Yviena Posted October 17, 2019 Share Posted October 17, 2019 4 hours ago, Miska said: I think @Yviena had some tunings to make it work with 3700X. This may have to do with AMD's CPU microcode revision and "favored cores" on 3rd gen Ryzen where some cores are better than others, and boost possibilities when all cores vs some cores are loaded. Here's some story and testing about the topic (you can find more if you want to): https://www.tomshardware.com/news/amd-ryzen-3000-boost-fix-cores,40398.html I think it's because EC 7 DSD256 needs 4 cores to make it work without dropouts, and the 3600x/3900x is 3+3/3+3+3+3 CCX cores, making it hit the additional latency with cores on another CCX, while 3700x/3800x is 4+4 cores so the latency associated with cross communication is not hit when I manually assign the cores to only one CCX. There's been talks about next Ryzen gen 3 to move to a unified L3 cache so maybe that will mitigate the latency penalty Link to comment
Yviena Posted October 31, 2019 Share Posted October 31, 2019 New KS version has worse IPC than earlier version due to hardware security fixes though... Link to comment
Yviena Posted November 1, 2019 Share Posted November 1, 2019 EC 512 will continue to be impossible unless miska reaches a breakthrough in parallelization or CPU reaches 7-8ghz clocks which is unlikely as higher density nodes are actually shrinking in frequency, the way forward is IPC, and architecture/parallelization unless we reach some breakthrough in materials like grapheme/nano carbon tubes. Link to comment
Yviena Posted December 4, 2019 Share Posted December 4, 2019 In my opinion HQplayer version 4.1.0.1 still has the best CPU usage, and behavior when manually selecting cores for EC7 modulator with my 3700x, idk why but newer builds still do not behave the same way as 4.1.0.1 with corepinning disabled even if miska said it should disable the changes done to core allocations... Like for example i can disable core pinning in the newest build select 4 cores for HQplayer to be on via affinity cpu usage around 21% i get drop outs but if i leave corepinning on, but use process lasso to set the affinity to those same cpu cores on program start , cpu usage is the same but then i get no dropouts... Link to comment
Yviena Posted January 25, 2020 Share Posted January 25, 2020 8 hours ago, Miska said: Windows seems to be doing something strange, and annoyingly hard to figure out why it behaves so differently with the same code... It only becomes visible when the machine is pushed near to it's limits... Quite possible that you may need to do two different build configurations one for windows, other Linux/Mac. Link to comment
Yviena Posted January 25, 2020 Share Posted January 25, 2020 3 hours ago, Miska said: Like?? Windows needs some sword and sledgehammer. Meanwhile, macOS is behaving sensibly... Windows scheduler has always been kind of stupid though, it's not even truly SMT-aware,, also could just be that due to windows quirks what works great for Linux/Mac works worse for windows, it also doesn't help that they dont really have a QA team anymore for windows... We can see in benchmarks also that windows struggles with higher amount of threads compared to Linux. Link to comment
Yviena Posted January 9, 2021 Share Posted January 9, 2021 Hmm in my case with HQplayer 4.8.1 Multicore DSP ticked on, off, or grayed my 5900x has problems with lots of drop-outs with EC7 DSD256, seems i still need to disable corepinning, and manually assign threads/cores. Link to comment
Yviena Posted January 13, 2021 Share Posted January 13, 2021 1 hour ago, Miska said: No dropouts, but the two cores are loaded to about 96%, so pretty high without too much margin. This is HQPlayer Desktop on Ubuntu 20.04. One thing to note is that in order to make most of the motherboard hardware work, I needed to use the Canonical's "hardware enablement" kernel. So in other words install "linux-lowlatency-hwe-20.04" meta-package. This brings in their latest short-term kernel, which is currently version 5.8. Also note that you need very latest motherboard BIOS that you can flash to certain motherboard models from USB memory stick without CPU or RAM installed. Because the motherboards have been shipped from factory with a BIOS that doesn't understand about 5000-series CPUs. That already gives DSD256... ;) Is there actually any reason why Intel seems to have a lower load with EC7 256, benchmarks seem to indicate that IPC/performance is actually higher around 5%ish with the new 5xxx series. Link to comment
Yviena Posted February 3, 2021 Share Posted February 3, 2021 Just curious but what is the per core usage for you guys running 10850-11xxx series with EC7 DSD256. Link to comment
Yviena Posted February 4, 2021 Share Posted February 4, 2021 4 hours ago, Miska said: Interesting is that core 0 + core 5 SMT? The cpu utilization per core looks close to what I get when manually assigning 2 cores only to HQplayer coupled with multi DSP ticked on around 70-80% per thread. Link to comment
Yviena Posted May 22, 2021 Share Posted May 22, 2021 1 hour ago, luisma said: Can the 11600K do EC modulators with poli sinc 3? Would it be best to go with Ryzen 5600X vs 11600K ? I might go with a cheaper build now and later on Jan 2022 get the 12th gen chip ext-3 seems to be very resource intensive when going from 44.1khz to 48khz especially with PCM 7xx rates on my 5900x it basically has the same CPU load as EC7 DSD256, and this is with PCM to PCM upsampling only. Link to comment
Yviena Posted May 22, 2021 Share Posted May 22, 2021 44 minutes ago, Miska said: Doesn't look too bad to me: This is on Ryzen 7 5800X. I dunno reported cpu load atleast for me is the same with 44.1 to 768khz ext3 , and EC7 DSD256. Link to comment
Yviena Posted July 4, 2021 Share Posted July 4, 2021 1 hour ago, lpost said: Now that some more months have passed, do you recommend the 5800X and B550/X570 motherboard? The Realtek 2.5G NIC issue still appears to be unresolved and I would think is a hw issue at this point. It seems the AMD processors are more efficient for HQP work but the supporting motherboards aren't quite up to par, yet. What Realtek NIC issue? Link to comment
Yviena Posted July 4, 2021 Share Posted July 4, 2021 7 hours ago, Miska said: I would recommend to avoid Realtek NICs in general if possible. I don't find anything wrong with my realtek 2.5G i'm using the 10050_05182021 driver tho Link to comment
Yviena Posted January 13, 2022 Share Posted January 13, 2022 XMP timings are mostly super conservative, Link to comment
Yviena Posted January 13, 2022 Share Posted January 13, 2022 1 hour ago, Miska said: XMP is pretty fast, but the default timings when XMP is not enabled are usually slow. For example usually for all my DDR4-3200/3600 it is DDR4-2133 timings without XMP. Ehhh I was thinking more the sub-timings which xmp does not really cover Link to comment
Yviena Posted September 28, 2022 Share Posted September 28, 2022 1 hour ago, Miska said: We'll know once both have been benchmarked with HQPlayer. But if I'd need to pick up either one, I would take 7950X or maybe 7900X. Seems like AMD has made bigger upgrade to their processors. In Intel's traditional tick-tock system, 13th gen seems to be a tick, just a small refinement. While AMD rolled out Zen 4. Intel's 12th gen was clearly a tock. But also depends on your OS. AMD is especially strong on Linux. Is the reason AMD is good on Linux because of scheduler being better or because of other stuff? Link to comment
Recommended Posts
Create an account or sign in to comment
You need to be a member in order to leave a comment
Create an account
Sign up for a new account in our community. It's easy!
Register a new accountSign in
Already have an account? Sign in here.
Sign In Now