AMD's Ryzen CPU's (Ryzen/TR/Epyc) & Vega/Polaris GPU's

Discussion in 'Hardware Components and Aftermarket Upgrades' started by Rage Set, Dec 14, 2016.

  1. hmscott

    hmscott Notebook Nobel Laureate

    Reputations:
    4,569
    Messages:
    15,958
    Likes Received:
    19,648
    Trophy Points:
    931
    There are other devices running through the 7nm process besides AMD GPU's, so Nvidia isn't necessarily waiting on AMD to pioneer the process at either foundry alone, just in general waiting for the first turns to come through and see how the process is yielding.

    Nvidia is likely pushing through test runs all the time themselves as well, but the foundry is getting feedback from more than AMD and Nvidia, so it all adds up to progress for everyone.

    Yup, we don't know until we know, and Nvidia and AMD are going to know way before we do. :)
     
    Vasudev likes this.
  2. Deks

    Deks Notebook Prophet

    Reputations:
    983
    Messages:
    4,038
    Likes Received:
    1,243
    Trophy Points:
    231
    35-40% increase in performance is just from the process itself... and that's over 16nm high performing process Nvidia used.
    14nmLPP as you know limited AMD's clocks quite a bit and produced high power consumption on lower frequencies, not to mention the fact that it resulted in low yields which forced AMD to increase the voltages (though, higher power consumption on AMD's end was also a result of 40% higher CU's).
    So, we are likely to see more than 35% increase in performance on 7nm (as a single chip).
    Think of 7nm Vega as if Vega was coming from 16nm TSMC process instead (which would probably mean same or very similar core clocks like we got on Pascal, unknown higher HBM frequencies, and probably lower voltages from the factory).

    Remember that Vega 56 is clocked about 32% less than GTX 1070 on the core alone where it was already comparable to GTX 1070, and bumping its HBM alone from 800MhZ to 900-940MhZ resulted in a pretty good performance increase... near 1080, for maybe 5W increase in power draw.
    Now, I don't know how much of a power draw the 40% higher CU's contributed to, but its possible that say Vega 56 on 7nm could exhibit around 67-72% core clock increase... or the performance bump on 7nm would result with say 30% increase going to the core and another 37% frequency increase going to the HBM - splitting between the core and HBM.

    Or at least, something to that effect.

    Overall, if Vega 56 on 16nm could achieve similar or slightly lower core clocks (and higher stock HBM frequencies) like Pascal did with much lower power draw (30% less on average, because that seems to be the main difference in overall power consumption between Vega and Pascal), then it would mean that Vega 56 would effectively equal 1080, and Vega 64 would essentially equal 1080ti (at virtually same or slightly higher power draw due to higher CU number).

    Now you need to translate those kinds of 'gain estimates' to 7nm.
    Besides, look at Ryzen... 14nmLPP severely limited its all core boost clocks which only resulted in roughly 200 MhZ increase from baseline... whereas Intel was able to push their all core frrequencies boost higher as their process was made on a high performing process.

    Similar changes would apply to 'hypothetical Vega' on 16nm.
     
    Vasudev and hmscott like this.
  3. Deks

    Deks Notebook Prophet

    Reputations:
    983
    Messages:
    4,038
    Likes Received:
    1,243
    Trophy Points:
    231
  4. jaybee83

    jaybee83 Biotech-Doc

    Reputations:
    3,027
    Messages:
    10,404
    Likes Received:
    7,264
    Trophy Points:
    931
    well this makes sense considering the leapfrogging strategy by AMD. highend vega was the last to come out, so now its midrange polaris turn to be succeeded by a newgen set of chips. guess well have to wait a bit longer for nextgen highend gpus by AMD....
    who knows, maybe the jump to 7nm will make Nvidia turn their 1080 class into the Ti class and their Ti class in turn into Titan class just like they did when introducing Titan/Ti several years ago.

    thatll leave us with their 1070 successor as the 1180/2080 with matching price hike, of course... yay....

    Sent from my Xiaomi Mi Max 2 (Oxygen) using Tapatalk
     
    Vasudev and hmscott like this.
  5. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,150
    Messages:
    4,706
    Likes Received:
    6,576
    Trophy Points:
    581
    I'll respond fully after I finish doing two long benchmarks to show software errors for monitoring software for Tanware their morning. Part of the response will be the response curves that show the trade off between improved performance or energy efficiency.

    Sent from my SM-G900P using Tapatalk
     
  6. Deks

    Deks Notebook Prophet

    Reputations:
    983
    Messages:
    4,038
    Likes Received:
    1,243
    Trophy Points:
    231
    Hmm... I have to wonder slightly about that.
    The head of Radeon group says they are going after high performance, and on the other hand, relatively unsubstantiated rumors seem to tell that Navi will be replacing Polaris as a mid-range GPU.
    Which one do you think is the more credible source?
    Head of Radeon division or rumors?

    Though, I have to say it does make a certain amount of sense for Polaris to be replaced.

    But hang on... if Navi will be using Infinity Fabric to interconnect two GPU chips for example... wouldn't that give AMD the ability to compete in the high end space as well?
    Looks to me like a single Navi chip could easily equate 1070/1080/1080ti level of performance (depending on the TDP brackets and core/hbm frequencies AMD gives each Navi chip)... with higher end being reserved for multi-chips connected through IF - and Navi might have lower number of CU's - not drastically lower, but just enough to help free up power for gaming capabilities).
    Navi could simply be shrunk down V56 for example with sufficiently increased frequencies on the HBM and core to equate V64 and possibly 1080ti.
    That... or possibly even shrunk down and overclocked Polaris with Infinity Fabric?
    Mid-range Navi might be using GDDR6, while high end would be using HBM.

    Remember that Vega supposedly underwent some internal optimizations, so if that gives it lets say 10% bump up in performance (which is possible), that would be on top of the 35-40% performance increase as allowed by 7nm, and then again EXTRA bump up to reach specified TDP (because of the 14nmLPP limitations).

    Will be interesting to see how much of this is wrong, and just how much correct.
    I'm not trying to raise anyone's hopes, I'm merely extrapolating (hopefully within reason) on what we know.
     
    hmscott likes this.
  7. jaybee83

    jaybee83 Biotech-Doc

    Reputations:
    3,027
    Messages:
    10,404
    Likes Received:
    7,264
    Trophy Points:
    931
    i guess its all in the definition of "high performance". if the goal is simply to beat 1080 and maybe 1080 Ti, then right now thats totally "high performance". once 1180/2080 is out, thatll only be midrange ^^ so depending on the point of view, both statements might be true in the end.
     
  8. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,150
    Messages:
    4,706
    Likes Received:
    6,576
    Trophy Points:
    581
    [​IMG]
    What do you see when you see this image? You can either reduce energy by 55%, or you can gain 40% performance at the same power draw. It is an and/or situation, trying to find an intermediate in the trade off. If you have a 35% performance increase and reduce power consumption by 50%, notice how that would intersect outside of the new curve for 7nm. THAT is what I was discussing. Please revise your comments, then restate them after you have done this, and please go back and read my comments on where performance increases came from in light of this new understanding.

    Source: https://fuse.wikichip.org/news/641/iedm-2017-globalfoundries-7nm-process-cobalt-euv/
     
  9. Deks

    Deks Notebook Prophet

    Reputations:
    983
    Messages:
    4,038
    Likes Received:
    1,243
    Trophy Points:
    231
    I understand what you're saying, but that doesn't make too much sense for GPU's... for CPU's yes.

    For GPU's, AMD will use 7nm TSMC process, which offers 35% increase in performance over 16nm TSMC process (and realistically, more on both ends when compared to 14nmLPP).

    And yes, I understand that 'usually', you get that kind of performance increase at same power draw... but my point was that 14nmLPP limitations resulted in Vega's lower yields, which ended up with higher stock voltages and the process was not suitable for driving high clocks in the first place.
    Shouldn't we take those factors into account as well, considering that TSMC has far better experience with GPU production?
    Plus, 1 company won't have the pressure of creating both GPU's and CPU's (aka Glofo - who struggled to supply proper yields for AMD gpu's for a long time now).

    http://www.tsmc.com/english/dedicatedFoundry/technology/7nm.htm

    That's why I'm saying that we're likely not getting the full picture in terms of process comparisons, performance enhancements and power draw reductions.

    https://www.overclock3d.net/news/misc_hardware/tsmc_starts_7nm_volume_production/1

    Granted, I could very easily be wrong.

    https://forums.anandtech.com/thread...-intel-on-leading-edge-process-nodes.2461425/

    "TSMC 7nm vs 16FF+ - 30-35% higher performance at same power or 55-58% lower power at same performance."
     
    Last edited: Jun 11, 2018
  10. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,150
    Messages:
    4,706
    Likes Received:
    6,576
    Trophy Points:
    581
    This is GF announcing how it works at an international transistor conference. This IS NOT up for debate. You either get more performance, or you cut energy consumption, or a mix. You can NOT get the max on both.

    Now, as I said, the HBM2 being used is 1.2GHz, which is 30% higher than the 950MHz or 800MHz variants used on the original generation of Vega. Overclocking the HBM2 was shown, in quite a few cases, to be as potent or more potent than your core overclock. So, you can easily say that 10% of the 35% comes from the memory bandwidth increase alone. It could be 15% or so.

    That means you have 20-25% of performance to account for. If they took 50% reduction in power, that left 5-10% of power savings on the table. Those would go into increasing the frequency. That can account for 5-15% of the boost in performance, leaving maybe 10-20% unaccounted for, depending on how sharp it curves up on total power relative to the normalized frequency.

    So, literally I have explained already the majority of the performance gains without any allusion and while complying WITH THE LAWS OF PHYSICS AND PUBLISHED PROCESS INFORMATION BY THE FOUNDRIES. You do not make the information agree with each other, which leads to absurdity. I've taken you to task before on this very topic.

    Now, for the last bit of performance, there are factors like making the architecture work with the node, using potentially a hybrid node which would increase the footprint of the die, so that even though transistors density increases, it is still spread out on the die, which can aide in allowing heat to also be transferred off die more efficiently (the problems of miniaturization and transistor density relative to heat are well known; if the density doubled, but the die size only decreases 31%, you are using 19% of space that otherwise would not have needed to be used, spreading out parts of the die to a degree using a hybrid node, which can be part of lessening the heat buildup in specific areas, which then can potentially allow for more frequency to be achieved). Then, you have the potential that the energy efficiency allowed for better results at a set energy draw than expected to make up the rest, or tweaks to caches and IF, etc. All of that makes sense.

    But, what makes no sense is saying they took it all in performance, rather than focusing on power reduction, yet magically cut the power consumption in half.

    I also brought up the use of higher stock voltages to qualify dies, not you originally. That, with smaller die sizes, allows for an easier way to absorb losses on die costs.

    As to what the process was suitable for, it is NOT all about high clocks. There is IPC, memory and cache bandwidths, latencies, etc. that all play into the performance calculations. If the IPC is high enough, a lower clock can outperform a higher clocked chip with a very low IPC. With that said, there is a bit of truth to this, and that too could play into the 35%, but is but one contributing factor to it.

    I already covered them being similar and I also covered that GF changed the transistor pitch to accommodate having designs able to be manufactured at either plant. GF has a slight density advantage at 7nm and a slightly better percentage on power reduction or normalized frequency increase at the same power draw. So, I will stick with my explanation over yours as it does better at getting the known factors accounted for, then pointing at what is left to explain, WHICH IS A BETTER EXPLANATION.
     
Loading...

Share This Page