Menu
What's new
By:
Filters

I7 6700k Slave Machine Sample Streaming Benchmarks

TintoL said:
I am surprised the improvement is so small.
Click to expand...
Yeah - I've never been able to explain why the read speeds and IOPS don't make much difference. When I saw no difference as a function of SSD read speeds 3-4 years ago I thought maybe IOPS would be the key but as I've started using NVME drives (which have much higher IOPS) I still haven't seen much improvement.

And that's not just for sample streaming - I do a lot of time-lapse photography and wind up with thousands of RAW camera images that need to have preview images generated and cached to manipulate them. I've tried putting my cache drive on NVME but doing so didn't make the process any faster. Same kind of thing as streaming samples but much larger file sizes. Neither one seems to benefit from the NVME drive.

The NVME drive did show 15% improvement in load times, though, so that's something. In truth, load times aren't an issue for me because I run from a template.

rgames

EDIT: Also, you are correct that CPU is not the bottleneck - I watched it while doing the benchmarks and it was only an issue at the really high buffer settings (512+512) with LASS. Max with everything else was around 50% - 60%.
 
To me, something has always smelled off about Kontakt performance, especially on loading patches. When there's no apparent bottleneck (all cores have plenty of headroom, filesystem cache is warm and disks are basically idle, no network I/O involved), one has to wonder just where the problem is. If this were Linux, I'd be able to more conclusively blame it on shitty software. On Windows, I feel entirely helpless. All I can do is stare at perfmon plots.

Is it like that on the Mac too? I've always just answered the "why do Kontakt patches load so slowly when every resource on my system has plenty of headroom?" with "Because Windows."
 
tack said:
"why do Kontakt patches load so slowly when every resource on my system has plenty of headroom?"
Click to expand...

I think this might depend on what the instrument actually contains BESIDES samples... like instantiating all those effects, filters, modulators, IR samples, reserving memory for groups, zones, modulators, etc... Sure that should be fast, but I would assume this part of Kontakt might not have received any important updates ever since K4.2 when binary format was introduced, or perhaps even before...
 
EvilDragon said:
I think this might depend on what the instrument actually contains BESIDES samples... like instantiating all those effects, filters, modulators, IR samples, reserving memory for groups, zones, modulators, etc... Sure that should be fast
Click to expand...
Or at the very least all that should be compute intensive, and I should see that in the graphs in the form of at least one core flirting with the 100% line.

I'd be much happier with that. At least I'd know I could always throw more hardware at it if I want it faster. But I can't actually find any apparent bottleneck in any system resource to suggest beefier hardware should make any difference. Even in the case of a cache stall (and so the CPU is waiting on RAM) this should manifest as utilization -- unless Windows counters don't work like Linux in this regard. That's worth looking into -- it could be a bad assumption.

What I see is most of the time cores are below 50% utilization during a heavy project load, at least on some of the heavier instruments like Mural and Sable. With some of the lighter ones (say woodwinds), not only do they load faster (unsurprisingly) but the CPU graphs show much higher utilization (surprisingly) than the heavier instances with multiple patches.
 
I concur that once you hit 110-135 IOPS you've topped out.
NVMe has benefits like embedded slots, great OS device too.
2 x decent SSDs in RAID 0 using RST is the same.
I use the 105k Phison controllers now.
8 months of trouble free gigs.

Try throwing STEAM Folder on an M.2 + OS.
PCM files love NVMe.
 
@rgames Strikes me this slave of yours could do pretty well as a single box with the DAW, and VEPro/samples all on it. What do you reckon ? (Soecifically using that M'board, 64GB RAM and a decent i7) - also, have you by any chance measured the power draw when you are giving it a hammering ?
 
Last edited:
Elephant said:
@rgames Strikes me this slave of yours could do pretty well as a single box with the DAW, and VEPro/samples all on it. What do you reckon ? (Soecifically using that M'board, 64GB RAM and a decent i7) - also, have you by any chance measured the power draw when you are giving it a hammering ?
Click to expand...
You certainly could use it as a single machine if your voice counts never exceed its limits. In general, I do exceed those limits, so I need a slave. Actually I picked up a second one with an even cheaper motherboard (ASUS Z-170E, $105 with $20 rebate, so an $85 mobo!).

Max power draw in Prime95 is around 150 W but you'll never hit that for DAW use. Max I've seen while actually using it is around 80 W.

The i7 6700k on the ASUS Z-170E hits 4.6 GHz and performs the same as the Asrock in all respects. That system is under $650 after the rebate.

With the two slaves and using my laptop as master, the setup is running all of EW diamond along with a ton of VSL, Kontakt and other libraries at a buffer of 128+128. When I get back to my desktop master I bet it'll do 96+96. Without PLAY I bet I could run everything at the same buffer (or better) with only one slave.

I haven't seen anything else that performs anywhere close to those two slaves and the total cost is under $1500 (assuming you re-use a case, power supply and drives).

So, as ever, you don't need to spend a huge sum of money on computers for running VIs. You just need a couple of cheap slaves.

rgames
 
@rgames Thanks v much - v useful. As a matter of interest, what is the voice count for one of those boxes when used as a single integrated box before you have to go to a slave ? As a starter system, I could kick off with one single box, then add a slave if and when needed. BTW, what is the power draw on idle for one box ? Sofar, this is looking like a great system. Now I just have to find a case and physical build that is relatively quiet and relatively light.
 
The voice count when running sequencer and samples on the same machine is always a bit less - how much less will depend on the project. My standard template has a few hundred tracks, a few dozen plug-ins (mostly EQs), four IR reverbs and a 15 or so group tracks for stems and drops the voice count by 10% or so when running the samples from the same machine.

Re: idle power draw, it's about 50 W.

rgames
 
tack said:
That's quite good. My 6700k system draws 130W while idle. Although I do have a monstrous GTX 980 Ti. :)
Click to expand...
That's certainly part of it - I use a $10 fanless video card. Fanless ensures that it won't draw much power. You can also use the onboard GPU on the 6700k but I think you get better overclock with the onboard GPU disabled - less heat generation on the die.

Do you disable the power-saving features? I also have all the power-saving features enabled (e.g. all C states enabled, speedstep enabled, etc). So my idle CPU speed is something like 1.3 GHz.

rgames
 
rgames said:
Do you disable the power-saving features? I also have all the power-saving features enabled (e.g. all C states enabled, speedstep enabled, etc). So my idle CPU speed is something like 1.3 GHz.
Click to expand...
I don't disable them, no. I have a couple icons on my desktop that makes it easy to switch between high performance and balanced power profiles, so when I'm in my DAW I run in high performance, and otherwise with balanced which lets the CPU clock down while idle. No C states are disabled in BIOS.

In the balanced profile, I idle at 800MHz. In high performance, I'm clocked at 4.5GHz. The difference between these is about 6W.
 
rgames said:
The voice count when running sequencer and samples on the same machine is always a bit less - how much less will depend on the project. My standard template has a few hundred tracks, a few dozen plug-ins (mostly EQs), four IR reverbs and a 15 or so group tracks for stems and drops the voice count by 10% or so when running the samples from the same machine.

Re: idle power draw, it's about 50 W.

rgames
Click to expand...

Thanks ! Very encouraging.
 
rgames said:
Yeah - I've never been able to explain why the read speeds and IOPS don't make much difference. When I saw no difference as a function of SSD read speeds 3-4 years ago I thought maybe IOPS would be the key but as I've started using NVME drives (which have much higher IOPS) I still haven't seen much improvement.

And that's not just for sample streaming - I do a lot of time-lapse photography and wind up with thousands of RAW camera images that need to have preview images generated and cached to manipulate them. I've tried putting my cache drive on NVME but doing so didn't make the process any faster. Same kind of thing as streaming samples but much larger file sizes. Neither one seems to benefit from the NVME drive.

The NVME drive did show 15% improvement in load times, though, so that's something. In truth, load times aren't an issue for me because I run from a template.

rgames

EDIT: Also, you are correct that CPU is not the bottleneck - I watched it while doing the benchmarks and it was only an issue at the really high buffer settings (512+512) with LASS. Max with everything else was around 50% - 60%.
Click to expand...
rgames said:
That's certainly part of it - I use a $10 fanless video card. Fanless ensures that it won't draw much power. You can also use the onboard GPU on the 6700k but I think you get better overclock with the onboard GPU disabled - less heat generation on the die.

Do you disable the power-saving features? I also have all the power-saving features enabled (e.g. all C states enabled, speedstep enabled, etc). So my idle CPU speed is something like 1.3 GHz.

rgames
Click to expand...

So would you recommend getting a simple graphics card, even for pure audio work, simply to take some load off the CPU? E.G. Can I buy a $10 GPU that fill fit into my Mini ITX?
 
Phryq said:
would you recommend getting a simple graphics card, even for pure audio work, simply to take some load off the CPU?
Click to expand...

Hi Phryq,

I used to add a cheap graphics card in the olden days for exactly this reason but, today, CPUs are so insanely more powerful that I would think the voice count impact of having or not having that card is negligible on a PC slave. I have no data to support this hypothesis.

That's if you're using it for music; for games it's another story.

For the main DAW computer, a lot of people use multiple screens. Most mobos don't seem to support more than two without a separate graphics card.

Kind regards,

John
 
Ok, thanks. I'm only on 1 screen. I wonder if a GPU would at least help by taking care of GUIs etc?

Anyhow, I'm hoping not, because it would just be 1 more layer of complexity building the machine.
 
rgames said:
...as I've started using NVME drives I still haven't seen much improvement.
Click to expand...

So, it's been two years since this was posted. Are people still really finding this to be true? It seems so strange...

Just to make sure - we're talking about NVMe drives running on the PCIe bus - not M.2 SATA drives (which are no faster than a regular SSD).

These two pages helped me better understand the difference:

goo.gl/ttxVCF

goo.gl/BuZaUs

(* see below...)

I'm about to build a new PC for VEP6 usage. I really want to believe that NVMe drives will help, but I'd also be happy to save some money!

Thanks...


* The forum software wouldn't let me post a link because I'm new to the forum. To get around this, links are shortened and unclickable - copy and paste them into address bar...
 
Last edited:
shmimptone said:
Wow! Thank you!
Click to expand...

I just have to say wow again. That is a fantastic analysis of some very complex phenomena.

Regarding your last comment in the document:

"DFD streaming could use a closer look in terms of the performance benefits at large scale, across a project consisting of 100+ tracks, and at aggressively lower preload buffer sizes."

I think that question is going to become more important. Although SATA drives perform well enough now, that's probably because sample libraries were developed for usage with them, no? Now that faster drives exist, libraries will push the envelope again...

As for who's the best qualified to answer those questions, I'd love to help, but I don't have your technical acumen. I hope you pursue it! Anyone thinking about rebuilding their rig has to decide whether to go NVMe or not, so I'm guessing this topic will continue to come up.
 
You must log in or register to reply here.
Top Bottom