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:
    4,574
    Messages:
    15,959
    Likes Received:
    19,649
    Trophy Points:
    931
    Dummy Dies? Threadripper examined (en)
     
    ajc9988 and TANWare like this.
  2. TANWare

    TANWare Just This Side of Senile, I think. Super Moderator

    Reputations:
    2,315
    Messages:
    8,876
    Likes Received:
    3,955
    Trophy Points:
    431
    Well there is a direct challenge to AMD as it seems he has caught them in a bit of a fib. Maybe with Ryzen 2 we will see the 32 core but now at 180w we would definitely need the Enermax capability of 500w TDP to handle a 32 core TR. Also a TR at 32 core would that not undermine their own Epyc processor?

    I do think though eve not OC'd it would trash a 7980XE, now that would be cool.

    Edit; thinking about it, I wonder if this is why Epyc requires stepping 2, in order to get the 4 CCX complex's running.
     
    Last edited: Sep 15, 2017
    hmscott, Papusan and Rage Set like this.
  3. Deks

    Deks Notebook Prophet

    Reputations:
    983
    Messages:
    4,038
    Likes Received:
    1,245
    Trophy Points:
    231
    Vega undervolt and overclock as done by german reviewers:
    https://translate.google.de/translate?sl=de&tl=en&js=y&prev=_t&hl=de&ie=UTF-8&u=https://www.hardwareluxx.de/index.php/artikel/hardware/grafikkarten/44084-amd-radeon-rx-vega-56-und-vega-64-im-undervolting-test.html&edit-text=

    Surpasses 1080 performance for a lower power draw than 1080.

    Hardcore undervolting of Vega 56:


    Something interesting to take from this video is the HBM/VRAM voltage.
    Its not that it does nothing... it seems that it acts like another core voltage modifier on Vega, because the VRAM/HBM is part of the same die along with the core, and if the VRAM/HBM voltage was set to too high levels (1.2 or 1.3V), then the lowered core voltage on states P6 and P7 does not remain low/or 'stick' unless the VRAM/HBM voltage is reduced as well (otherwise it gives the gpu the 'go ahead and just volt up to 1.2V regardless of P6 and P7 lowered voltages).


    Man if AMD can auto-tune the voltage out the door, even on the current manuf. process, it would STILL be as efficient or more efficient than Pascal.

    And if they made Vega on the same manuf. process as Pascal, and clocked vega on the core and HBM much higher, Vega would easily smash Pascal's entire lineup (because if Vega needs 1613 MhZ on the core and 1100 MhZ on HBM to surpass 1080 for a lower power draw with undervolting... what would happen if Vega was clocked to 2000 MhZ on the core and HBM to 1600 MhZ - or just the HBM is clocked to 1500 MhZ?).
    Seems like HBM overclock results in about 10% increase for every 200 MhZ roughly. So... HBM clocked to say 1400 MhZ or 1500 MhZ could easily give another 20% boost in performance while increasing power consumption by what... 10-15W?
    The core wouldn't need to be clocked higher than 16013 MhZ in that case... just leave it there and increase HBM.

    Also, the little or no difference in performance between Vega 56 and 64 on same core clocks also might explain the increased power consumption on 64.
    Among the previously mentioned things (such as high clocks not playing well with the manuf. process suited for lower clocks) seem to be the amount of GPU cores.
    The number of GPU cores on Vega is much higher than on GTX... goes well past 1080ti.

    Wouldn't it be better for AMD to reduce/disable the number of cores in that case and increase the clock and HBM frequencies instead while optimizing voltages?
    That might bring them well up to 1080ti and Titan range from just overclocking and less compute units while significantly cutting power consumption (optimized voltages makes Vega similarly or more efficient than Pascal... dropping the extra compute units or at least disabling them for gaming GPU's would likely drop power conusmption and allow the clocks to be set to higher levels).

    And ok... if AMD doesn't want to or can't even mess about with cutomizing each chip... why not disable the GPU cores instead via BIOS, increase HBM clocks dramatically (while less so with the core) and drop the freaking voltages?
     
    Last edited: Sep 15, 2017
    hmscott likes this.
  4. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,150
    Messages:
    4,707
    Likes Received:
    6,577
    Trophy Points:
    581
    shopping list.jpg
    Add to this list 8 vardar F4-120ER, VRM blocks, an XSPC Raystorm Neo sTR4 (black chrome), an EK 780 Classy block compatible with the 980 Ti Classy I own, a pack of 25 XSPC black chrome compression fittings, 25 PWM cable extensions, and you have a good idea of my build. On top of the photon res, I'll run a tube up with a cap to sit in the area where the two top radiators meet (that triangle area that is open), so that it is just removing the cap, screw in a compression fitting with tube on it, and top off when necessary. Thinking of switching the XSPC concentrate for Primochill Utopia and UV blue die. Have thoughts on this @Rage Set , @Papusan , @hmscott , @Mr. Fox
     
    Last edited: Sep 15, 2017
    Mr. Fox, hmscott and Rage Set like this.
  5. Robbo99999

    Robbo99999 Notebook Prophet

    Reputations:
    3,195
    Messages:
    5,523
    Likes Received:
    4,038
    Trophy Points:
    431
    Undervolted & Overclocked Vega56 outperforming GTX 1080 - sounds like a golden sample to me. AMD aren't gonna be using too much voltage for no reason - they gotta keep all those cards stable, golden sample here I think.

    EDIT: mind you there's only 17% performance difference between Vega56 and GTX 1080, in terms of Time Spy Graphics Score so in hindsight I don't find it surprising that a tweaked Vega56 can reach stock GTX 1080 performance: http://www.guru3d.com/articles_pages/amd_radeon_rx_vega_56_8gb_review,30.html
    That's stock GTX 1080 performance, so it's not saying Vega56 is better than GTX 1080, not at all.
     
  6. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,150
    Messages:
    4,707
    Likes Received:
    6,577
    Trophy Points:
    581
    Also, loop sequence:
    Two MCP50X DDC pumps in sequence -> CPU Block -> VRM1 -> VRM2 -> Rad1 with 8 iPPC 3000 -> GPU1 -> Rad 2 with 4 Vardars -> rad 3 with 4 Vardars -> res -> pumps. Nice clean loops. The reason to stick Rad1 there is not to cool the liquid, just so that the tube goes almost straight up as the tube from VRM1 goes from the north most hole to the west most hole on the VRM2, with the East (or right) most hole sitting under the right most hole with the rad positioned for inlet and outlet at the back of the machine. This then drops from the outlet to the VGA card. You use gravity to keep the majority of the line out of sight going from the graphics card to the inlet on the Rad 2 (my window is on the right to see the CPU side), then dropping the outlet tube directly down to the top hole of Rad3. This then goes from the bottom to the bottom of the photon res 170mm. From there, it goes to the bottom second hole of the res. There will also be a drain T right after the second pump.

    To fill, unmount the res and lay system on its side. Connect hoses to the T valve and the res top (that is why the m20 to g1/4", plus a 90 deg. rotary, to a compression fitting, tube, compression fitting, to female to female g1/4" fitting to either a cap or a g1/4" to M20 with cap; drain port is a Tvalve to ball valve with plug when not in use). With both the ball valve open and the drain port open, both with hoses suspended above the system and res oriented coming above the case, add fluid until you can't add anymore. This should allow it to run from Rad3 (lowest point), up through rad 1 before starting to hit issues, allowing for the majority to be filled before a pump is turned on to cycle. At the point the limit is reached, save a little room at the top of the res, close up the res and drain lines, and remove the drain line from the ball valve off of the T line. Stand the case up and mount the res, putting the add line up through the top of the case. Now it is time to cycle the pump 1 to push the air through the system, adding liquid as needed to the res. Repeat as necessary, then the normal tipping, etc. to work out air bubbles.

    Thoughts on the process @Rage Set ?
     
    Last edited: Sep 15, 2017
    hmscott and Rage Set like this.
  7. Deks

    Deks Notebook Prophet

    Reputations:
    983
    Messages:
    4,038
    Likes Received:
    1,245
    Trophy Points:
    231
    AMD is indeed overvolting their gpu's to increase yields... but 1.2V is for 'worst case scenarios'.. .and thus far, we hadn't seen 'worst case scenario Vega' in terms of voltage.

    Stock 1080 performance for a lower power draw than 1080... let that sink in for a second.
    Even an overclocked 1070 doesn't reach that.

    Also, I don't think this would be a golden sample at all.
    Most reference GPU's can undervolt easily enough by one degree or another. Manufacturer produced Vega GPU's will ship with better silicon by default, allowing for lower voltages.
    Most people undervolting Vega 56 report achieving these or very similar results on Reddit.

    Apart from that, Nvidia has optimized voltages out the door. AMD didn't/couldn't due to being a much smaller company and not having access to the same manuf. process as Nvidia (which can get away with an occasionally higher voltage on an early GPU).

    OEM made Vega's will likely ship with lower voltages - but this will probably be negated by the ridiculously high clocks which won't yield too much of a difference.
    HBM overclocks would make far more sense along with lower voltage and moderately increased core.
     
    hmscott likes this.
  8. Robbo99999

    Robbo99999 Notebook Prophet

    Reputations:
    3,195
    Messages:
    5,523
    Likes Received:
    4,038
    Trophy Points:
    431
    GPU performance for an architecture will be remembered most significantly for it's stock performance as sold, and until AMD can start releasing Vega products that run at NVidia performance levels at NVidia power levels then they've not equalled let alone 'won'. It's promising that undervolting Vega can produce some fantastic results that are equal or slightly better than stock untweaked Nvidia Founders Edition cards, but that's not really a win either. AMD need to do more work to produce products that are more competitive, I hope they can tweak Vega & subsequent architectures with that goal.
     
    hmscott likes this.
  9. TANWare

    TANWare Just This Side of Senile, I think. Super Moderator

    Reputations:
    2,315
    Messages:
    8,876
    Likes Received:
    3,955
    Trophy Points:
    431
    If there were a $399 Vega 56 I would have gone there but even the $499 with bonus is $549 so it lost out to the 1080 TI.
     
    hmscott likes this.
  10. Deks

    Deks Notebook Prophet

    Reputations:
    983
    Messages:
    4,038
    Likes Received:
    1,245
    Trophy Points:
    231
    Few things to note though:
    Nvidia's Pascal cards can't really undervolt without underclocking as far as I can tell because the clocks and voltages are tied together.
    AMD gpu's aren't running on optimized voltages and require manual tuning to see what the actual efficiency is... Nvidia voltages are optimized.
    As I said, comparing AMD and Nvidia at this stage is like comparing apples and oranges because they aren't running on same voltages or manuf. process (or clock speeds for that matter - AMD is running on slower clock speeds to boot, not to mention Vega is bandwidth starved, which makes HBM overclocking ideal for performance gains).

    As for the architecture, we know Vega shipped with several features that won't see the light of day until developers optimize for them.
    And, again, Nvidia by default has more frequent support from devs to take advantage of their hardware (which has been pretty much maxed out).
    Vega hasn't even begun to be optimized for by devs.
    The best results we could expect to see I think are upcoming drivers from AMD to optimize Infinity Fabric for games and improved functionality with Wattman.

    I don't care what the perception of Vega is at release because they are comparing an optimized Pascal to an unoptimized Vega.
    People seem to be more forgiving for Nvidia than they are to AMD, and also fail to realize that in DX12 and Vulkan, AMD has a pretty good lead.

    Here's a list of features that we could see with future developer support:
    http://www.pcgamer.com/the-amd-radeon-rx-vega-56-and-vega-64-review/
    "AMD notes five key areas where the Vega architecture has changed significantly from the previous AMD GCN architectures: High-Bandwidth Cache Controller, next generation geometry engine, rapid packed math, a revised pixel engine, and an overall design that is built for higher clockspeeds. "

    Vega as an architecture may have been optimized to run at high speeds, but the manuf. process itself goes contrary to that and results in much higher power draw at higher clocks (undervolting fixes that problem, but if you go past certain clocks, power consumption will skyrocket again as you will need to increase the voltage to stabilize those clocks).

    Other things such as 'next generation geometry engine, rapid packed math, a revised pixel engine' need to be optimized for by devs.
    Don't think a simple driver update from AMD could make a game use those features... only if a game is capable of them (which I doubt would be the case for any game on the market out today).
    I would hazzard a hypothetical guess that Vega's drivers are 'not particularly' ready due to issues with Wattman compatibility, and lack of Infinity fabric optimization for games.
     
    hmscott likes this.
Loading...

Share This Page