SMT Performance Benchmarks Continue To Show Benefit With AMD Zen 5/5C

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Ladis
    replied
    Originally posted by mahirzukic View Post

    7zip is just one such program which falls into the first category of 0% data sharing.
    You are cherry picking buddy.
    I already told you overwhelming majority of programs fall into the 3rd category where there is some data during between threads.
    ​Most of the algorithms have own set of data for each thread. Because sharing data is a slowdown, as the communication between the cores has to increase (and sux for Intel's small cores and AMD between chiplets). Also Apple shares a lot of data through the unified memory model (shared L2 cache mong CPU and coprocessor cores).

    Originally posted by zeealpal View Post

    Also, from a uArch perspective which has already been mentioned but not compared here is the transistor budget to achieve the results.
    • Apple M3-Pro: 37b transistors.
    • Apple M4: 28b transistors.
    • Ryzen 9 7945HX3d: 17.8b transistors.
    Obviously there are trade-offs for power vs area, general vs accelerator etc... but it's easy to say a processor is more performant / power efficient, but when it costs 50% more to make it's less of an achievement.

    ​But it doesn't cost 50% more for Apple, as he pays no middle man and orders directly from TSMC. That was the point.

    Leave a comment:


  • qarium
    replied
    Originally posted by mahirzukic View Post
    Once again I call upon @phoronix and Michael Michael to give us some actual numbers. For example amount of RAM used with HT in relation to without HT averaged over 300 something tests he has. I think he could do it with this very processor as well, so that we can compare the RAM increases with the performance increase (18%).
    exactly this is what we need.

    A number like 18% performance increase for a 5000€ cpu if you have to spend 7000€ more in ram is not as good as people may thing.

    in the past people had no choice intel had monopole and there was no high performance option without hyperthreating...

    also keep in mind for gamers its not only ram its also input latency and if you disable hyperthreating you get 5-6% better latency....

    people buy nvidia cards for nvidia reflex for better latency so why waste the latency on hyperthreating ?

    Leave a comment:


  • mahirzukic
    replied
    Once again I call upon @phoronix and Michael Michael to give us some actual numbers. For example amount of RAM used with HT in relation to without HT averaged over 300 something tests he has. I think he could do it with this very processor as well, so that we can compare the RAM increases with the performance increase (18%).

    Leave a comment:


  • qarium
    replied
    Originally posted by coder View Post
    Theoretically, you can spawn threads sharing the same code & data, with one not impacting the L2 cache available to the other. In practice, this level of sharing is probably quite rare and you're right that multiple SMT threads would reduce cache & memory bandwidth available to the other.
    "In practice, this level of sharing is probably quite rare"

    exactly... these people who defend SMT/hyperthreating they claim the exactly opposite of what you say.

    Leave a comment:


  • qarium
    replied
    Originally posted by chithanh View Post
    That is wrong on several levels.
    HyperThreading is an Intel-specific technology which does not exist on non-Intel CPUs.
    SMT does not use appreciably more RAM as long as the threads operate on the same data, thanks to shared memory. It does use more L2 cache and more memory bandwidth.
    And software support matters too.
    Distro choice for ARM is more limited.
    Proprietary Linux software support is not yet very good: Games, Zoom/WebEx/Slack/… client, AnyConnect VPN, etc. are predominantly x86-only

    The way it looks like currently, many users who plan to exit from x86 could just skip ARM, and go directly for RISC-V once it becomes a viable option.
    RISC-V is astroturfing if i want open-ISA i would go with OpenPOWEr

    also all existing RISC-V hardware is pure garbage in 2024 we see us and talk again in 2030...

    man i tested 7zip and the ram usage was double with hyperthreating.

    "Proprietary Linux software support is not yet very good: Games, Zoom/WebEx/Slack/… client, AnyConnect VPN, etc. are predominantly x86-only"

    proprietary linux software? very funny that can die.

    Games can use the x86 emulation.

    i am pretty sure you have intel stock shares at the stock market right ?

    "Distro choice for ARM is more limited."

    well qualcomm aim to mainline and upstream their driver code...

    and it really looks like ARM soon breaks the barrier. and it looks like Nvidia enters the ARM CPU market...

    and if you see the already existing ARM CPUs Nvidia has they are really fast compared to competition with much more core. i think nvidias 72core ARM cpu beats a 128core ARM CPU from ampere ..

    with nvidia going all in with ARM... it looks like ARM will make it and not RISC-V

    Leave a comment:


  • qarium
    replied
    Originally posted by mahirzukic View Post
    It is not. As I have already stated and also referenced a Wikipedia article: hyperthreading is a solution to inefficient core execution allocation. It may or may not use more memory, and it is usually faster.
    Michael did a review of Zen's 5 implementation of HT hugger the other day, here's the review:
    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

    As you can see 18% performance improvement on average. It did not test the memory consumption, but we can ask Michael to answer us here regarding that.
    i never said that you do not get this 18% more performance.

    but if you see the performance per dollar of a notebook you discover that notebooks with less ram are cheaper and any ram inefficiency from hyperthreating kills this performance per dollar metric.

    "It may or may not use more memory"

    what you really do not understand is that a "may" is already to much for me.

    then i buy the qualcomm elite x and there is no "may" anymore.

    Originally posted by mahirzukic View Post
    Also to refute the other statement, 32 GB without HT is not bigger than 32 GB with HT. In fact it is exactly the same, 32 = 32.
    you are clearly wrong here ... 32gb on a qualcomm elite X is clearly more ram than 32gb on a SMP/hyperthreating system like intel or amd systems

    you act like people are fools if intel release their non-hyperthreating cpus people will chose the intel cpu over a AMD cpu only because RAM is in fact expensive and any ram inefficiency from hyperthreating limits you what you can do with the amount of ram.

    i am 100% sure my next mobile device/notebook will clearly not have hyperthreating the qualcomm elite x are fast enough and have no hyperthreating and new intel cpus in mobile will also not have hyperthreating...

    Leave a comment:


  • chithanh
    replied
    Originally posted by coder View Post
    Uh... in exactly what ways does Hyper-Threading "Technology" differ from other SMT implementations? I was always under the impression that it's was just a trademarked name Intel slapped onto SMT, to make it sound Intel-specific.
    HyperThreading is how Intel does SMT. During the Netburst era, it was pretty bad, and delivering worse performance on a number of workloads when compared to HT disabled. Nowadays it is no longer as bad, but still practically all SMT implementations from the competition are better than HT.
    Originally posted by coder View Post
    Theoretically, you can spawn threads sharing the same code & data, with one not impacting the L2 cache available to the other. In practice, this level of sharing is probably quite rare and you're right that multiple SMT threads would reduce cache & memory bandwidth available to the other.
    Indeed. Plus having to run threads in lockstep like this would be contrary to SMT goal of filling unused execution units with work, and thus diminish its benefits.

    Leave a comment:


  • coder
    replied
    Originally posted by chithanh View Post
    That is wrong on several levels.
    HyperThreading is an Intel-specific technology which does not exist on non-Intel CPUs.
    Uh... in exactly what ways does Hyper-Threading "Technology" differ from other SMT implementations? I was always under the impression that it's was just a trademarked name Intel slapped onto SMT, to make it sound Intel-specific.

    Originally posted by chithanh View Post
    ​SMT does not use appreciably more RAM as long as the threads operate on the same data, thanks to shared memory. It does use more L2 cache and more memory bandwidth.
    Theoretically, you can spawn threads sharing the same code & data, with one not impacting the L2 cache available to the other. In practice, this level of sharing is probably quite rare and you're right that multiple SMT threads would reduce cache & memory bandwidth available to the other.

    Leave a comment:


  • mahirzukic
    replied
    Originally posted by qarium View Post

    isn't it funny that if you choose qualcomm elite x with similar performance compared to AMD Ryzen AI 370 then of course you do not have to pick cherry ?

    all these 3 systems have literally the same performance:

    Apple M3 PRO

    Qualcomm Elite X

    AMD ryzen AI 370....

    if you buy every one with 32gb ram then of course AMD ryzen AI 370 is not the best choice if you only see the CPU... because 32gb ram is more if you choose the Qualcomm Elite X or Apple M3 PRO

    you only need to care about the cherry pick part if you choose the AMD ryzen AI 370 because then you need to make sure you do not buy the device to run apps like 7zip or you need to buy a device with more ram.

    people in this forum claim if all 3 options have the same performance and literally the same power efficiency and maybe the same price then the amd ryzen AI 370 is the best option they claim what is plain and simple wrong. ARM wins even if the performance is the same and if the power efficiency is the same and even if price for a model with the same ram is the same... because in the end hyperthreating is a ram inefficiency ..,

    if you admit this then you can start performing cherry picking your software or you can buy a model with more ram but then you have to admit that now you spend much more money.

    all the claims that the AMD ryzen AI 370 is cheaper and better no longer couns if you compare 32gb ram devices with 64gb ram devices. these people in the end pay more money but they still claim ARM devices have no merit what so ever.

    these people are clearly fools. a AMD Ryzen AI 370 would need to be much faster and better to justify the price for the ram upgrade. and of course cherry picking software just to have excuse to not buy more ram does not count here.

    all the irrational X86 fanboys here maybe become very sad if AMD release ARM cpu to .. they could easily do it.

    i am pretty sure nvidia will enter the consumer market with ARM cpus to..
    It is not. As I have already stated and also referenced a Wikipedia article: hyperthreading is a solution to inefficient core execution allocation. It may or may not use more memory, and it is usually faster.
    Michael did a review of Zen's 5 implementation of HT hugger the other day, here's the review:
    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

    As you can see 18% performance improvement on average. It did not test the memory consumption, but we can ask Michael to answer us here regarding that.

    Also to refute the other statement, 32 GB without HT is not bigger than 32 GB with HT. In fact it is exactly the same, 32 = 32.

    Leave a comment:


  • chithanh
    replied
    Originally posted by qarium View Post
    ARM wins even if the performance is the same and if the power efficiency is the same and even if price for a model with the same ram is the same... because in the end hyperthreating is a ram inefficiency ..,
    That is wrong on several levels.
    HyperThreading is an Intel-specific technology which does not exist on non-Intel CPUs.
    SMT does not use appreciably more RAM as long as the threads operate on the same data, thanks to shared memory. It does use more L2 cache and more memory bandwidth.
    And software support matters too.
    Distro choice for ARM is more limited.
    Proprietary Linux software support is not yet very good: Games, Zoom/WebEx/Slack/… client, AnyConnect VPN, etc. are predominantly x86-only

    The way it looks like currently, many users who plan to exit from x86 could just skip ARM, and go directly for RISC-V once it becomes a viable option.

    Leave a comment:

Working...
X