• 0 Posts
  • 23 Comments
Joined 1 year ago
cake
Cake day: October 25th, 2023

help-circle



  • The thing about codec support is that you essentially have to add specific circuits that are used purely for decoding and encoding video using that specific codec. Each addition takes up transistors and increases the complexity of the chip.

    XMX cores are mostly used for XeSS and other AI inferencing tasks as far as I understand. While it could be feasible to create an AI model that encodes video to very small file sizes, it would likely consume a lot of power in the process. For video encoding with relatively high bitrates it’s more likely an ASIC would consume a lot less power.

    XeSS is already a worthy competitor/answer to DLSS (in contrast to AMD’s FSR2), so adding XMX cores to accelerate XeSS alone can be worth it. I also suspect Intel GPUs use the XMX cores for raytracing denoising.







  • How do you manage to draw this out for over 20 minutes? He even gets some of the basics like Zen 4 FCLK/UCLK sync wrong.

    It’s this simple:

    UCLK is the clock frequency of the memory controller

    MCLK is the clock frequency of the memory

    FCLK is the clock frequency of the infinity fabric interconnect.

    On Zen 1, these clock frequencies are always in sync.

    On Zen 2 and Zen 3, running UCLK and FCLK at the same frequency reduces memory latency by a significant number of clock cycles. The goal is generally to run at the highest possible UCLK and FCLK (whichever caps out lower is the limit).

    On Zen 4, running UCLK and FCLK at the same frequency provides no memory latency reduction. The goal here is to run at the highest possible MCLK and FCLK. UCLK = MCLK/2 is a very small performance deficit, so the tradeoff makes sense even if you only gain 10% MCLK.






  • Nowhere, like it should be. The launch was a disaster because AMD pushed the clock/power targets extremely high in order to “compete” with the 780 and 780 Ti, with the end result being a jet engine in terms of noise.

    It wasn’t until February before MSI and ASUS released better cooled cards, at which point the damage from reviews and lack of holiday sales had accumulated.

    In the 7970’s case, AMD had at least the benefit of being 3 months early compared to Nvidia.