HBM2 AMD iGPU's in Intel CPU's

Discussion in 'Gaming (Software and Graphics Cards)' started by Templesa, Nov 6, 2017.

  1. hfm

    hfm Notebook Virtuoso

    Reputations:
    1,637
    Messages:
    3,291
    Likes Received:
    1,039
    Trophy Points:
    231
    It can be due to many different aspects of the node and drivers, but the fact was that it was using a sizable amount more power for the same relative performance in real world use cases.

    That doesn't sound like a win unless they are bringing process improvements and optimizations to the architecture along with this implementation with Intel to improve their performance per watt.
     
  2. laserbullet

    laserbullet Notebook Evangelist

    Reputations:
    84
    Messages:
    593
    Likes Received:
    71
    Trophy Points:
    41
    Ha! I just realized that if this does threaten the 1050, that actually does put some pressure on Nvidia to stop dragging their damned feet and get Volta out the door. Who would have thought Nvidia would be receiving pressure from an iGPU at the start of the year. :D
     
    hmscott likes this.
  3. hmscott

    hmscott Notebook Nobel Laureate

    Reputations:
    3,631
    Messages:
    13,539
    Likes Received:
    15,677
    Trophy Points:
    931
    Likely to put a hurt on Xmas season for Nvidia with both AMD and now Intel with iGPU killers :D
     
  4. Deks

    Deks Notebook Virtuoso

    Reputations:
    872
    Messages:
    3,599
    Likes Received:
    746
    Trophy Points:
    181
    Drivers have little to do with Vega's power draw.
    Yes, they can modify power profiles for example at expense of performance, however, the manuf, process is the main culprit here, because it was designed for low power mobile parts, not high performing desktop ones.
    The fact that AMD was able to get the performance they did at the clocks they are pushing on Vega is pretty good. If you add undervolting to it, you get a really good performance for same or better perf. per watt like Pascal on a manuf. process that's not a good choice for high performance desktop parts.
    Undervolting is crucial here as it's not 'cheating' in any real sense of the word. Nvidia effectively put a stop to undervolting since they tied the clocks and voltages together... so you can't drop the voltage without dropping clocks, which negatively impacts performance.

    Look at AMD's touted number for Raven Ridge and Vega iGP for example. Those parts are clocked on low levels and still, they seemingly surpass Intel's mobile solutions in the same TDP in multicore (Even though Intel maintains a 10% advantage in single core due to 5% better IPC and 5% higher boost on single core, although I'm skeptical if Intel even has an IPC advantage considering most industry software is optimized for Intel to begin with and Cinebench still uses Intel compilers - but, regardless, that 10% in single core is pretty negligible) and the Vega igp seems to be equivalent to GT 1030 (MX 150).
    Vega seems to be AMD's basis for a scalable GPU architecture. Raven Ridge is where 14nm Glofo process actually shines through (unless of course AMD used 12nm LP process, but I doubt it as this APU is supposed to be released this year, and 12nm is slated for early next year).

    Either way, I would suggest to wait and see what 12nm refresh brings. Considering it bears the same LP designation as the upcoming 7nm process from IBM that Glofo will be using for Ryzen 2 (and specs indicate base clock of 5 GhZ), there's a chance this process is similarly designed for higher performing parts - in which case, AMD could easily bring a much higher performing Ryzen and Vega refreshes to the table with improved efficiency.
     
    Last edited: Nov 8, 2017
  5. hfm

    hfm Notebook Virtuoso

    Reputations:
    1,637
    Messages:
    3,291
    Likes Received:
    1,039
    Trophy Points:
    231
    Vega 64 is 14nm where 1080 is 16nm and it's still wiping the floor with the Vega in performance per watt. I don't see how AMD competes here
     
  6. Deks

    Deks Notebook Virtuoso

    Reputations:
    872
    Messages:
    3,599
    Likes Received:
    746
    Trophy Points:
    181
    Again, you are missing the point. There are clear differences between the types of manuf. processes used... 14nm Glofo is primarily made for mobile parts and low frequencies. 16nm TSMC is made with high performance parts in mind (and there's only 2 nm difference between the two processes).
    That's why performance per watt is better out the factory for Nvidia... but as I said before, by undervolting AMD gpu's in turn, you improve performance per watt close to or at Pascal levels.
    Pascal is nothing more than overclocked Maxwell... TSMC 16nm allows it to be clocked much higher than Maxwell at similar or better power draw... hence the performance advantage.
    AMD is forced to use lower clocks because the manuf. process will otherwise incur even higher power draw.... however, it does manage to deliver similar/same/better performance on lower clocks... but in order to achieve somewhat of a parity with Pascal, Vega needs to be unvervolted - that's why AMD provided Wattman for Polaris and Vega.
     
  7. yrekabakery

    yrekabakery Notebook Guru

    Reputations:
    22
    Messages:
    68
    Likes Received:
    64
    Trophy Points:
    26
    What I don't get is, like, Vega 64 is clocked some 60% higher than Fury X, but somehow it's only 30% faster? How does that make sense. Performance usually scales linearly with clock speed, at least that's what we see with Pascal being a higher clocked Maxwell, since at the higher clock speed every part of the GPU runs faster. So this makes me wonder if there's some other internal bottleneck going on with Vega. Could it be ROP-limited like Fiji was rumored to be? Vega still has 64, no increase there from Fiji (or Hawaii for that matter). Maybe memory bandwidth? It's actually lower on Vega than on Fiji.

    As a result, Vega's perf/FLOP is actually lower than Fiji. If Vega 64's performance actually scaled up linearly with its clocks/FLOPS increase, it would be nipping at the heels of the 1080 Ti, not be a 1080 competitor.
     
    Last edited: Nov 8, 2017
  8. Deks

    Deks Notebook Virtuoso

    Reputations:
    872
    Messages:
    3,599
    Likes Received:
    746
    Trophy Points:
    181
    Vega 64 IS a 1080ti and Titan X competitor... in professional software mostly.
    You have to bear in mind that most games were not written with AMD hardware in mind... also, Vega DOES have various features which games simply don't use (apart from Wolfenstein II it would seem where Nvidia has much lower advantage, and even there, more optimizations are needed by the devs as the game is not well ported it would seem).

    It could also be another issue with what Koduri mentioned (namely that the Infinity Fabric is not optimized for games).
    As for performance scaling... uhm, never have I seen a GPU (AMD or Nvidia) where clock increases result in linear increase in performance. With Vega, independent reviews confirm that HBM2 overclocking produces higher performance increase for a very low increase in power consumption (about 5w to 10W) vs clock increases (where a minor bump increases power draw by a much larger amount).

    Vega works better when undervolted on both core and HBM and then overclocked on HBM to ~940 MhZ (Vega 56) and ~1100 MhZ (Vega 64).
     
  9. yrekabakery

    yrekabakery Notebook Guru

    Reputations:
    22
    Messages:
    68
    Likes Received:
    64
    Trophy Points:
    26
    It's there with Pascal. Look at 1050 vs. 750 Ti, 1050 Ti vs. 950, 1060 vs. 970M, 1070N vs. 980, etc.
     
  10. Deks

    Deks Notebook Virtuoso

    Reputations:
    872
    Messages:
    3,599
    Likes Received:
    746
    Trophy Points:
    181
    I would appreciate some sources please
     
Loading...

Share This Page