Qualcomm SoC & ARMv8 custom core discussions

Discussion in 'Mobile Devices and SoCs' started by Nebuchadnezzar, Jan 20, 2015.

  1. Nebuchadnezzar

    Legend

    Joined:
    Feb 10, 2002
    Messages:
    974
    Likes Received:
    141
    Location:
    Luxembourg
  2. Wishmaster

    Newcomer

    Joined:
    Nov 16, 2008
    Messages:
    238
    Likes Received:
    0
    Location:
    Warsaw, Poland
    Taipan? Sounds funny when compared to Krait :)
    Wonder what they've come up with.
    Krait is a well done CPU so it'll be interesting to see what Taipan is all about.
     
  3. juicytuna

    Newcomer

    Joined:
    Jul 27, 2005
    Messages:
    71
    Likes Received:
    0
    Taipan, like Krait, is a type of snake so it fits in with their reptilian theme. I'm a bit puzzled by the presence of big.LITTLE. Wasn't it a fairly unique idea when ARM came up with the scheme for A15/A7? Now we have Qualcomm also agreeing this is the right way forward for mobile SOCs.
     
  4. Wishmaster

    Newcomer

    Joined:
    Nov 16, 2008
    Messages:
    238
    Likes Received:
    0
    Location:
    Warsaw, Poland
    One interesting bit - 815 is supposed to be an octa core with big.LITTLE but 820 is only listed as octa core.
    Besides, it looks like Taipan S2 is going to be 64bit but before that there will be Taipan S1 which isn't listed as 64bit. Something like that would be strange so I think that both Taipan S1 and S2 will be 64bit.
     
  5. McHuj

    Veteran Regular Subscriber

    Joined:
    Jul 1, 2005
    Messages:
    1,432
    Likes Received:
    553
    Location:
    Texas
    Don't believe everything you read on the internet.
     
  6. Ailuros

    Ailuros Epsilon plus three
    Legend Subscriber

    Joined:
    Feb 7, 2002
    Messages:
    9,420
    Likes Received:
    179
    Location:
    Chania
  7. Exophase

    Veteran

    Joined:
    Mar 25, 2010
    Messages:
    2,406
    Likes Received:
    429
    Location:
    Cleveland, OH
    Can anyone give even one reason why this should be taken credibly?
     
  8. Nebuchadnezzar

    Legend

    Joined:
    Feb 10, 2002
    Messages:
    974
    Likes Received:
    141
    Location:
    Luxembourg
    I wonder if that's the fixed revision or not.

    The roadmap is bogus an outdated, just opened a new thread to separate the discussion.
     
  9. Ailuros

    Ailuros Epsilon plus three
    Legend Subscriber

    Joined:
    Feb 7, 2002
    Messages:
    9,420
    Likes Received:
    179
    Location:
    Chania
    They usually start out relatively low and then increase performance as time goes by. In the absence of battery life scores my question would be if it throttles at all at that performance level.
     
  10. Erinyes

    Regular

    Joined:
    Mar 25, 2010
    Messages:
    647
    Likes Received:
    92
    I find it very interesting that as per this, Qualcomm is developing not one, but two custom cores to do big.LITTLE entirely with their own cores. The 815 seems to have something like a 4 A57 + 4 A53 type configuration and the 820 seems to have 8 full A57 type cores. However, as others have stated, this is an old and/or bogus roadmap so we should take it with a large grain of salt.
    Why do you say *yawn*? The Adreno 430 scores seem to be at QHD resolution compared to the 1080p of the A6 and as you say if we look at the offscreen 1080p results, it is quite competitive.

    I am also interested in seeing battery life results, especially given what we've seen with S805 and Adreno 420.
     
  11. Ailuros

    Ailuros Epsilon plus three
    Legend Subscriber

    Joined:
    Feb 7, 2002
    Messages:
    9,420
    Likes Received:
    179
    Location:
    Chania
    Manhattan 1080p offscreen:
    Adreno430 = 21.9 fps
    Adreno420 = 19.4 fps
    Apple A8 = 19.3 fps

    If the 430 scores don't increase then it's a 13% difference to the current 420 GPU, which deserves every "yawn" there is and I wouldn't suggest that Apple's next smartphone SoC will be in the same test by another lukewarm 13% faster than in the A8.
     
  12. Erinyes

    Regular

    Joined:
    Mar 25, 2010
    Messages:
    647
    Likes Received:
    92
    I see a value of 18.3 fps for Adreno 420. That puts The 430 about 20% ahead. And comparing T-Rex Offscreen results, it is 48.5 fps vs 40.7 fps, so again 20% ahead. Given that Qualcomm themselves claimed that the 430 would be only 30% faster than the 420, the fact that we're seeing 20% in real world benchmarks seem to about right (there may be some throttling in play as well so the gap could increase). This is nothing to sneeze at IMHO, especially considering that the 420 is itself about 40% faster than the 330. So that makes the 430 about 70-80% faster than the 330..which is a better comparison.

    Aside from the fact that it is competitive with the A8..why does it matter anyway..Android and iOS are two completely different markets so its like comparing Oranges to err Apples :lol:
     
  13. Ailuros

    Ailuros Epsilon plus three
    Legend Subscriber

    Joined:
    Feb 7, 2002
    Messages:
    9,420
    Likes Received:
    179
    Location:
    Chania
    I'm looking at "best scores".

    If 430 is by 30% faster (without an inch of throttling) than the cold boot peak result of a 420, then I'll truly tip my hat off to them. Given though how much 420 throttles if the former would be true their difference would be up to 80 and not 30%.

    It's by 40% faster than the 330 with or without throttling?

    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Samsung SM-G906 Galaxy S5 Prime +47%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Samsung Galaxy Note Edge (Adreno 420, SM-N915x, SCL24, SC-01G) -55%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Google Nexus 6 -21%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Motorola Droid Turbo XT1254, Moto Maxx XT1225 -37%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Samsung Galaxy Note 4 (Adreno 420, SM-N910x) -30%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=LG G3 (Adreno 420, F460) -33%

    Adreno 330:
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=HTC One M8 (2.4 GHz, 0P6B) -1%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=OnePlus One (A0001) -0.7%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Sony Xperia Z3 Compact Tablet (SGP6xx) +1.7%

    Shall I go on or are the trends clear enough?

    For the long term performance test Gfxbench3.0 loops through T-Rex 30x times (onscreen) in a row and records the last result. On these devices on estimate it takes roughly over an hour to complete the 30 runs. Now an Adreno330 gets an 1080p onscreen result at roughly 27-30 fps depending on device and sustains that score even after a reasonable amount of 3D runs. Au contraire the 420 scores from one side a similar score but at a much higher 1560p resolution, but after an hour or so throttles back down up to over 50% depending on device. I wouldn't call that exactly 30% faster; unless of course you're testing performance in it's own vacuum and power consumption in a totally different dimension in order for those two factors never to interact :roll:

    http://gfxbench.com/device.jsp?benchmark=gfx30&os=iOS&api=gl&D=Apple iPhone 6 Plus
    -3%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=iOS&api=gl&D=Apple iPhone 6 -0.8%
    http://gfxbench.com/device.jsp?benchmark=gfx30&os=iOS&api=gl&D=Apple iPad Air 2 -0.0%

    Given that you might compare those 19 frames the A8 gets in the iPhone 6 Plus to the Adreno420 and consider it competitive, but it's a damn shame the long term performance test doesn't also run with Manhattan; would you still suggest it would be able to sustain its original "cold" score. Given the data at hand I severely doubt it.
     
  14. Nebuchadnezzar

    Legend

    Joined:
    Feb 10, 2002
    Messages:
    974
    Likes Received:
    141
    Location:
    Luxembourg
    Trex is more power hungry than Manhattan in my testing.
     
  15. Ailuros

    Ailuros Epsilon plus three
    Legend Subscriber

    Joined:
    Feb 7, 2002
    Messages:
    9,420
    Likes Received:
    179
    Location:
    Chania
    Probably, because it uses more geometry?
     
  16. Nebuchadnezzar

    Legend

    Joined:
    Feb 10, 2002
    Messages:
    974
    Likes Received:
    141
    Location:
    Luxembourg
    Not sure on Adreno but on Midgard the load/store unit is stressed about twice as much while having only 70% of the ALU load compared to Manhattan. I dwell more on this in the Duke Exynos Forever article.
     
  17. Lazy8s

    Veteran

    Joined:
    Oct 3, 2002
    Messages:
    3,100
    Likes Received:
    18
    Outside of arithmetic, Manhattan is a lot more memory intensive. A few consecutive runs crashes the app on iPhone 6+ whereas the countlessly looped T-Rex endurance test has no issue. 1 GB nonsense...

    Still interesting how, among the 1080p display devices, the iPhone seems to be the only one to perform even higher with the lesser overhead of the onscreen Manhattan test than the off.
     
  18. Ailuros

    Ailuros Epsilon plus three
    Legend Subscriber

    Joined:
    Feb 7, 2002
    Messages:
    9,420
    Likes Received:
    179
    Location:
    Chania
    Here the first LG/S810 results: http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=LG H959

    Sw issues? Onscreen is actually at 1080p, meaning that offscreen results shouldn't have as big difference in the overhead test, or am I missing something?

    Nebu,

    LOL@DEF albeit it's actually reserved for DNF :p

    That's weird; between the first benchmark run and you clicking back a couple of buttoms to re-run the test, it doesn't have time enough to flush the memory? I'll give it a try on a 2GB device and see if the problem is probably elsewhere.

    The LG results I just posted above are similar in that regard.

    ***edit: ROFL it crashed on the 10th Manhattan run and yes instead of a last result it states OUT OF MEMORY. All other runs were on a stable 5.3 onscreen (Onda V989). If you can make 5 runs on the 6 Plus it's an achievement.
     
    #18 Ailuros, Jan 29, 2015
    Last edited: Jan 29, 2015
  19. wco81

    Legend

    Joined:
    Mar 20, 2004
    Messages:
    6,204
    Likes Received:
    264
    Location:
    West Coast
    Qualcomm reported today and guided lower earnings for the year because a major customer isn't going to use the 810.
     
  20. Nebuchadnezzar

    Legend

    Joined:
    Feb 10, 2002
    Messages:
    974
    Likes Received:
    141
    Location:
    Luxembourg
    Right should have used Duke Note Forver.

    This is first time I hear of memory issues with Manhattan, that sounds like a blatant memory leak.
    Samsung SLSI must be popping the champaign by now.
     
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...