hey devs,
Today i stumbled on something that piqued my interest, but i couldnt find an answer on google:

intel is throttling MMX instructions on their newer CPU's.
my question is: why?

i didnt find either the response to that question nor even the acknowledgement of that fact, yet...

(the interesting stuff starts at 8.30)

  • 1
    1. Intel makes the money with limiting features, but thats not the problem
    2. since intel released mmx hundreds of new instructions were added, the parsing takes a bit of time
    3. with SSE4 MMX is not supported anymore.
  • 0
    didnt think instructions had to be parsed.
    thought they were directly wired into the relevant component
  • 3
    I don't know the author nor am I a fan of YT videos related to benchmarkings if they don't explain stuff in a scientific, detailed way.


    That's ... Ancient.

    Usually a processor architecture lives for 5 - 6 years, till it's redesigned.

    2 generations old is a lot.

    It's a guess, but I don't think Intel crippled MMX in any way.

    Phenom and Intel at that time were two completely different mindsets regarding architecture, and Phenom was - all in all - a desaster for many reasons.

    What is true is that when AVX is used, the core frequency must be lowered.

    Depending on width of instruction there are 3 frequency levels:

    L0 - nominal speed, 128 bit
    L1 - 256 bit
    L2 - 512 bit

    There are further distinctions, regarding the type of operation (heavy / light) and the usage of an FPU if I remember correctly.

    I think AMD does the same, not sure though.

    Mr Torvalds was pretty clear that AVX 2 and all it's complications are ... Shit.
  • 1
  • 0
    thanks. thats a lot of info to take in, but im not exactly sure if this concerns MMX per se.
    however what im sure of, is that i love my floats.
    but then thats what gpu's are for...
  • 0


    AVX isn't MMX, but both belong to the category of SIMD extensions.
  • 0
    all right, but if i understand it right, only avx needs lowering cpu clock.

    and now that i found a description how RISC processor handles opcode it makes all the sense in the world as to why intel wants to drop support for mmx, correct me if im wrong:

    its all in the name. complex instructions.
    theres a layer of logic between the original opcode and what the processor actually does.
    and all that needs to be encoded in hardware.
    any instruction set would make the instruction decoder exponentially bigger, and with every layer of transistors, slower.

    intel has trouble with miniaturizing their process.
    so theres a limit on how much stuff they can cram into the chip.

    so they make MMX obsolete to cut down on the no. of transistors they need for the instruction sets.

    also single thread performance is a selling point for intel, so any way they have to cut down latency is a boon for them.
  • 0

    finally, since they seem to have trouble with increasing the precision of their manufacturing, and thus their core count and cache, they go for the next best thing: increasing data throughput per process through new instructions. AVX 512.

    again, correct me if im wrong, im just throwing my ideas at the wall, see what sticks...
  • 1
    Ok. I'm currently bit fluffy in the brain so let's take this out one by one.

    I will drop some references, as brain is only partially working, and I don't claim to be correct at all.


    If you wanna dabble into a deeper understanding, I'd recommend anandtech. They usually have a very detailed explanation of stuff.

    But this is very very low level. It will require research, as with nearly every architecture some things change.

    Generally speaking, the idea of x86 extensions is to optimize complex operations.

    Like in programming, you can deal best with complexity if the operations have a fixed ruleset.

    So... An x86 extension is just a function with fixed parameters.

    Hence, many compilers can give a _limited_ auto optimize utilizing SIMD instructions and other extensions.

    The tricky part is now how the CPU handles it.

    That's the part where it get's funky.

    Depending on architecture, usually a decoder cracks down and optimizes the incoming stream of commands to a set of operations.

    These operations can be transformed, e.g. split or merged, and then get "executed".

    In a nutshell: What goes in will be optimized internally by the CPU and then executed.

    The reason why X86 extensions are necessary is because of physics.

    Frequency of a processor is not unimportant, but it's limited by physics and hence there was a need to "circumvent" the limitations.

    Cache Architecture, hyper threading, X86 extensions, transistor design, research of conductor materials, speculative scheduling... Tons of other stuff.

    The goal is to squeeze out more with less.

    MMX is redundant since SSE 1 came out.

    It's design was made for the architecture at that time - and lived in reality I think only 4 years or so.

    That's the problem of X86: the further the architecture advances, the more "troublesome" it is to keep all the extensions supported (CPU & software). Yet they need them to beat physics.
  • 0
    this is cookies for my brain. many thanks.
    sorry if i kept you away from (surely) well deserved rest.
  • 0
    @bad-frog nah.

    Brain is currently fluffy because of cold and sinusitis.

    Weather is crazy here in Germany, seems like my body played the "I can be a bitch, too" card.
  • 0
    fak. body needs some time off too it seems.
    i was told zinc is a good thing against colds.
    its used against covid too.
    ginger can help. albeit i couldnt say what dosage.

    well, in a way i could.
    father eats it all the time and while his concubine got sick with covid he didnt get it without any sort of precautions.
    well, apart from eating ginger.

    the posology would be: each morning, onto the sandwich, quantity: as much as if you would put tomato onto it.
    from experience: works well in a tea too. but it has to be fresh in any case.
    when used in tea, theres no upper limit. just let it infuse for some time. its actually quite tasty if you add honey.

    saw a paper on the effects of various compounds found in food, and gingerol was among the best.
  • 0

    there is another paper where you got a breakdown of various compounds in edible plants, but i havent got the time to find it right now.
    i will come back to you once i find it (in approx 2 hrs)
  • 0

    ginger (Zingiber officinale) is good for you, for the cold and against covid if consumed fresh or infused from fresh.
    in any case it wont hurt, and it is proven it will help.

    also you shouldnt OD' on it if you eat it like you would comfortably eat fresh tomato, or if you infuse it.
    eat in sandwich or salad so it goes down better.

    now the rest:
    i wont find the original.
    i read it almost a year ago, and plenty of newer papers got out since then. (last april)

    here is the closest i found:

    here is the interesting part:

    in the latter you got 3 caracteristics:

    -anti-inflammatory (self explaining)

    -inmmunomodulatory (prevents you from dying from cytokine storm. its when your immune system freaks out and nukes the body)

    -ACE inhibitor (hinders the virus in infecting the cell)

    you want all three to be yes, especially the last one.
  • 1
    Watch Gamer's Nexus for actual informative reviews and benchmarks
  • 1
    tx for the advice. i already do:)
Add Comment