AMD Radeon (big NAVI) again rumored to get 80 Compute units

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Aug 3, 2020.

  1. Fediuld

    Fediuld Master Guru

    Messages:
    515
    Likes Received:
    264
    GPU:
    AMD 5700XT AE
    Deleted member 213629 likes this.
  2. cucaulay malkin

    cucaulay malkin Master Guru

    Messages:
    201
    Likes Received:
    93
    GPU:
    3080 (1) 1070 (2)
    no wonder mi100 has IF links,mi100s are linked with an IF bridge
    even RVII Pro cards are linked with IF bridge
    Maybe they'll think about using it on Navi,but CF support is dead

    your logic is weird.with old gcn they made more of an attempt to compete in the high end than with navi.
     
    Last edited: Aug 4, 2020
  3. Is just a re-tool of existing SOC in the consoles far as the AIBs are concerned. Really a niche product?
     
  4. Fediuld

    Fediuld Master Guru

    Messages:
    515
    Likes Received:
    264
    GPU:
    AMD 5700XT AE
    MI100 is CGN (Arcturus an evolution over GCN Vega) not Navi and don't confuse Crossfire with Infinity Fabric.
    We know from the publication releases IF2 and IF3 are effectively making the whole system a big APU between various parts were even the VRAM can be used by the CPU.

    Also on IF3 is relying the upcoming MCM generation of GPUs from AMD, while we have no idea how it NV will use the MCM on their cards from next year (2021).

    So comparing IF with Crossfire is like comparing Q6600 glued cores to Zen 2 based CPU.
     
    Last edited: Aug 4, 2020

  5. cucaulay malkin

    cucaulay malkin Master Guru

    Messages:
    201
    Likes Received:
    93
    GPU:
    3080 (1) 1070 (2)
    so how is this IF link used on navi 21 as opposed to linking two cards ?
     
  6. -Tj-

    -Tj- Ancient Guru

    Messages:
    17,039
    Likes Received:
    1,863
    GPU:
    Zotac GTX980Ti OC
    ^
    Via pcie?
     
  7. cucaulay malkin

    cucaulay malkin Master Guru

    Messages:
    201
    Likes Received:
    93
    GPU:
    3080 (1) 1070 (2)
    IF with no bridge ?
    and how is this going to do anything more than CF ?
     
  8. Denial

    Denial Ancient Guru

    Messages:
    13,152
    Likes Received:
    2,648
    GPU:
    EVGA RTX 3080
    Infinity Fabric is more than just a physical bus, it's also a communication protocol that can be applied across different bus's like PCI-E.
     
    -Tj- and Fediuld like this.
  9. k3vst3r

    k3vst3r Ancient Guru

    Messages:
    3,428
    Likes Received:
    49
    GPU:
    Nitro+ 5700XT
    CF no longer works because of software support, game devs don't wish to add additional work by programming for it. IF is different it's hardware solution to scale up multi gpu dies, no software needed IF links the dies together at hardware level, so basically you get CF levels of performance without software problems associated with CF solution. AMD can just keep gluing more dies together to increase gpu performance using IF will also be cheaper than using monolithic gpu dies.
     
  10. JamesSneed

    JamesSneed Maha Guru

    Messages:
    1,083
    Likes Received:
    449
    GPU:
    GTX 1070
    There is cache coherency which will be great for easing certain programming tasks.

    For the M100 in particular each GPU has two Infinity Fabric links, running at 1333 MHz, with a per-direction bandwidth of 42 GB/s. This brings the total bidirectional bandwidth to a 168 GB/s—more than twice the PCIe 4.0 x16 limit of 64 GB/s.

    So in short its better than CF because of reduced latency, increased bandwidth, and better communication protocol which allows things like cache coherency. Basically its better than CF in every measurable way.
     

  11. cucaulay malkin

    cucaulay malkin Master Guru

    Messages:
    201
    Likes Received:
    93
    GPU:
    3080 (1) 1070 (2)
    on hardware level it seems
     
  12. JamesSneed

    JamesSneed Maha Guru

    Messages:
    1,083
    Likes Received:
    449
    GPU:
    GTX 1070
    y
     
  13. negates software layer for driver support is entirely self contained hardware & protocol *gosh :D
     
  14. Venix

    Venix Maha Guru

    Messages:
    1,405
    Likes Received:
    524
    GPU:
    Palit 1060 6gb
    @Denial the thing of the gtx 480 was that it was running too hot 95 Celsius or so and was the first card ever to have a sticker warning about serious burns in touch ... The 250 watts was not that much of an issue if it was maxing out to 80-85 degrees.... But the combination of that first seen sticker temperatures to warrant it etc gave bad rep to the card was different times i remember seeing the 9700 pro requiring 4 pin molex and thinking darn that's too much .
     
  15. Fediuld

    Fediuld Master Guru

    Messages:
    515
    Likes Received:
    264
    GPU:
    AMD 5700XT AE
    AMD didn't use bridges since 2013, everything goes through PCIe.

    Infinity Fabric is hardware interconnect of multiple components for the AMD ecosystem.
    Upcoming Infinity Fabric 2 (Zen 3/RDNA 2/CDNA) and especially Infinity Fabric 3 (Zen 4/RDNA 3) are making the whole system effectively a behemoth APU where everything is accessible and shareable, including the VRAM to the CPU. (look at the new consoles).

    Also AMD said last October that all these are coming with the new I/O designed for the Zen 3 which is a very intelligent chip similar to the one used on PS5 which has direct access (via IF) on the NVME drives and sound chips also (where ever applicable). PS5 uses a 1 CU graphic chip for it's Tempest sound processing which can be ordered to use it's process power to boost graphics when not used and is connected through IF to the rest of the system.

    Ofc those going down the Intel/Nvidia route won't benefit from anything of the above.
     
    k3vst3r likes this.

  16. cucaulay malkin

    cucaulay malkin Master Guru

    Messages:
    201
    Likes Received:
    93
    GPU:
    3080 (1) 1070 (2)
    so how is this going to function on rdna2 ? I'm not asking for zen 3 or console audio information so stay on topic.
    how will this differ from the usual AFR/SFR ?
     
  17. Fediuld

    Fediuld Master Guru

    Messages:
    515
    Likes Received:
    264
    GPU:
    AMD 5700XT AE
    It won't work on non AMD system and it would work the same as if you have 1 GPU.
     
  18. cucaulay malkin

    cucaulay malkin Master Guru

    Messages:
    201
    Likes Received:
    93
    GPU:
    3080 (1) 1070 (2)
    not what I'm asking.
    since you corrected me that CF and IF is like q6600 to zen 2 how is this IF link thing on rdna2 gonna be 15 years ahead of what we have now ?
     
    Last edited: Aug 5, 2020
  19. Fediuld

    Fediuld Master Guru

    Messages:
    515
    Likes Received:
    264
    GPU:
    AMD 5700XT AE
    Still I cannot understand what you cannot understand. AMD the last 3 years used Infinity Fabric to link successfully multiple CPU chiplets to a single logical unit (CPU), is now using IF to link GPU chiplets and CPU chiplets in the same machine into a big APU.

    The GPU "chiplets" could be multiple monolithic graphic cards or multiple MCM GPUs. The system would be agnostic and a game will see the whole lot as 1 GPU.
    Also wrote the term "APU" because the CPU will have direct access to the VRAM of the GPUs as a single entity.

    SLI (and Crossfire) idea is archaic tech designed at the start of the century by 3dFX. We are 20 years later and things moved on.

    Now how Nvidia and Intel would try to combat this, I do not know. For certain AMD will keep IF locked to it's ecosystem as is a big boon to sell products especially servers.
     
    Deleted member 213629 likes this.
  20. Denial

    Denial Ancient Guru

    Messages:
    13,152
    Likes Received:
    2,648
    GPU:
    EVGA RTX 3080
    Most of this isn't happening anytime soon.

    They'll use IF to communicate across PCI-E to the GPUs but if you run multiple-GPUs it's not going to show as one big GPU. They are still really far away from doing that and it would require massive, sweeping changes to the scheduler of the GPU + it would probably never happen to discreet GPUs, only across chiplets themselves.

    Nvidia already has NVLink - Nvidia's problem is that it can't get NVLink in CPUs outside of IBM.. which is probably where the whole buy a stake in ARM rumor is going. Intel had Omniband but they recently discontinued it.

    It will probably offer similar benefits that NVLink does over the old SLI bridge on Nvidia cards.
     
    Last edited: Aug 5, 2020

Share This Page