AMD Exoscale heterogeneous processor: 32 cores and a GPU with HBM2 in one package

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Apr 9, 2020.

  1. Hilbert Hagedoorn

    Hilbert Hagedoorn Don Vito Corleone Staff Member

    Messages:
    38,996
    Likes Received:
    7,685
    GPU:
    AMD | NVIDIA
    DeskStar, user1, Alessio1989 and 3 others like this.
  2. cryohellinc

    cryohellinc Ancient Guru

    Messages:
    3,302
    Likes Received:
    2,582
    GPU:
    RX 5700 XT/GTX 1060
    DeskStar likes this.
  3. ATi Lazarus

    ATi Lazarus New Member

    Messages:
    3
    Likes Received:
    2
    Surely the ultimate aim for Intel or AMD APUs it to win as many segments as possible and this includes the high end. There are a few segments it's unlikely to win over, things like very specialist server and AI stuff as the ratios of CPU vs GPU vs memory is very different but on the whole you have low cost builds which lend themselves perfectly to APUs all the way up to having massive CPU performance, massive GPU performance and enough HBM3 stacked system memory to be used as the unified system memory for both CPU and GPU so we can get rid of slow and energy expensive off chip DRAM right?

    So lets say AMD builds an APU with 16 CPU core, 120CU and 32GB HBM3. I know this sounds crazily expensive to achieve right now but in a few years when we get to the densities and power reductions of 5nm EUV I really don't think it will be.

    The costs of separately putting together this high end system would be roughly as follows:

    CPU (16 Cores Zen 4) = $500
    GPU (120CU RDNA 3.0 with 16GB GDDR6) = $1200
    RAM (32GB DDR5) = $400 (at least)
    MOBO (high end board) = $600 (a specialist embedded design)
    Water Cooler = $300 (very high end cooling required to control temperatures)

    Total = $3000

    So could AMD make this chip and a profit if they priced it around $1700? I would say yes very easily given they can make almost half of this chips for the XSX for probably less than ~$100!!! The XSX is 360mm^2, so with a move to 5nm I would project this chip to be less than 600mm^2 which is certainly big but doable, nvidia have been making chips bigger 800mm^2 for years.
    Plus if you are selling the entire system kit for around $3000 then there is a lot of profit to be made and a lot of performance for the customer. Win-Win!

    With 5nm densities and efficiencies plus Zen 4 power efficiencies plus RDNA 3.0 being 50% more power efficient than RDNA 2.0 which is 50% more efficient than RDNA 1.0, I think this is all doable and saleable too, so you could have half those specs for half the price. Clocks might have to be tamed a little to keep temperatures under control but I bet the bandwidth and location of the memory more than makes up for lower clocks.
    And maybe the CU count won't be the most important thing to focus on, maybe it will be the number of accelerators they can increase on a self-contained giant, higher performance APU that has insanely wide memory bandwidth plus ultra low latenties from it being on-die.

    I'm not saying this is a perfect fit for all use cases but it would for a lot of gamers, workstations, programming, high-res professional image editing, 3D rendering, etc.
    It would be perfect for something like the Apple Mac Pro 2022/2023 or maybe MS wants to being a desktop Surface machine to compete with Apple in this space?

    Possibly 5nm for all this is a little optimistic and we need to wait for 3nm but I would say this is where we are heading and the winner will be the first company to truly embrace this idea and then push it to the max. We haven't seen a powerful APU come to the desktop yet, it's only low cost, low power stuff but even though AMD seems to be doing all the ground work to set this up maybe it will be Intel that brings something truly special first? There are already rumours they are playing around with big.LITTLE CPU cores, they attempting with build a decent scaleable GPU plus their work on Foveros shows they understand what the future needs such as memory on die but again it's just down to who can nail it first. I guess it's just 'watch this space' but it that's certainly what the ARM based chips appear to be doing when you look at the likes of Fujitsu's A64FX, x86 has to be careful to not be caught napping. If Intel or AMD don't make this chip soon then I think Apple will using ARM or even RISC V architecture to complete relinquish their dependence on Intel or AMD for all of their ranges.
     
  4. yeeeeman

    yeeeeman Member

    Messages:
    40
    Likes Received:
    14
    GPU:
    Geforce 9600GT 512MB

  5. Denial

    Denial Ancient Guru

    Messages:
    13,014
    Likes Received:
    2,420
    GPU:
    EVGA 1080Ti
  6. user1

    user1 Ancient Guru

    Messages:
    1,602
    Likes Received:
    543
    GPU:
    hd 6870
    not really the same thing, fundementally its just pcie connectivity between the cpu and the gpu for that one, this will use the native data fabric/ protocol of the cpu and gpu. (very big gb/s) likely with a shared memory controller.
     
  7. icedman

    icedman Maha Guru

    Messages:
    1,037
    Likes Received:
    110
    GPU:
    MSI Duke GTX 1080
    perfect now scale it down and make something that is console like in performance for desktop/htpc use i love my 2400g in terms of cost for performance but it leaves much to be desired in the graphics area and the low number of CU's as well as ddr4 are big limiters. If they can build this for consoles and sell the entire console for around 500$ then there's no reason they cant do this for desktop use.
     
  8. Khaled Tarakji

    Khaled Tarakji New Member

    Messages:
    1
    Likes Received:
    0
    GPU:
    5700 XT
    U all don't get the important Point of this article, No one done it before.
    Till now all APUs have a Single chiplet for GPU and another one for CPU on the Same die.
    Even Intel with Vega are two different Chips.

    Here they mean in the Same silicone a CPU/GPU together!!!

    No one ever Made such a Thing. And i think the bandwith between the CPU/GPU are incredible with 0ms latency.
     
  9. RoboJ1M

    RoboJ1M New Member

    Messages:
    4
    Likes Received:
    0
    GPU:
    GT430
    Infinity fabric, which is PCIe I believe, but yes, AMD is far far ahead on this stuff than Intel
     
  10. RoboJ1M

    RoboJ1M New Member

    Messages:
    4
    Likes Received:
    0
    GPU:
    GT430
    No, it's the other way around
    All integrated graphics are on-die currently

    AMD designed HSA, enabling real heterogenous compute.
    Not that it's closed, Intel could do it too.

    But so far, there's no sign of MCM APUs, meaning chiplets and scalability.

    But trust, chiplets is not a limitation it's the answer
     

  11. RoboJ1M

    RoboJ1M New Member

    Messages:
    4
    Likes Received:
    0
    GPU:
    GT430
    AMD have been building towards this on their zero budget for decades.

    But everybody choose Intel instead.
    Luckily, AMD have backed the right technologies, took the risks and yes, very soon APUs will replace off board GPUs.
    Soon, off board GPUs will look as dated as off board cache (Google it! )
     
  12. RoboJ1M

    RoboJ1M New Member

    Messages:
    4
    Likes Received:
    0
    GPU:
    GT430
    Beautiful.
    Absolutely glorious.
    I've been watching AMD and the RTG slowly work towards implementing HSA 1.0 which is a world without GPUs, where is PCs are powered by exascale APUs.

    I want one.
    I want to write code for it.
     
  13. user1

    user1 Ancient Guru

    Messages:
    1,602
    Likes Received:
    543
    GPU:
    hd 6870
    The infinity fabric can refer to both the protocol (which can run over amd's gmi controllers, which are not pcie and are using for inter-die connections on the package like this product, or it can run over pcie as used in Dual socket boards. ), and to the internal data fabric.
     
  14. Andrew LB

    Andrew LB Maha Guru

    Messages:
    1,173
    Likes Received:
    187
    GPU:
    EVGA GTX 1080@2,025
    Umm.... ever heard of nVidia Tegra? Putting CPU and GPU on the same die with memory since 2010.
     
  15. MorganX

    MorganX Member Guru

    Messages:
    129
    Likes Received:
    9
    GPU:
    Red Devil Vega 64
    Everyone is just overlooking the importance of the memory stacking. This would also most likely leverage a unified memory architecture as well.
     

  16. Venix

    Venix Maha Guru

    Messages:
    1,358
    Likes Received:
    502
    GPU:
    Palit 1060 6gb
    Computation is free , almost free consumption wise , what is really expensive is moving the data around i guess cost less pico joules to request data from hbm on the package than the system ddr maybe even tenfold.
     

Share This Page