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:
    4,882
    Messages:
    17,073
    Likes Received:
    20,976
    Trophy Points:
    931
    Full AMD Computex 2018 coverage, starts at 09:58, sound starts at 13:07. ThreadRipper coverage starts at 01:10:27...

    AMD at Computex 2018
    AMD
    Streamed live 7 hours ago
    Watch AMD’s press conference at Computex 2018 in Taipei, showcasing high performance product leadership and innovation.
     
    ajc9988 likes this.
  2. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,191
    Messages:
    4,883
    Likes Received:
    6,823
    Trophy Points:
    581
    Go to 8:20 on the second video, where he says infinity fabric is leveraged for multi-die! I think I just about finished with that statement! That got me excited. It practically guarantees what I said about Navi. Two to four Navi dies, the HBCC and shared HBM2 is really going to help get that game going more, just saying.

    Edit: He referred to IF to connect GPUs. So far, IF is in PCB and AMD hasn't used a cable to connect Xfire in years, instead switching to PCIe. The die shown only has a single die with likely 32GB of HBM2 (which some reported to be running at 1.2GHz on the HBM2). Good stuff.
     
    Last edited: Jun 6, 2018
    hmscott likes this.
  3. Deks

    Deks Notebook Prophet

    Reputations:
    1,001
    Messages:
    4,145
    Likes Received:
    1,381
    Trophy Points:
    231
    Yeah, it looks like Navi (the consumer version) might be the first using IF to connect multiple GPU dies (it also might explain why they won't be releasing it until next year)... or at the very latest, the GPU's using EUV litography (7nm+) will have a multi-GPU-die design with IF.

    Vega on 7nm (the one being sampled right now for AI) still seems like a big monolithic die... and according to what was said in the article, they are expecting 1.35 times performance increase for half the power draw.
    1.35 times performance increase... what does that translate to?
    35% performance increase over Vega with half the power draw (150W)?

    35% increase in performance is consistent when compared to 16nm TSMC.
    Vega 56 and 64 weren't made on 16nm though... they were made on a 14nmLPP process not suited for high frequencies or desktop parts.
    So, I'm thinking 35% performance increase is a given, yes, but to reach 150W on this new (high performance) process, I think AMD will have more clock headroom (or at least, they should).
     
    Last edited: Jun 6, 2018
    hmscott and ajc9988 like this.
  4. ajc9988

    ajc9988 Death by a thousand paper cuts

    Reputations:
    1,191
    Messages:
    4,883
    Likes Received:
    6,823
    Trophy Points:
    581
    So, for IF, the guy that presented for AMD said Vega 7 with IF this year. I am taking the company insider at his word given at one of the largest presentations of the year for the company. He said he wanted to introduce the Vega 7nm instinct (around 8:05), then described it, then including IF to connect GPUs as a distinguishing factor. That means specifically 7nm Vega multi-die, potentially, this year. Even though for commercial, it makes sense because this was Vega 7nm potentially used as the pipe cleaner at GF, although they may do limited runs there and TSMC for the rest, don't really know. So, also testing multi-die on the more expensive, higher margin products makes sense as well. Public perception is as important as the performance, because some wouldn't even consider AMD due to the name and how far they fell, so to speak. Navi was always going to be next year, though. EUV likely will not go into effect for Navi. Reason I say this is that EUV, due to less masking, and less patterning, can take other techniques to develop. Even with Fabs expected to pick up volume EUV between Q1 and Q3 next year, depending, I would err on the side of caution, instead saying Navi would be 7nm without EUV, just because we also do not know if the EUV transistor pitch and patterning techniques will match between GF and TSMC on the EUV 7nm, just on the 7nm without the EUV.

    As to Vega 7nm, they kept transistor count. This means pure die shrink, making it the size of 242 sq. mm., I believe, which is like 20% bigger than the Ryzen die size. You can call it monolithic, but it really isn't anymore, or at least not when you cut the size in half. Nvidia, as a contrast, increased the cuda cores by 40%, matching the prior gens Ti card, and increased the die size by 27%, matching the die size of the 1080 Ti, rather than the 1080, for its 1180. That is a monolithic die.

    Now, the Vega 64 is about that of the 1080, and the 1080 Ti is about 25% more powerful, with the 1180 about 10-15% over the Ti, supposedly. This means the Radeon Instinct should be better than the 1080 Ti on a chip that is around half that die's size. Now, it seems they got 50% power reduction & 35% performance increase, something I was not expecting. I know with the power reduction, it reduces heat, which can allow better clocks, and that Vega runs HOT (this is beyond dispute at this point that Vega is a furnace), but I would not have expected both. If this does bear out, I will need to revise my beliefs on AMDs ability on multi-die chips coming up. I also agree with the clock headroom. But, I need to take a moment to point out that 10+% of this IS coming from the faster HBM2. Normal HBM2 is clocked in the 800s or up to 950, whereas this HBM2, IIRC, is 1.2GHz. It is practically 30+% increase in speed, plus it has the full four stacks with 32GB and massive bandwidth. So, keep this in mind that not all of the benefit in performance is coming from the die shrink alone, as some is related to the memory used. This really makes it only around Ti levels (without the memory speed) on the core performance, while using less wattage. Not that bad, but when viewed from this angle, it makes a lot more sense for what we are seeing.

    So there is some benefit here from the shrink, AMD took almost all benefit to reduce voltage, but in doing so took off the heat bottleneck, most likely, that allowed for some additional performance (if you look at TSMC's comparison of 16nm and 10nm to 7nm, I think it was, I think I said it was 35% and around 55% on energy; which, if true and correct, places about 5% on the process for performance, and 10+% on memory speed and bandwidth. If they also figured out how to better control voltages and not pump them full well beyond what they need (as seen on every recent AMD product so that more qualified for being incorporated onto PCB for the dies), then that would put a little less taken on the energy savings and the correlating increase in the transistor performance. Either way, not a bad showing. But, that is what my thoughts are on it.
     
    hmscott likes this.
  5. hmscott

    hmscott Notebook Nobel Laureate

    Reputations:
    4,882
    Messages:
    17,073
    Likes Received:
    20,976
    Trophy Points:
    931
    So, both TR2 24c / 32c will draw more power than the lower core count TR1 in absolute terms, but less per core (perhaps), and Intel 28c is just a power sucking (600w+?) thermal demon. :)

    MSI X399 MEG Board for 300W of Threadripper 2 | Computex 2018
    Gamers Nexus
    Published on Jun 6, 2018
    The MSI X399 MEG Creation motherboard is built with a 19-phase VRM -- configured as a doubled 8-phase + 2 VSOC phases, for Threadripper 2 accommodation. Threadripper 2 CPUs will be power hungry at 250-300W, and so motherboards are being designed around for TR2 CPUs on the TR4 socket. MSI's X399 MEG Creation is built with IR 3555 power stages for a 16+3 VRM design (doubled 8-phase), using an IR35201 and IR35204 for the PWM controllers. Learn more in the video!


    28 Cores of Bulls#!t - Intel's "5GHz" Parlor Trick
    Gamers Nexus
    Published on Jun 6, 2018
    We talk about Intel's 28-core, 5GHz CPU unveil at Computex, which we found disingenuous and misleading. Intel used a Xeon processor for its "new" desktop demo, likely the 8176 Xeon CPU, then pretended that the CPU was a brand new desktop part. The company also pushed the narrative of a 5GHz overclock on 28 cores, but neglected to note that they were using a chiller to achieve these temperatures. It's an old CPU -- rebranded as a new one, at that -- and it's overclocked with an unnamed chiller that operated below -10 degrees Celsius.


    Intel's Surprise 28 Core CPU - WE RAN IT!
    Linus Tech Tips
    Published on Jun 7, 2018


    Be Quiet! Dark Rock Pro Threadripper 2 Cooler | Computex 2018
    Gamers Nexus
    Published on Jun 6, 2018
    Be Quiet! showcased the Dark Rock Pro TR4 cooler at Computex 2018, a new cooler suitable for Threadripper 2 by way of large coldplate. Be Quiet!'s showcase this year included focus on a refreshed series of cases -- the Silent Base and Pure Base 601 and 801 cases and the Dark Base 900 rev 2. Aside from this, the showcase highlighted just the new Shadow Wings 2 fans, positioned between the Silent Wings and Pure Wings fans.
     
    Last edited: Jun 7, 2018
    bennyg, ajc9988 and jaybee83 like this.
  6. jaybee83

    jaybee83 Biotech-Doc

    Reputations:
    3,065
    Messages:
    10,436
    Likes Received:
    7,342
    Trophy Points:
    931
    haha i LOVED that video of GN on the 28 core BS by intel :D
     
    ajc9988 and hmscott like this.
  7. hmscott

    hmscott Notebook Nobel Laureate

    Reputations:
    4,882
    Messages:
    17,073
    Likes Received:
    20,976
    Trophy Points:
    931
    2nd-Gen Threadripper...32 CORES!?
    TechLinked
    Published on Jun 6, 2018


    32 Core Threadripper, 7nm Gaming GPUs ARE Coming!

    Gamer Meld
    Published on Jun 6, 2018
    AMD brings it all with a 32 core HEDT, Threadripper CPU! Plus, an update in Intel's 28 core CPU, AMD's first 7nm GPU and Vega 56 Nano! Stay tuned...


    He must have missed AMD's 7nm announcement for Epyc. :)

    FRYZEN Threadripper Cooler
    Optimum Tech
    Published on Jun 6, 2018


    Thermoelectric Liquid AIO Cooler - Cooler Master Prototype!
    Optimum Tech
    Published on Jun 5, 2018
     
    Last edited: Jun 7, 2018
    ajc9988 likes this.
  8. hmscott

    hmscott Notebook Nobel Laureate

    Reputations:
    4,882
    Messages:
    17,073
    Likes Received:
    20,976
    Trophy Points:
    931
    Asrock B450 Gaming-ITX/AC & Radeon RX Vega 56 Graphics Card
    Hardware Unboxed
    Published on Jun 7, 2018
     
  9. yrekabakery

    yrekabakery Notebook Deity

    Reputations:
    400
    Messages:
    1,294
    Likes Received:
    1,264
    Trophy Points:
    181
    Well this is interesting: Jen-Hsun Huang and Lisa Su are second cousins

    7:52 in this video for ppl who understand Mandarin

    https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=en&ie=UTF-8&u=https://www.cw.com.tw/article/article.action?id=5061946&edit-text=&act=url

     
    hmscott likes this.
  10. hmscott

    hmscott Notebook Nobel Laureate

    Reputations:
    4,882
    Messages:
    17,073
    Likes Received:
    20,976
    Trophy Points:
    931
    "Jen-Hsun Huang immigrated to the United States at age 10. In high school, he discovered the magic of computers and after studying computer science and chip design at Oregon State University, where he met Lori, he moved to Silicon Valley to work at AMD and then LSI. "

    Jen-Hsun Huang, NVIDIA co-founder, invests in the next generation of Stanford engineers

    Jen-Hsun Huang (MS ’92 Electrical Engineering) discusses his company and breakfast habits.
    By Stanford Engineering staff, October 1, 2010
    https://engineering.stanford.edu/ne...er-invests-next-generation-stanford-engineers
     
Loading...

Share This Page