I have no idea what I’m doing with regard to overclocking. I only the bios settings to cause a kernel panic and I was able to cause the thing to not even post . I am now seeing 5.2Ghz on cores where I didn’t see that before tinkering last night. so, maybe progress? I also know how to turn off e-cores finally but that didn’t seem to make a difference. I’m not sure for since-MG(a) if there is any advantage to turning e-cores on / off. I do not run any eq or convolution. Just the conversion to DSD and ship it on NAA.
When you say offset of +1 or +2 that doesn’t translate to what I need to do in the BIOS. I’m researching this frantically and causing myself some confusion.
Memory profile is XMP II and I’m running that as fast a freq as I can. I can run it at a slightly lower freq and lower latency. I’ve not tried that.
I have no touched the HQP config yet. That was next on my list once I find that documentation. I’m running embedded from the bootable image so it’s HQPlayer embedded kernel.
I forgot I could do this. OK, I’m going to try and push this onto the e-cores as I’m trying to move 16/44.1 to DSD256. Maybe that will help. e-cores are currently just sitting there laughing at me.
It has been asked every now and then. Technically possible, but interpreting that into effects throughout the audio band is known but rather complex topic. So I’m not sure how useful that is compared to the current table? Certainly interesting in a way, but not so straightforward to interpret in another way.
I see. Basically there are few things to try before reaching out for your wallet. 1) BIOS 2) HQP config.
btw, “nblocks” I’m guessing only make sense for GPU systems
offset +1 or +2 → this is when we increase the p-core ratio by +1 or +2 notches. there are few ways to do it… do it for all p-cores, do it selective per p-core, do it for “groups”.
sinc-M is always fixed length in number of taps, regardless of conversion ratio. sinc-Mx is always fixed length in time, regardless of conversion ratio. Meaning that sinc-M gets less steep as ratio is increased. While sinc-Mx keeps the same steepness regardless of conversion ratio.
sinc-MG(a) and sinc-L(l/m/s) work like sinc-Mx, that they scale depending on conversion ratio, so that the frequency response is always the same (meaning more taps are needed higher the conversion ratio).
What do you have as E-cores offload?
If you like above filters, using a powerful Nvidia GPU can make a huge difference.
This was a bug fix in a recent release that corrected the length calculation. Not that it would have been bad before, now it just works as it should…
With these filters, CUDA offload to a suitable Nvidia GPU is best option. Because these are precisely the cases where the Nvidia GPUs with thousands of CUDA cores excel. For example the RTX A4500 I’m using on my i9-14900K has 7168 CUDA cores and still very reasonable 200W maximum power consumption. RTX 4090 is a beast, with 16384 CUDA cores, but consumes up to 450W of power.
Please note though that number of CUDA cores alone doesn’t mean much if you compare between GPU generations, because each generation tends to almost double the performance of single CUDA core.
Advantage of the professional series of GPUs is that you get more RAM (and possibly ECC).
@IgorSki
I found the documentation I needed. But I do not have time to play at this level for a while. I’ll paste it here for others to find:
I found where to set voltage offset… but my BIOS lets me do it per clock frequency? Not sure how to maths this into what HQPlayer is doing / needs.
I just tested moving this from Default (where they are not used) to filter (where they max out at 100% @ 4Ghz and we still stutter). So… doesn’t seem to “solve” at least not for -MGa
A GPU might very well be the right solution here. The GPU will cost more than the machine and that’s not exactly an impulse purchase.
I only listened for about a minute but it didn’t stutter. I might be right on the edge though as I do see my e-cores jump to 100%. My p-cores, although not well distributed, stay below 100%. Here are my configs:
BIOS tweaks:
P-Cores can go to 5.2Ghz
E-Cores can go to 4.0Ghz
RAM is optimized for overclock XMP II
HQPlayer Embedded:
Short buffer: Short
Multicore DSP: On
E-cores: DSP Pool
Blocks per cycle: 11
Modulator: ASDM7EC-super
Oversampling: sinc-MGa
Rate: 12288000
since we’re sharing:
It bounces between e-cores close to 100% @ 4Ghz
Yes, this is the good one. But certainly a lot of information for a “quicky”. You may also check particular CPU overclocking examples from the same source. Notmally those articles are less generic and more to the point. But regardless, with Linux Server “nitty gritty” overclocking is rather time consuming trial and error anyway.
If you are on ASUS motherboard, and referring to this documentation, the “per core voltage” is a rather complex matter and can “brick” your gear if set too high or too low - normally BIOS will warn you with red flags when you go off boundries… The “per core ratio limits” is an easier strategy.
You may check this one out, as you may have these settings available on your CPU. “Intel turbo ratios” strategy works well with HQPlayer (at least in my OC trials)
Appreciate the link! When I am ready for frustration 'brick, reboot, try again" I will dive in. I suspect I will need to claw out more performance once I have a longer listening session. I don’t think this is solved yet. But we are worlds better than where I was last night.
BTW it is still rather potent set-up. Not bad at all. But, indeed, if I would go further, I’d consider a GPU for this one and that’s gonna be an investment. Jussy has just posted above a very nice explanation on GPU’s…
You will have more slack for variations using full buffer unless the delay annoys you too much.
These two will like have a pretty big difference. Many times I get nice performance with E-cores: filter and Blocks set to default, or something like 4 or 6. But the most optimal settings will always depend on the hardware and the particular settings.
This is in category “feedback to Intel”…
Possibly if you can find fast enough AMD, it will be better in this respect. But that will be hard since AMD is overall still behind in single core performance. So you’d be looking for “AMD core with Intel speeds”.
My 14900K (non-overclocked) still just chokes trying to do ASDM7EC-super 512+fs to DSD1024. The regular ASDM7EC-super works without issues though. Even though the 512+fs is few instructions less to run per sample…
I’m not sure if it works for you due to overclocking, or just plain KS model is enough?
. I am now seeing 5.2Ghz on cores where I didn’t see that before tinkering last night. so, maybe progress? I also know how to turn off e-cores finally but that didn’t seem to make a difference. I’m not sure for since-MG(a) if there is any advantage to turning e-cores on / off. I do not run any eq or convolution. Just the conversion to DSD and ship it on NAA.
