that is pure speculation.
     
    BRiT, Alexko and iroboto like this.
  15. Max McMullen

    Newcomer

    Joined:
    Apr 4, 2014
    Messages:
    20
    Likes Received:
    105
    Location:
    Seattle, WA
    The hardware used in the demo was an Intel 4600 GPU with an NVidia GTX 750. The average utilization of the Intel GPU was 70% and the NVidia GPU was 99%. The demo used a fixed partitioning of the post-processing pipeline, there's room for improvement with dynamic partitioning. Dynamic partitioning also seems necessary to deal with the variety of configurations across all consumer machines, including thermal limits that can vary more from machine to machine than the actual GPU hardware.

    While maximizing performance across multiple processors isn't an easy problem to solve there are simple workloads where iGPU+dGPU makes sense already: guess how the OS implements hybrid graphics support? By exposing the feature at the API I'm expecting a lot of new ideas to come from the developer community on how to use the extra horsepower.
     
    Osamar, BRiT, Alexko and 2 others like this.
  16. 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
    Is it problematic if the iGPU and dGPU don't share the same feature level support?

    Also if the iGPU is running full tilt, couldn't this impact CPU performance since the iGPU and CPU share the same memory bus? Especially if the iGPU starts doing blending operations etc. that require lots of bandwidth.
     
  17. Max McMullen

    Newcomer

    Joined:
    Apr 4, 2014
    Messages:
    20
    Likes Received:
    105
    Location:
    Seattle, WA
    It's up to the application to deal with feature capability differences. The feature support for each GPU can be enumerated by the Direct3D 12 API. These days so many GPUs are FL 11+ that dealing with capability differences is much easier than it used to be.

    Your point regarding iGPU performance isn't unique to multiadapter, it's possible for heavy iGPU utilization to decrease CPU performance and visa-versa. The asteroids demo Andrew Lauritzen wrote to show off DX12 last SigGraph demonstrates reduced CPU load allows the GPU to clock up. IIRC this was a thermal limitation.

    The general statement that's true is with both an iGPU and a dGPU, there are more shared resources in the system that can be bottlenecks. Managing those bottlenecks is one of the hard engineering problems necessary to solve for maximizing performance. In this case system memory bandwidth, CPU performance, and thermal limits are some of the bottlenecks with increased pressure in an iGPU+dGPU setup.
     
  18. Andrew Lauritzen

    Andrew Lauritzen Moderator
    Moderator Veteran

    Joined:
    May 21, 2004
    Messages:
    2,553
    Likes Received:
    633
    Location:
    British Columbia, Canada
    Thanks for clarifying the hardware Max :)

    I agree that there are a lot of open questions about in which cases this will be a win and by how much, but exposing it to applications is the best way to start exploring the space of possibilities.

    There are too many different combinations of hardware to try and draw any sort of general conclusions from desktops with high end dGPUs that are ~10-20x faster than the iGPU to some laptops where the dGPU is roughly the same speed as the iGPU. (There are even cases where the dGPU is actually slower than the iGPU but as those are fairly stupid configs I'll ignore them for now).

    Certainly the cases with beefier iGPUs (Iris Pro, etc) are likely to be the most compelling. Otherwise you're often talking about roughly as many FLOPs as you could already get with AVX2 if developers bothered to make use of that :)

    But I'd argue that even with lower end iGPU configurations (i.e. the baseline that you get in most desktop CPUs these days), there are still some possibilities due to a few factors:
    1) iGPUs still have access to fixed function hardware for graphics such as the rasterizer and texture sampler.
    2) iGPUs can often run closer to the CPU timeline with lower latency on getting results. This is especially true in cases like the one demonstrated in Max's talk where you're not submitting a deep pipeline of iGPU work at all and rather letting it idle for part of the time. This sort of setup has some similar benefits to low-latency async compute, except you can do async graphics/3D too.
    3) iGPUs usually don't need to copy memory around when interacting with the CPU.

    For these reasons I think there's a number of interesting things to pursue in terms of using the iGPU to do low-latency graphics tasks in support of CPU work (either to set up the main rendering, or not related to graphics at all). I'm excited to see what developers come up with now that there's a first-class API to do it!
     
    BRiT likes this.
  19. 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
    Will be interesting to see what comes of this. It would seem in most gaming PCs the dGPU is vastly superior to the iGPU (if there even is an iGPU), so unless the iGPU has some inherent advantage over the dGPU for certain tasks (perhaps due to proximity to the CPU - shared address space, cache etc.) it seems hard to justify for all the complexity it involves. This would be way cooler if Intel hadn't strangely castrated all the desktop iGPUs.

    Anyhow maybe one day I'll enable the GPU of my 3770K in the BIOS :)
     
  20. Davros

    Legend

    Joined:
    Jun 7, 2004
    Messages:
    16,866
    Likes Received:
    4,190
    ???
     
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...