New Zen 2 Leak

Discussion in 'AMD Processors' started by DuronBurgerMan, Dec 4, 2018.

Thread Status:
Not open for further replies.
  1. DuronBurgerMan

    DuronBurgerMan [H]ard|Gawd

    Messages:
    1,095
    Joined:
    Mar 13, 2017
    I suspect the bandwidth and memory controller issues won't be as bad as they are with the 2990X, which produces very schizoid performance depending on workload due to its strange configuration. At least all dies will have direct memory access through the I/O die, and won't have to make the extra hop.

    However, we are seeing, in effect, a doubling of cores without a corresponding doubling of bandwidth. No doubt AMD has better, faster memory support, and has addressed some latency issues with their memory controller too (though moving to the I/O die might cancel out some of those improvements). However, they will not have anywhere near double the bandwidth. So I don't doubt that there will be certain instances where core scaling will run into a bandwidth wall.

    Still, that doesn't matter overmuch. So you won't see 100% core scaling. So what? It won't be schizoid like the 2990X, at least. And you're still getting more cores for the money. It's a good move by AMD.
     
    tayunz likes this.
  2. JoseJones

    JoseJones Gawd

    Messages:
    591
    Joined:
    Jun 6, 2012
    Yeah, I agree with you that Ryzen & X570 with DDR4 could be memory starved with dual channel DDR4.


    I'm not confused at all. Here's another video that backs me up saying early in the video at around 2 minutes that the chances of any releases at CES 2019 are zero - he says Computex in June at the earliest with a Q3 fall release possible. There's just no reason to believe otherwise from anything AMD themselves have said.





    and here's an article that backs me up too:

    https://www.techspot.com/news/77769-our-take-amd-zen-2-cpu-navi-gpu.html
     
    Last edited: Dec 11, 2018
  3. NKD

    NKD [H]ardness Supreme

    Messages:
    7,119
    Joined:
    Aug 26, 2007
    For Navi yea sure. I think that will be mid year unless they are well ahead and are ready to mass produce as of now and have enough capacity. But as far as zen 2, I would not be surprised at all if they announce some zen 2 chips with March-April availability along with motherboards.
     
  4. Pieter3dnow

    Pieter3dnow [H]ardness Supreme

    Messages:
    6,172
    Joined:
    Jul 29, 2009
    Techspot = hardware unboxed weird that they have the same opinion in the matter ;)

    I wish I was as talented as the folks from Techspot I would make 12 different websites with the same opinion :)

    Since there no specs available for X570 and people presume it is just the same thing with a different number is weird if that one is specifically for the 16 core AM4 and not deal with obvious issues stemming from having a 16 core product on another platform for the same reason I guess ?
     
  5. schmide

    schmide Limp Gawd

    Messages:
    162
    Joined:
    Jul 22, 2008
    Isn't this a totality?

    Any system operating on a memory bound workload, will be bound by memory.

    STREAM is designed to test the memory subsystem.

    No magic sauce will circumvent a memory bottleneck. This goes for Intel, Arm, nVidia, AMD, IBM, etc.
     
  6. luminousone

    luminousone [H]Lite

    Messages:
    115
    Joined:
    Feb 13, 2011
    His response get a 30% grade, very few real world workloads look like stream, pure sequential memory copy's with zero computational load are rare. Databases, jump around at utter random, but will generally hit certain parts of the dataset more often then others, the top branches of btree's will tend to get into the cache, and get used over and over again, where as the lower leafs of btree's will tend to always be in memory, meaning memory access are generally packed together and bursty, things like coherency, processor CAS(conditional assignment) latency, speculative read ahead, pre-fetching are bigger effects on performance then memory channels per core, after the baseline minimum has been reached(and 8 channels, hell 4 channels hits this on current platforms).

    Similar memory access patterns for virtual machines, or machine learning, or hell your office suite can be observed.

    The biggest performance difference with epyc and threadripper 1000/2000(excluding wx series) chips is having memory in every NUMA domain and a processor scheduler that isn't windows. Threadripper 3000 WX will be a very interesting beast with the NUMA domain memory access issue being made moot.
     
    thebufenator and schmide like this.
  7. juanrga

    juanrga Pro-Intel / Anti-AMD Just FYI

    Messages:
    2,357
    Joined:
    Feb 22, 2017
  8. Zarathustra[H]

    Zarathustra[H] Official Forum Curmudgeon

    Messages:
    26,541
    Joined:
    Oct 29, 2000
    A lot of these pre-launch leaks are pure fabrication. Just practice... well... zen, and wait for the official information. We'll get it sooner or later.

    5.1Ghz, 16 core, even if just a single core max turbo, does seem a little optimistic to me, but I've been surprised before. I'll reserve judgment until we have real silicon through.

    That said, I'd be happy with "only" 8 cores if I can get above 5ghz clocks.

    I just wish there were more PCIe lanes...
     
  9. notarat

    notarat [H]ard|Gawd

    Messages:
    1,624
    Joined:
    Mar 28, 2010
    The current amount isn't "horrible"...
     
  10. DuronBurgerMan

    DuronBurgerMan [H]ard|Gawd

    Messages:
    1,095
    Joined:
    Mar 13, 2017
    Some of them probably are. AMD and TSMC claim a 1.25x performance uplift/watt from 7nm vs 14LP. 14LP brought Zen to 4GHz max boost (4.1 XFR). So 5GHz max boost, 5.1GHz XFR at AMD's 95 watt TDP (which isn't really 95, but whatever) would be in line with that. 1.25x 3.6 base clock would be 4.5, but the leak says 4.3 base. This makes some sense, too, because the leaked 16 core 3850x part is rated at 135 watts. 2x 95 watt would be 190 and would blow that out. They'd need to sacrifice some base clock speed to stay within even AMD's nominal 135 watt spec.

    If it was top quality silicon - cherry picked shit - it would plausible.

    I'd think a 4GHz base, 5GHz max single core turbo is more likely, though. Easier to stay within the power envelope that way. Similarly, a 4GHz base, 5GHz max boost for a 12 core 3700x would make more sense too, for a 105 watt TDP.

    One thing I think the leaks are probably right on - even if they are outright fucking fabrications (entirely possible) is that AMD will be carrying over the chiplet concept, and have a smaller 2 channel I/O die for AM4. This means I think 12 and 16 core products are very likely now.
     
  11. Zarathustra[H]

    Zarathustra[H] Official Forum Curmudgeon

    Messages:
    26,541
    Joined:
    Oct 29, 2000
    Yes it is. 16 lanes only. Plus a small amount to the chipset. That's enough for a GPU and nothing else. I have at least 4 more expansion cards I'd want to use in any desktop build I go after.
     
  12. Pieter3dnow

    Pieter3dnow [H]ardness Supreme

    Messages:
    6,172
    Joined:
    Jul 29, 2009
    Really odd but this somewhat more reliable.
    It depends on the pricing but if we get more then 8 cores and it coincides with better silicon thus higher frequencies why not?
    I can understand why AMD is doing this as it looks now AMD will create a landscape where developers would be stupid not to optimize for the extra cores and Intel will have a tougher time competing then just ramping up the frequency.
    If the ipc rumours are true and AMD can keep this up for the next 5 years people will forget about Intel ....
     
  13. tangoseal

    tangoseal [H]ardness Supreme

    Messages:
    6,549
    Joined:
    Dec 18, 2010
    They fixed it mostly. It's not the chip so much as windows scheduler. Nvidia drivers played a huge role. Also the latest iteration of ryzen master solved a huge issue with the memory pooling via changes to how it dedicates threads to particular numa nodes etc...

    Linux has ZERO of the memory pool issues that windows has. So your assessment sounds mostly like hear say at this point. These problems should hopefully be resolved in threadripper 3000 of which I will be getting as well.[/QUOTE]
     
  14. DuronBurgerMan

    DuronBurgerMan [H]ard|Gawd

    Messages:
    1,095
    Joined:
    Mar 13, 2017
    It still has issues. Yes, yes, the issues are mitigated by better scheduling (or, made worse by shitty scheduling on Windows - take your pick). I never really bothered reading up on the Nvidia driver issue - but it wouldn't surprise me if you were right about that. Fact remains, half the cores have much greater latency than the other half, and this still produces issues, because it's a fundamental hardware/design issue. I'm sure it's worth the tradeoffs for certain workloads - else AMD wouldn't have made them.

    When we have 4 (or more? Rome has up to 8) chiplets connected to a single I/O die as we have with Zen 2, this problem goes away entirely. You won't have that latency penalty on half the cores. That part is not hearsay or rumor, that is fact - working chiplet silicon was already demonstrated by AMD. The scheduler can be retarded and you'll still see the benefits.
     
    Sulphademus likes this.
  15. noko

    noko 2[H]4U

    Messages:
    4,013
    Joined:
    Apr 14, 2010
    One way to simulate memory bandwidth restrictions for dual channel 16 core is to go single channel 8 core and see how the current Ryzen perform. Cannot really use games as a benchmark since 6 core to 16 core the game would use the same bandwidth for the most part and dual channel would be fine unless one loads and use the other cores for something else while gaming.
     
  16. Pieter3dnow

    Pieter3dnow [H]ardness Supreme

    Messages:
    6,172
    Joined:
    Jul 29, 2009
    Leaks explained in more detail.

     
    Randall Stephens and mbelue like this.
  17. bobzdar

    bobzdar [H]ard|Gawd

    Messages:
    1,539
    Joined:
    Jun 6, 2003
    Yeah, that's why they make threadripper, for edge cases, of which adding 4 expansion cards certainly is one in 2018/2019. Side bonus is you'll get 64 cores if you want/need them end of 2019. I'm seriously considering moving to TR4 just for that capability, but reality is I'm only using all 8 cores on certain workloads now, so the am4 path to 12-16 cores should be fine. I would like to build a 3df zephyr benchmark to see how much can be gained by going from 8 cores up to 16-32 and how that compares to CUDA scaling, but only so much time in a day.
     
  18. Zarathustra[H]

    Zarathustra[H] Official Forum Curmudgeon

    Messages:
    26,541
    Joined:
    Oct 29, 2000
    Yeah, I just wish there were something between Ryzen and Threadripper.

    I don't need a gazillion cores. As long as a chip has at least 6C/12T, I think I'd take clock speed over core count. I also do the want to deal with infinity fabric/Numa considerations with ram being on different dies with latency etc.

    So, currently in my system in addition to the GPU I have a sound card, a 10G Ethernet adapter and two PCIe SSD's.

    I could easily get rid of the sound card. I keep it for old times sake, because I don't like the on board sound, but honestly, I have an external DAC anyway, so I don't really need it.

    The 10G Ethernet may be a little unusual in the client world, but I use it for a fast direct link to my NAS.

    The two PCIe SSD's don't feel that extreme to me either. I ran out of space on my 400GB Intel 750, so I added a 1TB Samsung 970 Evo.

    None of this feels like it is all that crazy and corner case.

    I'm actually astonished that so many people have just the GPU in an expansion slot, and that's it. So much of making our systems custom and unique to us has to do with hardware expansion to meet our needs. Take that away and the PC is a shell of it's former self.
     
    TheFlayedMan and drescherjm like this.
  19. bobzdar

    bobzdar [H]ard|Gawd

    Messages:
    1,539
    Joined:
    Jun 6, 2003
    I mean, 1900x would probably fit you perfectly and doesn't have numa issues like the high core count processors. That's what I had considered to jump to the platform and then go more cores later.
     
    Sulphademus likes this.
  20. Sulphademus

    Sulphademus Limp Gawd

    Messages:
    296
    Joined:
    Mar 18, 2010
    Sounds like Mr. Z needs a low core/high speed TR build next year!
     
  21. cyberguyz

    cyberguyz Gawd

    Messages:
    706
    Joined:
    Aug 28, 2014
    I'll most certainly be looking at the 3000 series equivalent to the 1950x (3950X?). I'm skipping right over the whole 2000 series - improvements for me were too incremental.
     
  22. mbelue

    mbelue Limp Gawd

    Messages:
    205
    Joined:
    Apr 11, 2017
    I get nostalgia sometimes about the way I used to build PCs with disparate cards for disparate purposes. Course I kinda miss having separate 2d and 3d graphics cards as well, but that's really going too far. I remember going to LAN parties and looking for those external VGA pass-though's to see who was "really serious."

    Looking at a lot of ATX motherboards and wonder why they exist. All these slots hanging off of the PCH so severely handicapped as to not have much of a reason to exist.

    Seems like the vast MAJORITY of people only NEED ITX boards but would be served well enough by mATX.

    The problem seems to lie in the manufacturers not making premium examples of mATX, neither in motherboards nor in cases. Everything in this form factor is budget oriented, DESPITE the fact that very few people will ever use more than that.

    There's either some kinda fetish about full-ATX or a stigma about mATX that has so many people extremely proud about the amount of empty space in their cases.
     
    Zarathustra[H] likes this.
  23. Zarathustra[H]

    Zarathustra[H] Official Forum Curmudgeon

    Messages:
    26,541
    Joined:
    Oct 29, 2000
    Yeah, but they skipped the 8 core part without Numa issues in the 2xxx series, so now the choice is older lower performing 1900x or newer faster chips but with Numa issues

    I'd LOVE a balls to the wall binned 8 core Threadripper with very high clocks. I'd pay more for that than a 16 core part with lower clocks.
     
    Astral Abyss and Sulphademus like this.
  24. bobzdar

    bobzdar [H]ard|Gawd

    Messages:
    1,539
    Joined:
    Jun 6, 2003
    Yeah, but I think the 1900x is fine, it has roughly the same performance as a 2700X but with the extra pci-e. A potential 2900X wouldn't give too much of a bump, only ~5% if it kept similar clock scaling (estimated 4.0 base 4.4 xfr vs 3.8 base 4.2 xfr). Probably why they didn't bother. I think we'll see a high clocked 16c tr4 end of 2019, but a 1900X would be good to get on the platform if you don't need the cores. It'll do 3.9 all core and 4.2 single/dual at stock (with xfr). It doesn't have the nice clock scaling of the zen+ chips, though it does allow boost on 2 cores instead of single like the non tr4's, so it basically gives 2700X cpu performance on tr4, which isn't too shabby. It's also around the same price as a 2700X, so really the cost is in the motherboard and extra 2 ram sticks. I'm wishing I had gone that route (hindsight is 20/20 afterall), but I may still sell my 2700x and c6h and go 1900x/tr4.
     
  25. DuronBurgerMan

    DuronBurgerMan [H]ard|Gawd

    Messages:
    1,095
    Joined:
    Mar 13, 2017

    Now this is interesting... now he's saying no I/O die?

    That doesn't track - not really. If there's going to be a separate GPU die or a second CPU die... there must be a controller to connect them. Unless they are making a second chiplet design - one with specially-worked out I/O in it - this doesn't work. And even if they did that, it would have some of the same drawbacks as the 2990X.

    IMHO, if there is no I/O die for AM4, it wouldn't make sense to increase core count. Just leave TR as the higher core count product, in that case.
     
  26. Pieter3dnow

    Pieter3dnow [H]ardness Supreme

    Messages:
    6,172
    Joined:
    Jul 29, 2009
    Something to look forward to at CES I guess at least that part is news if we hear it ;)
     
    drescherjm likes this.
  27. IdiotInCharge

    IdiotInCharge Not the Idiot YOU are Looking for

    Messages:
    7,934
    Joined:
    Jun 13, 2003
    Here's one: you need a GPU; maybe you have two SSDs in M.2 slots, maybe one is NVMe/3DXpoint, other is cheaper/larger SATA, use built-in *1Gbps/5Gbps/10Gbps LAN, and sound can be off the board either in analog form or out through optical.

    Assuming you only need two RAM slots, you could get away with ITX here, though you'll want to be picky about enclosure, fan setup, and CPU and GPU cooling.

    [*haven't seen any sub-ATX boards that do this yet, though there's certainly nothing stopping them]
     
  28. Mega6

    Mega6 [H]ard|Gawd

    Messages:
    1,325
    Joined:
    Aug 13, 2017
    Will Zen2 have 3D chip stacking? Similar to Intel's announced stacking?
     
  29. Pieter3dnow

    Pieter3dnow [H]ardness Supreme

    Messages:
    6,172
    Joined:
    Jul 29, 2009
    From what I read not and the way it is described here : https://www.cnet.com/news/intel-3d-chip-stacking-could-get-you-to-buy-a-new-pc/

    Sounds like a test bed for what they intent to produce. There is a good change that this won't scale to larger dies.
    If you watched AdoredTV youtube about chiplets you might have seen some form of it using the interposer for certain parts.

    And it is pretty limited to what is stacked:
    The one thing we have not heard about how the Ryzen 3000 series is using the chiplets without the IO block that is there on the server counterpart (Rome).
     
  30. luminousone

    luminousone [H]Lite

    Messages:
    115
    Joined:
    Feb 13, 2011
    I would have to bet they will have a smaller IO die for Ryzen, and maybe for Threadripper. I would imagine that their is enough value proposition in a couple extra IO die variants, more so considering they will be using 14nm for them.

    2000 series 12 and 16 core parts do not have the NUMA issue, both enabled dies on these parts have attached memory channels. Further the TSV is not the same as the 24 and 32 core parts, as 2 infinity fabric links are used between the enabled dies, This will actually perform better then even the 1900x, including in single thread use.
     
  31. mesyn191

    mesyn191 2[H]4U

    Messages:
    2,942
    Joined:
    Jun 28, 2004
    Or they're using the IF bus which we already know can work without a connector die through the package. No additional controller needed in that case.

    GPU's are highly latency tolerant as are their work loads (games). It'd be no more of an issue than it is now for a dual die CPU/GPU APU.

    For CPU's yeah there could be NUMA issues...but it depends on how bad the latency would be. If the latency is low enough (needs to be around typical main memory access latency) it won't matter much and there'll be minimal NUMA issues. If latency is around where its currently then yeah it'll have the same NUMA issues as current TR/Epyc's which requires good OS/software support to hide the latency and work well.

    AMD is supposed to have greatly improved the IF bus though with Zen2.
     
  32. mesyn191

    mesyn191 2[H]4U

    Messages:
    2,942
    Joined:
    Jun 28, 2004
    No.

    And Intel's die stacking tech, while damn nifty, isn't something that will be practical or affordable for use in a desktop performance part for a long time yet. If ever. Its massively power/heat limited which is why they're focusing on tiny low power integrated devices for it first.

    Expect Intel to do make use of their EIMB to do Zen/IF bus-esque MCM's at some point in late 2019 or early 2020 but probably server focused and probably real expensive.
     
  33. Pieter3dnow

    Pieter3dnow [H]ardness Supreme

    Messages:
    6,172
    Joined:
    Jul 29, 2009
    The problem is the video released after(to good to be true) it clearly says that for consumer Ryzen there is no such solution as an IO block(Rome). I think he even suggested that the Ryzen 3000 are monolithic.

    That there is a solution is clear because there are no 12 core zen 2 dies ;)
     
  34. DuronBurgerMan

    DuronBurgerMan [H]ard|Gawd

    Messages:
    1,095
    Joined:
    Mar 13, 2017
    They could. They'd need a different CPU die for that, though, which strikes me as somewhat inefficient. Cheaper, probably, to do a second I/O die that is, in essence, 1/4 of the Rome I/O die. Then use defective Rome I/O dies for TR.


    I don't see a major issue with doing the 'integrated' GPU this way, per se, except that again, this necessitates more redesign of the CPU die.

    The improvements to the IF were almost certainly needed to negate latency issues created from moving to the chiplet format. If AM4 is going to be dual CPU die, the NUMA issues will hurt AMD going forward.

    Nonetheless, AMD almost has to increase core count on AM4, given rumors of Intel's 10 core mainstream chip.

    IMHO, the smaller I/O die for AM4 makes much more sense. Then you fab one chiplet at TSMC and use binning/salvaging to get your various permutations across Epyc, TR, and AM4. And with APU implementations, you merely swap in crappy salvaged Navi GPUs for the second CPU die. AND you use GlobalFoundries for I/O dies across the range, eating at the wafer supply agreement. One I/O die for Rome, which could also be used for TR in the case of some defects (only half the I/O die would be needed for TR) and a second die for AM4 that is much less complex.
     
    Sulphademus likes this.
  35. Krenum

    Krenum [H]ardForum Junkie

    Messages:
    14,617
    Joined:
    Apr 29, 2005
    Blah Blah Blah. Wait for CES.
     
  36. Pieter3dnow

    Pieter3dnow [H]ardness Supreme

    Messages:
    6,172
    Joined:
    Jul 29, 2009
    Well this is the first time in a long time that there is something to talk about on the AMD front ;)
     
  37. horrorshow

    horrorshow [H]ardness Supreme

    Messages:
    5,629
    Joined:
    Dec 14, 2007
    What? No.

    AMD threads are always happening because people wanna root for the underdog....

    The Fury and Vega threads in particular have been a huge deal in the past etc.

    (Not to mention the many, many Ryzen threads over the past 3 years)

    Until CES though, we're just blowing hot air and throwing darts at an IPC board....
     
    shaolin95 and juanrga like this.
  38. DuronBurgerMan

    DuronBurgerMan [H]ard|Gawd

    Messages:
    1,095
    Joined:
    Mar 13, 2017
    Speculation is fun, provided people understand that it is just speculation.
     
    juanrga and Sulphademus like this.
  39. Pieter3dnow

    Pieter3dnow [H]ardness Supreme

    Messages:
    6,172
    Joined:
    Jul 29, 2009
    Well the promise of parts we all know already been presented when they announced Rome and showed it had some pretty good improvements were all just hoping that the desktop Ryzen 3000 series is something that delivers on expectations.
    This is by far the biggest "jump" for AMD, Ryzen 1000 was still good but it is exciting that what has not happened since a long time there are improvements in the 3rd release of the "same" architecture.
    We have not seen this before or it was so long ago we all forgot about it ;)
     
  40. horrorshow

    horrorshow [H]ardness Supreme

    Messages:
    5,629
    Joined:
    Dec 14, 2007
    Look at my sig.

    No one wants a killer 8c/16t 4.6+ Zen2 more than me lol

    But why gossip for another month, when we know all will be revealed in January?
     
Thread Status:
Not open for further replies.