• How AA Robotics works [Embodied AI via Frontier LLMs] (Was: For iGPU which acts as a APU RAM is shared)

    From Mild Shock@janburse@fastmail.fm to comp.lang.prolog,sci.physics,sci.logic on Fri Oct 30 10:44:48 2026
    From Newsgroup: sci.physics

    Hi,

    Ok, this is interesting. 2-3 years ago I
    had a vision. Some poor sweating waiters and
    waitresses, in Hotel Resort, could easily be
    replaced by robots, serving the drinks.

    This would be an absolute win for both sides,
    but the obstacle is probably also a natural
    language interface. And in the future LLMs
    could be key to solve the problem, even multi-

    lingual. Now here some benchmarks with

    Welcome to the era of embodied AI:
    Introducing AA-Robotics v0.1
    https://artificialanalysis.ai/articles/aa-robotics

    "Frontier" models. Well not extremly "Frontier"
    they are already out of the lab and used for toy
    explorations, like here from artificialanalysis.ai.

    Woa! That was quick. 2-3 years from back
    then when I read Determined: A Science of Life
    Without Free Will, by Robert Sapolsky.

    Have Fun!

    Bye

    P.S.: They have also benchmarks comparing the
    new N1X Laptop with a RTX 5070. The RTX 5070
    still wins. So lets give it 2-3 more years

    until the form factor and energy footprint
    has become even smaller.

    P.S.S.: Here the "simple" LLM brain architecture:

    Welcome to the era of embodied AI:
    Introducing AA-Robotics v0.1
    https://artificialanalysis.ai/articles/aa-robotics

    [Picture Architecture]

    P.P.P.S.: There are currently also variants of
    LLM architectures making LLMs more responsive,
    bringing down certain decision loops to a few

    milliseconds. Even with a N1X and like 45 tokens/sec
    this is still a little slow to watch pouring a
    bottle of liquid into a glass:

    [Film Robotarm]

    Mild Shock schrieb:
    Hi,

    For iGPU which acts as a APU RAM is shared,

    an integrated GPU (iGPU) used as an APU
    relies entirely on shared system RAM.
    Because the processor and graphics chip
    live on the same silicon die, they do not
    have their own separate video memory
    (like discrete GPUs do with dedicated VRAM).
    Instead, the iGPU carves out a portion of
    your regular system memory to store textures,
    frame buffers, and other graphics data.

    While AMD originally coined Accelerated
    Processing Unit (APU) in 2011 to highlight the
    fusion of CPU and GPU on a single die,
    modern AI computing has redefined what
    it means to "accelerate" data. When Intel
    and Apple discuss their latest processors,
    they frequently highlight dedicated hardware
    blocks called NPUs (Neural Processing Units) or
    Neural Engines right alongside the CPU and GPU.

    https://www.articsledge.com/post/accelerated-processing-unit-apu

    Bye

    Mild Shock schrieb:
    Hi,

    So memory bandwidth is no issue at all for
    this kind of backtracking problem. Its all
    about arithmetic bandwidth . And since I am

    not using vectors or matrices, it is still
    poor. You can do the math. The CPU has 16
    logical cores (8 physical with hyperthreading).

    The GPU single-threaded is around 5 times
    slower than the CPU single threaded, but
    their number is greater. Now you can do

    the math as follows:

    16 * 11.4 / 1.7 * 5 = 536.47

    Pretty much the number of shader cores (*),
    i.e. 512, that the Ryzen AI 7 350 laptop
    offers in its integrated GPU Radeon 860M.

    LoL

    Bye

    (*) Unified shading units (also referred to
    as unified shaders, shader cores, or stream
    processors) are flexible processing components

    inside a graphics processing unit (GPU) that
    can handle any type of rendering or compute task.

    Mild Shock schrieb:
    Hi,

    We recently implemented a parallel -C-WAM
    on a GPU backend and could demonstrate an
    estimated 11.4 Giga Lips. In this post we
    report a further experiment, this time
    presenting a parallel -C-WAM on a CPU backend,
    that can lift specialized Prolog, currently
    to 1.7 Giga Lips performance.

    Having an excess number of threads is a
    bad idea. What if we do context switching
    on our own? With this approach we could
    bring down the execution time of 128 Hack
    VMs by 33%. We estimate for the test which
    had 11.4 GLips on the GPU, that we reach
    1.7 GLips on the CPU.

    Bye

    See also:

    Parallel -C-WAM: 1.7 Giga Lips on a CPU
    https://medium.com/2989/8a984e75af44

    Mild Shock schrieb:
    Hi,

    TSMC's 2nm (N2) node and the 7 GHz target
    represent a monumental milestone in silicon
    manufacturing, shifting away from legacy
    FinFETs to advanced Gate-All-Around (GAA)
    nanosheet transistors.

    This evolution allows chip designersrCoparticularly
    in the PC and AI sectorsrCoto push processor
    clock speeds previously thought impossible
    on standard nodes

    AMD confirms Zen 6 rollout for its July 22 AI event
    https://videocardz.com/newz/amd-confirms-zen-6-launches-in-less-than-two-weeks-starting-with-epyc-venice


    Production of the 6th Gen EPYC family is
    already ramping up. AMD says Venice is its
    first high-performance computing product
    manufactured using TSMCrCOs 2nm process
    technology.

    Bye

    Mild Shock schrieb:
    Hi,

    For marketing purposes people
    typically look at the race towards
    2nm, and we find:

    A month ago, Apple lost its exclusivity
    on 3 nm smartphone processors with
    MediaTekrCOs Dimensity 9400 chip, integrated
    in the Vivo X200 Pro smartphone. Qualcomm
    is also in the race with its recently
    unveiled Snapdragon 8 Elite and set to
    power the Xiaomi 15 Pro in 2025. However,
    Apple should regain its position as innovation
    leader in 2026 with the release of the
    iPhone 18, which should feature the A20
    chip built on TSMCrCOs 2 nm process."

    But there is a vertical vias revolution
    going on as well, some SOCs typically
    being at 18 layers now:

    Zooming Into a CPU (It's Incredible)
    https://www.youtube.com/watch?v=Bez-2cvYja0

    imec has coined the term CMOS 2.0:

    LEUVEN (Belgium), MARCH 12th, 2026 rCo Imec,
    a world-leading research and innovation hub
    in advanced semiconductor technologies, has
    launched a first-of-its-kind consortium with
    26 European university groups that will jointly
    work on the technology roadmap beyond
    CMOS scaling (CMOS 2.0).
    https://www.imec-int.com/en/press/imec-launches-university-consortium-around-next-generation-chips


    So we might see more mobile grade GPUs.

    Bye





    --- Synchronet 3.22a-Linux NewsLink 1.2