Asrock A620 6000mt performance worse than 4800mt

Jump to latest Follow Reply
Status
Not open for further replies.
R

Rekonn

Seniorius Lurkius
32
  • #1
    Hi, I just built a pc to mine Monero with xmrig. I got ram rated for 6000mt, but when go into the bios and change the DRAM Profile Setting to use that, I get worse performance, a lower hashrate from xmrig. Based on some YouTube videos, I expected to see 1-2k higher hashrate from using Expo 6000 Aggressive settings, but for me it gets worse.

    Motherboard: ASRock 620I LIGHTNING WIFI
    CPU: AMD Ryzen 9 7900
    RAM: G.SKILL Flare X5 Series (AMD Expo) DDR5 RAM 32GB (2x16GB) 6000MT/s CL30-38-38-96 1.35V

    The BIOS has two settings I've been experimenting with:

    DRAM Profile Setting - has values JEDEC 4800 and EXPO 6000
    DRAM Performance Mode - has values AMD AGESA Default, Competitive, and Aggressive

    With ram at 4800 and AGESA default settings, I get a hashrate of 10k-13k. At Expo 6000 AGESA default I get 6k-11k. At Expo 6000 Aggressive I get 5k-10k.

    My board came with the 1.28 bios. After upgrading to last non-beta bios 2.02 I saw no change. I just upgraded again to the latest bios, 2.08.AS01 [Beta], and saw improvement for Expo 6000 Aggressive to 6k-11k. 4800mt AGESA performance stayed same at much better 10k-13k hashrate.

    Any recommendations for what to try next to use ram to full potential?
     
    • Angry
    Reactions: grstanford
    malor

    malor

    Ars Legatus Legionis
    16,093
  • #2
    On the prior generation, it was important to set FCLK to half the clock rate of your RAM. Nearly all systems would support overclocking from 3200MHz to 3600MHz DDR4, but you'd take a major speed hit doing so unless you also bumped your FCLK from 1600 to 1800.

    I don't know if that applies to the AM5 chipsets, but it's probably worth a try. I'm not sure how the MT/s figure translates to MHz, however, you'll have to figure that part out.
     
    • Like
    Reactions: continuum
    R

    Rekonn

    Seniorius Lurkius
    32
  • #5
    Looks like you were right! Found a program called "mbw" that run from command line
    unix.stackexchange.com

    How can I observe memory bandwidth?

    I have an embedded Linux ARM system that is showing significantly less throughput than expected on both Ethernet and USB. I suspect the memory may be contributing. Is there a way to observe the mem...
    unix.stackexchange.com unix.stackexchange.com

    At 4800 AGESA that test shows values like ~17k, vs 6000 Aggressive shows 21k.

    1709344720545.png
    1709344783516.png
     
    R

    Rekonn

    Seniorius Lurkius
    32
  • #8
    grstanford said:
    Why are we giving coin miners advice on how to work their machines properly?!
    Click to expand...
    I'm just a guy that decided to finally replace a gaming rig I built in 2011 to play starcraft 2 with a machine that would also be good at mining a crypto I really like. What's wrong with that? And even if I was some big mining operation, (I'd already have expertise and wouldn't need to ask for help on forums, but let's just go with the example), what would be wrong with helping that?
     
    • Like
    Reactions: pasorrijer
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #10
    Most tasks care about latency more than bandwidth.

    I have the timings in the image working on a 7800X3D with 2x 48 GB (96 total), and on a 7950X3D with 2x 32GB (64 total).

    Those are significantly better than defaults, and are fairly but not hugely aggressive (6400 was not entirely stable on the XMP kit). I would be sure to test it extensively if you do this. Some of those timings are usually OK, but are pretty far off of what it will ship with.
     

    Attachments

    • even more timings.png
      even more timings.png
      240.2 KB · Views: 5
    • Like
    Reactions: Rekonn
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #11
    grstanford said:
    Anyone who has tried to purchase a graphics card at a fair price for the last few years knows exactly what is wrong with it! Miners and scalpers are the scum of the earth.
    Click to expand...
    If he is very concerned about memory performance, would that not mean he is very likely mining on the CPU, and not contributing to the GPU shortage?
     
    • Like
    Reactions: pasorrijer and Rekonn
    G

    grstanford

    Ars Praefectus
    3,797
  • #13
    Rekonn said:
    Monero is mined using only CPU. I didn't even buy a graphics card for this build, just using integrated.
    Click to expand...
    OK in that case I'll apologize, however unless you are powering this thing off solar panels its still an overall negative to society in my opinion.

    Anyway, you are using an A series chipset. These are meant for office and granny PC's, not performance PC's. They didn't matter for GPU mining back in the day because the GPU did all the work, not the CPU. You'll want at least a B series motherboard and a decent one at that if you want to extract maximum memory performance on AMD.
     
    • Like
    Reactions: diabol1k
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #14
    There is the power aspect, but a modern AMD system uses about as much electricity as an old school light bulb even under full load. Being memory latency bound usually means not fully loading the processor.

    In theory the energy can be put to better use, but I am not going to throw that stone while lighting up the property all night, running a steam humidifier, and using what is likely a few hundred watts for the 5 air filters I have running (two of which are decently large commercial units.)

    I did replace an 11900K with a 7950X3D partially for power reasons, and that was a large drop (it uses about a third the power while running, and completes equivalent work in about a third the time). If you have long running tasks, that adds up.

    Anything I do on the 4090 is also a bit of a power draw, but for most tasks I have found better algorithms than brute force for my AI work by now, so it is mostly sort and search against memory on the CPU at this point (or running in the cloud on real AI hardware, using a model which is not mine).
     
    • Like
    Reactions: pasorrijer
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #15
    malor said:
    On the prior generation, it was important to set FCLK to half the clock rate of your RAM. Nearly all systems would support overclocking from 3200MHz to 3600MHz DDR4, but you'd take a major speed hit doing so unless you also bumped your FCLK from 1600 to 1800.

    I don't know if that applies to the AM5 chipsets, but it's probably worth a try. I'm not sure how the MT/s figure translates to MHz, however, you'll have to figure that part out.
    Click to expand...
    AMD says set the FCLK to auto, I think they added a buffer for AM5. It performs best when set as high as possible, but that is usually less meaningful than finding minimum memory timings.

    Personally I only bother overclocking memory, I do not even have PBO on as it is a small difference for relatively a lot of power, and undervolting is a difficult to assess stability risk.

    MT/s is double MHz for DDR memory. You probably mean 3200 or 3600 MT/s, not MHz (3200MHz is DDR 6400 memory, 3600MHz would be DDR 7200 memory).
     
    malor

    malor

    Ars Legatus Legionis
    16,093
  • #16
    cerberusTI said:
    MT/s is double MHz for DDR memory. You probably mean 3200 or 3600 MT/s, not MHz (3200MHz is DDR 6400 memory, 3600MHz would be DDR 7200 memory).
    Click to expand...
    It was sold to me as DDR4-3600 MHz, so that's how I think about it. On DDR4, FCLK needs to be half that number.

    They've been doubling signal rates on the same-speed RAM, more or less, for ages. The underlying clock speed of the actual cells was 200MHz back in the pre-DDR days, and it stayed near there for like twenty years. They just kept stacking it twice as wide with each new generation of DDR, and signaling it twice as fast, while the actual physical cells stayed right around the same wall-clock latency. They did climb in speed a little bit, but not that much.

    DDR5 apparently didn't just double the signaling rate again, but I think the underlying physical modules still aren't that much faster than 200MHz.
     
    Last edited:
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #17
    malor said:
    It was sold to me as DDR4-3600 MHz, so that's how I think about it. On DDR4, FCLK needs to be half that number.

    They've been doubling signal rates on the same-speed RAM, more or less, for ages. The underlying clock speed of the actual cells was 200MHz back in the pre-DDR days, and it stayed near there for like twenty years. They just kept stacking it twice as wide with each new generation of DDR, and signaling it twice as fast, while the actual physical cells stayed right around the same wall-clock latency. They did climb in speed a little bit, but not that much.

    DDR5 apparently didn't just double the signaling rate again, but I think the underlying physical modules still aren't that much faster than 200MHz.
    Click to expand...
    What has not improved much over time is latency. The sense amplifiers to read an active row are not much faster than they were 25 years ago.

    I have been buying 10ns memory for quite a while. The CL number keeps going up to compensate for the higher clock rate, but DDR 6000 CL 30 is the same wall time as the DDR 400 CL 2 I had back in like the year 2000.

    Effectively this means the delay between issuing the read and the first data byte is the same as it has always been.

    I think DDR4 was a little bit faster towards the end of its run, and DDR3 was substantially so (like half, I put some very low latency memory into a 4790K for a huge benefit on some tasks, which I want to say was like 6ns), but the lower clock means they do not transfer as much data in a given time, so later memory catches it pretty quickly during a burst and is almost always going to be better overall.

    DDR5 may not reach that point as it has on chip ECC (also something to watch out for if your measured speeds decrease when set more aggressively, it could be correcting errors, I am not really sure what that looks like in terms of timing) and some other features which will eat into this.
     
    • Like
    Reactions: malor and continuum
    C

    cerberusTI

    Ars Tribunus Angusticlavius
    6,449
    Subscriptor++
  • #18
    malor said:
    It was sold to me as DDR4-3600 MHz, so that's how I think about it. On DDR4, FCLK needs to be half that number.

    They've been doubling signal rates on the same-speed RAM, more or less, for ages. The underlying clock speed of the actual cells was 200MHz back in the pre-DDR days, and it stayed near there for like twenty years. They just kept stacking it twice as wide with each new generation of DDR, and signaling it twice as fast, while the actual physical cells stayed right around the same wall-clock latency. They did climb in speed a little bit, but not that much.

    DDR5 apparently didn't just double the signaling rate again, but I think the underlying physical modules still aren't that much faster than 200MHz.
    Click to expand...
    This is a pretty good overview of how various common memory works if you have a little bit of time.

    There is very little Verilog in that, it is mostly just a general summary with some circuit diagrams.
     
    G

    grstanford

    Ars Praefectus
    3,797
  • #19
    cerberusTI said:
    There is the power aspect, but a modern AMD system uses about as much electricity as an old school light bulb even under full load. Being memory latency bound usually means not fully loading the processor.

    In theory the energy can be put to better use, but I am not going to throw that stone while lighting up the property all night, running a steam humidifier, and using what is likely a few hundred watts for the 5 air filters I have running (two of which are decently large commercial units.)

    I did replace an 11900K with a 7950X3D partially for power reasons, and that was a large drop (it uses about a third the power while running, and completes equivalent work in about a third the time). If you have long running tasks, that adds up.

    Anything I do on the 4090 is also a bit of a power draw, but for most tasks I have found better algorithms than brute force for my AI work by now, so it is mostly sort and search against memory on the CPU at this point (or running in the cloud on real AI hardware, using a model which is not mine).
    Click to expand...
    It's a waste of power/electricity in the sense that it does nothing useful in exchange for the power/electricity consumed (hence my hope its powered by solar comment), your light globes, humidifier etc all serve useful purposes including your 11900K and 7950X3D (they have uses other than generating funny money).

    And yes, wasting power is a concern, since more demand on the power grid increases prices for everyone else on that grid and puts more pollution into the environment for no good reason.
     
    • Like
    Reactions: steelghost
    R

    Rekonn

    Seniorius Lurkius
    32
  • #20
    grstanford said:
    It's a waste of power/electricity in the sense that it does nothing useful in exchange for the power/electricity consumed (hence my hope its powered by solar comment), your light globes, humidifier etc all serve useful purposes including your 11900K and 7950X3D (they have uses other than generating funny money).

    And yes, wasting power is a concern, since more demand on the power grid increases prices for everyone else on that grid and puts more pollution into the environment for no good reason.
    Click to expand...
    Where would you put the US military on your scum of the earth scale? Lot of energy wasted, for example, over two decades to free Afghanistan from the Taliban, to give back to the Taliban. Biggest polluter too.
     
    R

    Rekonn

    Seniorius Lurkius
    32
  • #22
    grstanford said:
    I'm fine with the US Military given what the world would be like if we didn't have them as the worlds policemen. They benefit civilized society by keeping it civilized.
    Click to expand...
    Please, take a few minutes to see what US foreign policy really accomplishes, and who benefits. Vid is a couple years old, so doesn't have most recent Russia/Ukraine conflict, but still an excellent summary.

    View: https://www.youtube.com/watch?v=HbDIdPyfPjM
     
    • Angry
    Reactions: grstanford
    DaveB

    DaveB

    Ars Tribunus Angusticlavius
    7,274
  • #23
    Rekonn said:
    Where would you put the US military on your scum of the earth scale? Lot of energy wasted, for example, over two decades to free Afghanistan from the Taliban, to give back to the Taliban. Biggest polluter too.
    Click to expand...
    How did you see fit to insert fucking political bullshit into this discussion? Please stop polluting the forum with politics! Whether far left woke or far right MAGA keep your opinions out of here and go argue with all the left and right wing nuts on social media. Nobody really gives a damn what your politics are here.

    You came here for help on a technical issue, please stick to that. If someone's comments annoy you, highlight their avatar and click "Ignore" and you won't have to see their posts.
     
    Status
    Not open for further replies.
    Jump to latest Follow Reply
    Jump to latest Follow Reply