New Zen microarchitecture details

Discussion in 'CPUs and Overclocking' started by Dresdenboy, Mar 1, 2016.

  1. JustMe21

    JustMe21 Senior member

    Joined:
    Sep 8, 2011
    Messages:
    234
    Likes Received:
    5
    The only expectation I have with Zen is that it will perform better than Bulldozer. I purchased AMD processors for many years, but Bulldozer always looked like a dud even before it came out. The reason was that even though AMD could beat Intel's Integer performance, their FPUs have sucked since the Pentium days and with Bulldozer was sharing an FPU between 2 cores, that just went from bad to worse. Also, extensions like 3DNow never performed on par with Intel's counterpart. But, if AMD can offer Ivy Bridge performance, at least 3.5 GHz and an 8 core processor for around $300, I'd certainly pick one up. I'd be tempted to go 4 core if they can get it to 4 GHz and around $250.
     
  2. Ajay

    Ajay Diamond Member

    Joined:
    Jan 8, 2001
    Messages:
    4,642
    Likes Received:
    1,264
    Honestly, I hope this is good enough to for AMD to fetch a good price for Zen CPUs. Getting $500+ for the top binned part would be a major win for them financially (and that's the place the need to win - if they are to continue as an MPU/SoC developer).
     
  3. itsmydamnation

    itsmydamnation Golden Member

    Joined:
    Feb 6, 2011
    Messages:
    1,702
    Likes Received:
    601
    No the main prize is selling 64C/128T 2P servers to cloud, enterprise and web 2.0. In all of those markets the already demonstrated clocks are sufficient. Being able to clock Zen high enough to fight intel quads on the left and enough throughput to fight Intel 8-10 cores on the right would just be extra cream. The one advantage AMD has this time VS bulldozer is that most of really IPC hungry app's are now significantly better in terms of threading. So clock disadvantage to intel quads shouln't hurt as much as it might of a few years ago.
     
  4. piesquared

    piesquared Golden Member

    Joined:
    Oct 16, 2006
    Messages:
    1,568
    Likes Received:
    387
    V good points. Id add to that the muticore era in regards to gaming has also just begun with low level APIs able to extract parallelism quite well. High single thread performance on many cores seems pretty good. :)
    I think its important to remember that the idea of low level APIs would have been in discussion within AMD for a fair while, having a working copy with Mantle way before anyone else, so there were likely design decisions made with those in mind also. AMD takes gaming very seriously, so i think ZEN will be a great gaming chip as well as getting them back into datacenters. Anyway, Lisa Su, Mark Papermaster and team look pretty confident and looking forward hearing the rest as Lisa says 'the best is yet to come'.
     
  5. Phynaz

    Phynaz Lifer

    Joined:
    Mar 13, 2006
    Messages:
    10,047
    Likes Received:
    740
    Word is Mark Papermast is doing a deep dive at Hotchips.

    Confirmed. Last session of the last day :(
     
  6. KTE

    KTE Senior member

    Joined:
    May 26, 2016
    Messages:
    478
    Likes Received:
    130
    It's Day 2, by Mike Clark, Lead Zen Architect

    Straight after the other two exciting presentations:

    Inside 6th generation Intel Core code named Skylake: New Microarchitecture and Power Management by Jack Doweck

    POWER9: Processor for the Cognitive Era by Brian Thompton


    I can't bloody wait for David's write up!

    Oh yea, little birdy tells me that frequencies are on target and +3GHz is game.

    Sent from HTC 10
    (Opinions are own)
     
    sirmo likes this.
  7. Dresdenboy

    Dresdenboy Golden Member

    Joined:
    Jul 28, 2003
    Messages:
    1,730
    Likes Received:
    554
    I already tried once, but the new system didn't accept it. Will try again soon. :)
     
    krumme likes this.
  8. raghu78

    raghu78 Diamond Member

    Joined:
    Aug 23, 2012
    Messages:
    3,992
    Likes Received:
    1,203
    yeah zen's primary goal is to gain share in servers. Now it all comes down to clocks. GF 14nm is not quite mature and a lot will depend on whether they can get it to good shape by Q1 2017. I think if AMD hit 2 Ghz base clocks for 32c/64t SKUs they would be competitive with Skylake EP (28c/56T).

    yeah.

    http://www.hotchips.org/program/

    A New, High Performance x86 Core Design from AMD Michael Clark AMD
     
    #2658 raghu78, Aug 19, 2016
    Last edited: Aug 19, 2016
  9. sirmo

    sirmo Golden Member

    Joined:
    Oct 10, 2011
    Messages:
    1,008
    Likes Received:
    365
    Server is lucrative and an important market for AMD to recapture. But where Zen really looks promising to me is AMD's APUs. AMD has never been able to make competitive APUs due to being held back by poor Bulldozer based cores. And if Zen has the perf/watt, AMD's APUs could finally show their true potential. Add some HBM2 into the mix, and you have a unique product no other company can rival.
     
  10. swilli89

    swilli89 Golden Member

    Joined:
    Mar 23, 2010
    Messages:
    1,425
    Likes Received:
    834
    Yeah even a 2C/4T Zen APU with ~800 Polaris/Vega Cores with HBM2 would be an absolute beast in both outright performance, form factor, and performance/watt especially with HBM2 being more energy efficient that traditional memory systems. You could have a Nintendo Wii sized Mini PC that could play games @ 1080p for probably pretty cheap. Apple surely is waiting for this with baited breath in their obsession with small form factors.
     
    Scholzpdx and sirmo like this.
  11. sirmo

    sirmo Golden Member

    Joined:
    Oct 10, 2011
    Messages:
    1,008
    Likes Received:
    365
    I could totally see it.. Apple who loves to solder RAM in their MacBooks.. being told, they don't have to solder it on the PCB it's already soldered onto the CPU core. ... "where do we sign?" :p
     
  12. raghu78

    raghu78 Diamond Member

    Joined:
    Aug 23, 2012
    Messages:
    3,992
    Likes Received:
    1,203
    I agree completely. AMD has the opportunity to deliver a gaming PC on a single package. I think they will wait till 2018 to introduce HBM2 to APUs due to cost and yield reasons. I would like to see a Zen+ combined with Vega and 16GB HBM2 on a single package. Thats the future.
     
    scannall and sirmo like this.
  13. Ajay

    Ajay Diamond Member

    Joined:
    Jan 8, 2001
    Messages:
    4,642
    Likes Received:
    1,264
    Good point. I hope AMD and it's partners can muster enough strength and breadth in infrastructure to move in on some of Intel's Cloud computing market.
     
  14. F-Rex

    F-Rex Junior Member

    Joined:
    Aug 11, 2016
    Messages:
    17
    Likes Received:
    3
    Perhaps amd don't want to let Intel know at what frequencies zen Will ship.
    If i was AMD i would not disclose frequencies until zen launch. Let's Intel worrying a little bit.
     
  15. turtile

    turtile Senior member

    Joined:
    Aug 19, 2014
    Messages:
    410
    Likes Received:
    25
    I doubt we'll see HBM2 on any new product in 2018 since AMD has plans to move to the next generation memory that year. We'll probably only see HBM2 is someone like Apple asks for a custom design.
     
  16. rtsurfer

    rtsurfer Senior member

    Joined:
    Oct 14, 2013
    Messages:
    733
    Likes Received:
    15
    What's next gen after HBM2...?
     
  17. Dresdenboy

    Dresdenboy Golden Member

    Joined:
    Jul 28, 2003
    Messages:
    1,730
    Likes Received:
    554
    Might well be the case. But there's another point: power consumption or efficiency. With those power optimized FP execution units (see my latest blog), uOp cache, high density design (also targeting lower power), there is a good chance for scenarios with Zen cores using less power for a task than BDW cores. This could result in less throttling or higher boost clocks.
     
  18. turtile

    turtile Senior member

    Joined:
    Aug 19, 2014
    Messages:
    410
    Likes Received:
    25
    They haven't stated which technology they will be using. My assumption is that HBM is too expensive to be worthwhile in APUs (at least for low end).
     
  19. jpiniero

    jpiniero Diamond Member

    Joined:
    Oct 1, 2010
    Messages:
    5,356
    Likes Received:
    402
    Navi (GPU) has "Next Gen" memory... my guess would be stacked memory on the die, maybe with the GPU die on top of the memory.

    Whether this would be suitable or cheap enough for a consumer APU is anyone's guess. HBM2 for sure isn't though. I don't know why people think otherwise.
     
  20. rtsurfer

    rtsurfer Senior member

    Joined:
    Oct 14, 2013
    Messages:
    733
    Likes Received:
    15
    Interesting, thanks for the information. I thought we were gonna settle with HBM2 for a few years, guess not.

    The price argument definitely makes sense though.
     
  21. turtile

    turtile Senior member

    Joined:
    Aug 19, 2014
    Messages:
    410
    Likes Received:
    25
    The interposer has to be pretty expensive to implement. I think 7nm will allow for something different than what we've seen before. Apparently, Global Foundries is moving to 7nm next in 2018 but we'll have to see if that happens on time.
     
  22. Abwx

    Abwx Diamond Member

    Joined:
    Apr 2, 2011
    Messages:
    8,561
    Likes Received:
    364
    More than a good chance since they explicitely stated that Zen was consuming a little less than BDW during the Blender rendering comparison at 3GHz..
     
  23. sirmo

    sirmo Golden Member

    Joined:
    Oct 10, 2011
    Messages:
    1,008
    Likes Received:
    365
    HBM2 isn't that expensive. Interposers are pretty cheap, it's not what's significantly adding to the cost of an HBM solution. Interposer can be done on a 90nm process or similar (watch AMD's presentation on it). What adds the cost to an HBM solution is the extra tooling to do the packaging, and the special memory required.. and there are only a handful of packaging places that can do the work. The additional steps in the process of putting dies on interposer also lower yields, because each step adds to the failure rate. Part of the reason AMD offers a free license for HBM2 is so that they can lower the cost of manufacturing. Packaging and economies of scale from SK Hynix.

    Why I mentioned Apple is because Apple more than any other manufacturer would be willing to pay the cost of the difference for the benefits HBM2 offers. Lower power, better bandwidth, and the biggest advantage to Apple in their pursuit of ever thinner devices the space savings.

    Take this mobile Intel CPU: http://ark.intel.com/products/93336/Intel-Core-i7-6970HQ-Processor-8M-Cache-up-to-3_70-GHz with its $620 pricetag. An HBM2 Zen based APU could be built for that price range while providing the cost benefit of simplified motherboard (no need for the memory bus and PCB traces) and integrated HBM2 memory.

    I would argue that AMD could build HBM2 based APUs and still have better margins than they do on HBM2 GPUs.

    Not to mention this same product could have applications in high density blade servers. And other benefits like iGPU not being memory bandwidth starved.
     
    #2673 sirmo, Aug 19, 2016
    Last edited: Aug 19, 2016
  24. Enigmoid

    Enigmoid Platinum Member

    Joined:
    Sep 27, 2012
    Messages:
    2,909
    Likes Received:
    24
    How much HBM2 RAM will such a system have? Remember the RAM is shared. 8 GB is simply not enough for the consumer anymore, least of all going into the future. Motherboards are dirt cheap. That would save maybe $10.

    Intel's mobile CPU prices are not accurate. Remember the desktop prices from ark intel are the prices you pay the retailer (who gets a cut). CPUs sold in bulk to a OEM are far cheaper.
     
    zentan likes this.
  25. sirmo

    sirmo Golden Member

    Joined:
    Oct 10, 2011
    Messages:
    1,008
    Likes Received:
    365
    4 stacks of HBM2 RAM is 16Gb. Which is plenty for most applications. Apple ships 8Gb or less on most of their computers.
    Similar is the case for GPUs as well. Sapphire doesn't buy a GPU from AMD for what we pay in the store. Also there is additional costs in GPUs, like VRMs, heatsinks, card manufacturing.. etc.