AMD Announces Radeon Instinct Including VEGA GPU powered MI25 Accelerator

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Dec 12, 2016.

  1. Hilbert Hagedoorn

    Hilbert Hagedoorn Don Vito Corleone Staff Member

    Messages:
    40,627
    Likes Received:
    8,993
    GPU:
    AMD | NVIDIA
    AMD today unveiled its strategy to accelerate the machine intelligence era in server computing, alongside the products announced is the MI25 which seems to be offering 25 TFLOPs (16-bit) of performance, and is using a confirmed VEGA GPU. New Radeon Instinct accelerators will offer organizations powerful GPU-based solutions for deep learning inference and training.

    http://www.guru3d.com/news_story/am...ct-including-vega-based-mi25-accelerator.html
     
  2. AlmondMan

    AlmondMan Master Guru

    Messages:
    695
    Likes Received:
    128
    GPU:
    5700 XT Red Devil
    Interesting. But it does say "2X Packed Math" on there - might just mean it's a card with 2 GPUs on there.
     
  3. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    5,748
    Likes Received:
    2,194
    GPU:
    HIS R9 290
    It could mean that, but does it matter? This isn't a gaming GPU; on a server, things like Crossfire are kind of irrelevant.
     
  4. Picolete

    Picolete Master Guru

    Messages:
    320
    Likes Received:
    99
    GPU:
    R9 290 Sapphire Tri-x
    That card looks to long for HBM it might be double GPU
     

  5. AlmondMan

    AlmondMan Master Guru

    Messages:
    695
    Likes Received:
    128
    GPU:
    5700 XT Red Devil
    I was thinking about the notion of using the listed comparison to gauge performance of Vega against Pascal. If its two chips, then a Vega is about the same as a Titan X.
     
  6. kosh_neranek

    kosh_neranek Master Guru

    Messages:
    341
    Likes Received:
    0
    GPU:
    Palit GTX 1070@2101 Boost
    And in other words, AMD is once again on par with nvidia although a year behind.. :( STEP UP AMD,please! Give us ZENs for better price yet same power as Intel and try releasing your GPUs a bit earlier and I will after years get a full AMD setup..
     
  7. fry178

    fry178 Ancient Guru

    Messages:
    1,683
    Likes Received:
    247
    GPU:
    2080S WaterForceWB
    doesnt look good.
    or why does amd release info on 2 products, not available soon and/or not needed for 99.9% of the consumers out there (and most guru3d readers), like the 32C server cpus and now a pro-gpu.

    Hey AMD, how about some info on a cpu supposed to be released NEXT MONTH.??
     
    Last edited: Dec 12, 2016
  8. chispy

    chispy Ancient Guru

    Messages:
    8,954
    Likes Received:
    1,165
    GPU:
    RX 6900xt / RTX3090
  9. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    5,748
    Likes Received:
    2,194
    GPU:
    HIS R9 290
    I know what you mean, but keep in mind we aren't their money-makers. They're trying to appeal to potential investors and large corporations that will rake in millions of dollars worth of sales. AMD will likely make more money from a single company than they will from all potential enthusiast customers combined. The to-be-released CPUs aren't mainstream products.
     
  10. Ryu5uzaku

    Ryu5uzaku Ancient Guru

    Messages:
    7,026
    Likes Received:
    242
    GPU:
    6800 XT
    They really can't. Tech wise they were ahead for a long time. But they gained so little money when they did good that they fell kind of behind in the end tech wise. While Nvidia has been doing good even when doing bad.
     

  11. AlmondMan

    AlmondMan Master Guru

    Messages:
    695
    Likes Received:
    128
    GPU:
    5700 XT Red Devil
    Really, you are asking this question in all seriousness?
     
  12. cyclone3d

    cyclone3d Master Guru

    Messages:
    419
    Likes Received:
    1
    GPU:
    ASUS R9 390
    The slide says that the MI25 is "passively cooled" though.

    So even if it is a dual chip setup, that is Titan X performance for a single chip when passively cooled.

    What happens to max clocks when it is actively cooled?
     
  13. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    7,652
    Likes Received:
    581
    GPU:
    Inno3D RTX 3090
    This is an announcement for HPC people so that they can plan their deployments accordingly. Vega is obviously a single chip card, the 2x Packed Math is a performance metric for doing a specific type of calculation, not an indicator that this is a dual GPU card. The more interesting bit was about memory, it refers to it as "high speed cache" or something like that. I believe we could be seeing mixed memory configurations.

    On the other hand it might be just AMD announcing that the GPUs are more flexible than the competition in this regard (you need to combine this with news of the other HPC card they have with the SSD onboard).

    12.5Tflops for <300W with passive cooling is impressive. If this is a 4096 shader configuration it has a clock of 1525MHz.
     
  14. AlmondMan

    AlmondMan Master Guru

    Messages:
    695
    Likes Received:
    128
    GPU:
    5700 XT Red Devil
    It's meant to sit in a rackenclosure in a serverfarm. The passively cooled part makes sense in that environment as you don't want movable parts inside your servers if at all possible. And it's made possible by rackenclosures being fitted with highperformance, very noisy, fans back and front that make a constant high-airflow through the cases.
     
  15. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    7,652
    Likes Received:
    581
    GPU:
    Inno3D RTX 3090
    Or it might be a mixed solution for the HPC crowd. 16GB of HBM and 32GB of GDDR5x would make a lot of sense for some configurations. The memory info AMD gives for it also mentions "high bandwidth caches and fast memory".
     

  16. Dygaza

    Dygaza Master Guru

    Messages:
    536
    Likes Received:
    0
    GPU:
    Vega 64 Liquid
  17. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    7,652
    Likes Received:
    581
    GPU:
    Inno3D RTX 3090
    Turns out "it might be dual, it might be single, it might be GDDR5x, it might be HBM, it might be cool, it might be hot" WCCF crap was wrong.

    What a surprise.
     
  18. fry178

    fry178 Ancient Guru

    Messages:
    1,683
    Likes Received:
    247
    GPU:
    2080S WaterForceWB
    @schmidtbag
    i know, i know. just funny that they wanna "get somewhere" with the consumer stuff, yet i dont see them do whats needed.

    @chispy
    i know about the even since it was first announced.
    dont really care for another amd slide/10s of vid showing one benchmark.

    if your product can keep up with the competition/beat it, why hide the facts?
    how many info have we seen from Nv way before their last release?
    how much info was out before release of the 4oo series?

    and how did they perform vs what was shown to "us" in the month/weeks before?


    reminds me of Lexus taking 10y to build the LFA to be able to compete with porsche/lambo/ferrari etc, only to see their taillights once they released the next "years" model.
     
    Last edited: Dec 12, 2016
  19. chispy

    chispy Ancient Guru

    Messages:
    8,954
    Likes Received:
    1,165
    GPU:
    RX 6900xt / RTX3090
    +1 This i agree ...
     
  20. pharma

    pharma Ancient Guru

    Messages:
    1,671
    Likes Received:
    485
    GPU:
    Asus Strix GTX 1080
    It's unfortunate the benchmark slide is using GP102 (Titan Pascal) and Maxwell since they lack the double rate FP16. They should have compared "Radeon Instinct" against P100 which is Nvidia's product for deep learning ...
     

Share This Page