PC specs for lowest possible latency?

Discussion in 'PC/Mac Builders, Mods, Peripherals - New' started by chillbot, Jan 8, 2019.

  1. chillbot

    chillbot Forum Bot

    2,124
    3,976
    Feb 6, 2014
    La Canada Flintridge, CA
    I wanted to ask this in the "buffers" thread but it's already at 4 pages and didn't want to hijack it.

    Cost is not an issue. Assume a decent audio card and just focus on the PC. What would you assemble to give yourself the best chance of having super-low latency without pops or clicks.

    Just hypothetically. (And because I'm getting a new PC built.)
     
    whiskers likes this.
  2. zolhof

    zolhof Active Member

    380
    561
    Jun 27, 2014
    Silent Hill
    There are so many variables when it comes to low latency. Hardware-wise, for audio production, if you decide to go full balls to the walls i9 9900k (get a decent power supply and cooling solution), stay away from motherboards with PLX chips, as they add latency.

    The Gigabyte Z390 AORUS Xtreme is (imho) currently the best motherboard money can buy. If you won't be doing any extreme overclocking, you are safe with the AORUS Master, which is also very reliable and $200 cheaper. For Thunderbolt check the Gigabyte Z390 Designare variant.

    Also, skip Nvidia GPU cards to avoid driver headaches. I managed to tame my GTX 1070 but it wasn't fun. My next build will be Radeon just for a change.

    There's also the Threadripper 2990WX which costs 3x the price of the 9900k and, for raw power, can't be beaten. It didn't go too well in DAWbench though "Both chip ranges (1st and 2nd gens) are designed for certain tasks and optimized in certain ways, which ultimately makes them largely unsuitable for low latency audio work, no matter how much they exceed in other segments."

    A few good sources of information:

    http://www.scanproaudio.info/2018/0...bench-just-a-little-bit-of-history-repeating/
    http://www.scanproaudio.info/2018/10/19/intels-i9-9900k-and-the-coffee-lake-refresh/



    Plot twist:


    Hope that helps as a starting point!
     
    chillbot likes this.
  3. tack

    tack Damned Dirty Ape

    2,330
    1,924
    Aug 3, 2013
    For lowest latency I'd avoid the Threadrippers. I have a 2950X and like it, but I can't squeeze it on very low buffers, whereas the 8700k I came from could be pushed further. As far as the CPU goes, it really comes down to single core performance. I think the 9900k would make a fine choice with an all-core OC.
     
    whiskers and chillbot like this.
  4. OP
    OP
    chillbot

    chillbot Forum Bot

    2,124
    3,976
    Feb 6, 2014
    La Canada Flintridge, CA
    Very interesting, thanks.

    Cost aside would you consider 9940X @ 3.3GHz OR 9920X 3.5Ghz or still stick with the 9900k?
     
  5. tack

    tack Damned Dirty Ape

    2,330
    1,924
    Aug 3, 2013
    Given the primary goal to optimize for lowest possible latency definitely still the 9900k. You can get some fairly good overclocks on it.
     
    whiskers and chillbot like this.
  6. whiskers

    whiskers I should probably be studying :\

    1,419
    1,720
    Sep 5, 2018
    US
    will be interesting to see what's in store re: zen2 and Icy Lake too...
     
  7. Nathanael Iversen

    Nathanael Iversen Active Member

    333
    201
    Mar 25, 2013
    I have an i9-9900k, overclocked to 4.7Ghz. I run my ~600 track Cubase 9.5 template across two VEP slave machines at 128 samples. VEP uses an extra buffer, so 256 practically. Works pretty well for me. That said, when I have a full orchestra mock-up going, the real-time system is 90% in Cubase. I don't get drop outs, but I am maxing out what it can do. Insert effects seem to really chew up the available Real-time performance. The less I use, the lower it is. That is only so workable.

    I will upgrade to a 9940XE or one of the Zen2 processors later this year to just get more cores. The 9940XE will overclock to over 4Ghz, so it will be plenty fast enough. This MOBO+RAM will get turned into a 64GB VEP slave and have a long life in that capacity, so it will all work out.
     
    JohnG and chillbot like this.
  8. Nathanael Iversen

    Nathanael Iversen Active Member

    333
    201
    Mar 25, 2013
    To directly answer you question, "Specs for low latency":

    - I look for boards with Intel on board NIC's - they have better drivers & lower latency
    - I look for motherboards without WiFI (all I do is turn it off in the BIOS anyway)
    - I turn off all ASMedia controllers in the BIOS and only use the ports supported by the SouthBridge chip
    - I turn off all onboard sound in the BIOS, and any other hardware controllers that i can.
    - I would use motherboard graphics if I could, but I have too many monitors, so I use AMD Radeon. I get consistently lower latency without the big spikes I got with NVIDIA (WIN 10 here)
    - I use the Focusrite RedNet PCIe card as my interface. It goes into one of the 16x slots directly connected to the CPU. Rock solid 1.7ms latency for 128 channels of I/O. I gave up USB and won't ever go back. Audio goes straight to the CPU, not even through the SouthBridge. Some may claim the same for their Thunderbolt stuff, but I run close to 40 channels of I/O and it is just solid. As I used more and more of my old RME interface, it got to where only certain short USB cables would work. Lots of people have 8ch I/O and it works great, but that isn't my setup, and for the extra channels, I have been very happy with Dante and its practically limitless expansion. Super fast, just works, and I'm digital all the way to my monitors, so there's only one A/D in, and one A/D out in the whole system.
    - I have a Midas M32 console with a Dante card in it to provide low-latency monitoring for recording live in the studio. Cue mixes go to Rednet AM2 boxes scattered about the studio by the drums and the piano. The only latency is the 1.7ms AD and .83ms for the Midas. The MIDAS is an FPGA-based design and has fixed latency. There is a D/A at the AM2 boxes, of course. Total latency is certainly under 5ms for real-time recording in the studio. Professional drummers in the room have never complained.

    Getting low latency is more complicated if your channel count exceeds the little mixers built into 8ch I/O boxes. You either need an analog console, or to plan your digital I/O very carefully. If you want to spend a few 100k, you can get an SSL L300 console and Dante gear like the good folk at VSL - they have fixed latency under 2 or 3ms from any in to any out for recording an entire orchestra at 96k....

    You may already have all the IO stuff to satisfaction, so apologies if that veered too far off the CPU/MOBO/RAM dimension.
     
    JohnG and chillbot like this.
  9. OP
    OP
    chillbot

    chillbot Forum Bot

    2,124
    3,976
    Feb 6, 2014
    La Canada Flintridge, CA
    No I'm still working on the I/O but I've been told that cost-wise there's not much benefit in replacing my old MOTU 2408mk3s. I use 3 of them connected to a PCIe-424 card, I also won't touch USB. Do you have an opinion on them? I use 56 channels out (via ADAT) and 12 channels in (via ADAT and SPDIF) and mix everything externally on two Yamaha 02R96s.

    EDIT: I should add I also don't use VEPro in the traditional sense. I have 24 ADAT channels coming out of the slave computers but I use them like external synths. They receive MIDI from the DAW and output audio directly to the mixers, not back into the DAW.
     
  10. Nathanael Iversen

    Nathanael Iversen Active Member

    333
    201
    Mar 25, 2013
    Chillbot,
    I haven't used MOTU since the original 828 box - my first interface! So I don't have any opinion based on direct experience. Quick thoughts:

    1) If it is working, it is unlikely that conversion quality or such is holding you back. Nothing to gain there except extra cost. But it is an older rig... when do you upgrade... I get it. Not broken... but could it be better? Hard to justify on paper.
    2) Mixing out of the box keeps that load off your DAW. If you put it back into your DAW, presumably that is going to have some CPU "tax" implications.
    3) Hi-count I/O is a pretty limited game compared to the ubiquitous 8 channel IO with ADAT boxes. There is Dante, there is MOTU's AVB system. If you are on Mac, Apogee. Many vendors have ability to stack boxes, ie. Lynx and others. Live concert sound is transitioning from MADI to Dante. I went with Dante over AVB because of how much more it is deployed and the extensive 3rd party ecosystem. It seems that if you want/need to run 32+ channels of I/O, we are in the minority and pay "pro" prices to do so.
    4) The new Allen&Heath SQ series takes Dante cards and runs at 96Khz. My studio runs at 48Khz since that's what the samples do, but it is nicer than my M32. Dante itself is not limited, but sometimes I/O cards are limited to 32 channels at 48 or 96khz, so you do have to watch that. The PCIe card runs 128 ch at full data rate.
     
    chillbot likes this.
  11. Yay Level 1 Techs! Love those guys and gal of course ;)
     
  12. No thoughts or love for SoundGrid?
     
  13. Nathanael Iversen

    Nathanael Iversen Active Member

    333
    201
    Mar 25, 2013
    No love for Soundgrid. Totally proprietary, one vendor, no ecosystem. Why am I paying them to build low spec PCs and put their name and proprietary network protocol on them? Seems like exactly the sort of thing to be orphan technology in 5 years. Maybe they will get that ecosystem in place.

    Dante has it already, and the whole live sound community embraced it, with dozens of vendors producing products for it. I think Soundgrid is taking off for Waves, but I'm not that interested in being locked into one vendor. I am also not a plugin junkie. I am not of the opinion that some special EQ or something is going to transform my tracks. I am also not a sound-designer with exotic needs. For just working, I use the stuff built into Cubase. If I need color, I use Slate Digital's stuff - I have the subscription. I've never worked with the original hardware gear, so have no opinion about accuracy of emulation. This stuff sounds good to me, so I use it.
     
    JohnG likes this.
  14. HelixK

    HelixK Active Member

    261
    287
    Nov 10, 2018
    Are you hosting any instruments on the i9 machine? If so, how many Kontakt instruments plus insert effects you can load and play before getting drop outs? How much of an improvement you get from overclocking?

    I was hopeful for the Zen2 but I don't know what to think now... I had the 1950X and sold it because the real-time audio was abysmal compared to my 8700k. No improvement on the newer 2050X-2090, so what makes you think the Zen2 is going to be any better compared to the 9900K or 9940XE?
     
  15. JohnG

    JohnG Senior Member

    7,706
    4,163
    Nov 13, 2007
    California
    @Nathanael Iversen

    Thanks!

    I'd emulate a lot of what you describe, but I'm still running the now old-school system where sounds don't return to my DAW but go directly from slave PCs to a separate Pro Tools rig. Unfortunately that negates some -- perhaps a crucial "some" -- of what VE Pro can do to help reduce latency overall.

    Maybe the thing to do is rewire everything, but that would no doubt drive anyone to self-harm...
     
    Nathanael Iversen likes this.
  16. Nathanael Iversen

    Nathanael Iversen Active Member

    333
    201
    Mar 25, 2013
    I do host instruments, but mostly synths. I have no idea on total Kontakt instruments; not important to my application. I'm sure it performs similarly to any published test if I ran that test. The only thing that is important is, "Does my template work for the things I write". It does - but I almost max out the CPU doing whatever it is that I do. So I'll buy more CPU down the road. The i9-9900K is still a "value oriented" processor for what we do, so it was a bit of an experiment. It will make a great VEP sample server.

    I never ran the system without an overclock, so I can't possibly say. If the CPU can be overclocked, it is - that being the point of buying it. I purchase chips that can run all core over-clocks above 4.0Ghz.

    Zen2 has newer micro-architecture that seems to address the limitations of Threadripper, but I have no idea. It will be thoroughly tested within days of release by dozens of sites. I am not a "fanboy". I just need stuff that works.
     
    HelixK likes this.
  17. Nathanael Iversen

    Nathanael Iversen Active Member

    333
    201
    Mar 25, 2013
    Indeed. Tearing the studio apart to install the Dante and get everything routed was a big deal. The one thing that I've learned is that there are a lot of ways to do this. And local workflow needs drive most of the differences. I'd be loathe to criticize anyone with a working system. I know mine inside out, and it does what I need, but what does that mean to anyone else? That's harder to say for sure. If I integrate a Protools machine, I'd probably look to stick a Dante card in it, but that presumes all the work I've done, not a "fresh build". That's the other complication. This all looks different as a transition than from a "green field build". Those considerations are not insignificant.
     
    JohnG likes this.
  18. JohnG

    JohnG Senior Member

    7,706
    4,163
    Nov 13, 2007
    California
    Right again. But I really wish my latency weren't making things so mushy!
     
    Nathanael Iversen likes this.
  19. HelixK

    HelixK Active Member

    261
    287
    Nov 10, 2018
    What a coincidence, AMD just teased the Zen 2 at CES2019! They didn't show much but there was a very promising comparison between an early sample 3rd gen Ryzen and an i9-9900K. It's a single die-cpu and that's a great move because this is going to reduce the latency from the intercore communication between the L3 cache and the actual CPU core, giving the snappiest experience possible. And significantly less power consumption compared to the i9.

     
  20. Nick Batzdorf

    Nick Batzdorf Moderator

    14,852
    2,166
    Sep 14, 2004
    Los Angeles
    John, how much latency are you experiencing? Is it because you're monitoring through Pro Tools?
     

Share This Page