What's new

NVMe vs SATA: Will it make Kontakt faster?

tack

Damned Dirty Ape
On the subject of the benefit of NVMe over a modest SATA SSD, specifically as it relates to Kontakt, I've commented here and elsewhere that:

At least with Kontakt's compressed samples, NVMe is completely wasted. Decent SATA SSDs are too for that matter: I bottleneck my CPU decompressing the samples as they're being loaded in long before I bottleneck my storage.

Having now investigated this a bit more, I need to moderate my position. The truth is it can help, depending on what you're measuring.

There are of course a lot of aspects to Kontakt performance (another big one being DFD streaming with various preload buffers), but I wanted to answer the one I personally found most annoying: can an NVMe make my project load times faster?

So I've done a bit of benchmarking and technical analysis (you know, for fun) trying to suss out what's going on (as best as I can given Kontakt's a black box).

https://docs.google.com/document/d/1wL8XYGgd_O9fomMrK1EpSnZJeQwhVOAn91e82byj8s4/edit?usp=sharing


The doc is a bit eye-glazing so I'll just cut to the chase:
  • Where "project load" is measured as "I can start using my DAW now" the answer is no.
  • Where "project load" is measured as "all samples are fully loaded into memory and the system is idle" the answer is yes. Unless you run everything purged. :)


Note that this is all quite specific to Kontakt. The ability to fully take advantage of system resources depends quite a lot on software design. So these findings aren't going to be transportable to other VIs (e.g. UVI, Omnisphere, etc).

I've posed some questions in the doc that I'd be keen to hear ideas on. Notably the fact that disk utilization % doesn't seem to map predictably onto actual disk usage metrics (throughput and IOPS) relative to the benchmarked capacity at equivalent block sizes. I proposed a possible explanation as to why that might be (the section titled "Kontakt's File Access Pathology") but I'm a bit skeptical that's the real explanation.


Cheers!
 
hey @tack -

interesting, thanks for sharing. some time ago I've compared loading times on a mac pro in between an EVO 840 and a RAID0 of 4 EVO. in the end, I didn't see any difference that would justify to create a RAID in order to make a kontakt instrument load faster - when you have an SSD already. from what I understand that was different for you? which version of kontakt did you test with?
 
It'll be interesting to see how Spitfire's new sampler performs/compares. We might be using that more than Kontakt in the future!
 
in the end, I didn't see any difference that would justify to create a RAID in order to make a kontakt instrument load faster - when you have an SSD already. from what I understand that was different for you?
Not really that different for me as I run everything purged. In this case all my painful wait times are entirely CPU bound and NVMe was no help. But when it comes to loading those samples into memory, NVMe did much better. (Though the data suggests Kontakt could do better on slower SSDs if it made a bit more effort.)

It'll be interesting to see how Spitfire's new sampler performs/compares. We might be using that more than Kontakt in the future!
Yeah, I'm interested as well. I'd be even more interested if they were going to open it up to third parties a la Kontakt. And I'd be ecstatic if Spitfire decided that building a sampler isn't their core business and released it as open source.
 
I wonder though if opening it up to third parties is a synonyme of the platform being hacked and cracked copies appearing?
 
I wonder though if opening it up to third parties is a synonyme of the platform being hacked and cracked copies appearing?
If the security of their copyright protection hinges on the secrecy of the software's source code then it is already doomed to fail.

It's analogous to developing a cryptosystem (such as TLS that secures most of the Internet): if the security of the encryption depends on the secrecy of the algorithm, as opposed to the key material being fed into the algorithm, then it's intrinsically insecure.
 
Nice work - good technical depth and test methodology.

I was obsessed with maximizing sample streaming and minimizing load times when I built my last slave and then the Cubase DAW. I first RAID'ed everything, then tried allocating libraries across JBOD based on likely use & streaming density. It was a PIA and when I went back to a Mac with my current iMac Pro, I just took the 4 850 EVO's and organized everything by type of library so I could actually remember where shit is - Strings, Winds&Perc, Brass&Hybrid, Ensembles&Keys. Guess what? - I have never noticed any significant difference regardless of what I did.

I did put the Spectrasonics STEAM folder and my Garritan CFX lib on the internal NVMe SSD in the iMP. The big Keyscape patches load a little quicker and the CFX is a LOT faster - the only library that really made a difference.

I have come to the conclusion that the big cost difference for NVMe isn't remotely worth it unless you make music by a stopwatch.
 
I noticed the Phison E8 Controller specs/benches like most M.2s but loads Keyscape faster than my Samsung’s that bench higher.
But that’s the Magician software so no telling if that’s from the company they bought that uses caching or just faster transaction times.
I’m interested in such details because I use Dual Live Mode during performances.
Until I got the Phison based M.2 I was limited to loading one Multi per song.
Now I’ve got 2 and 4 way splits instantly.
The new CC # sure helps too.

Anyone looking for new NVMe M.2s should check out the Phison Devices.
MyDigitalSSD is the cheapest, others are coming.
But this is only in reference for STEAM and one shot samples from my Host Bidule.
Every extra second counts.

Omnisphere really gave us a fantastic update.
CCs and resizing were my only 2 complaints.
Both were addressed, and then some.

Nice review on Kontakt and Storage.
 
ok but technically for the streaming, not the loading, it would help, right? AFAIK, it takes multiple processes to start really maxing out the bandwidth of those NVMe drives.. for example, just streaming one patch from kontakt would never get close to using the possible speed. But what about if kontakt was set up to use little ram and stream as much as possible.. if there were like, 16 kontakt instances, with 16 patches each, a huge orchestration, all streaming off the one drive.. wouldn't NVME *then* make a difference? Or am i in la la land and sata would still have headroom? I am trying to work out for myself now, as i am finally getting an imac pro, whether to bother investing in something like an OWC thunderblade as an external TBolt3 drive, which is roughly 2-2.5gb/s transfer speeds..or whether to just get a couple of cheap samsung 2Tb T5's.. I will not write to the drives often, these will be purely for streaming samples so endurance should be fine...
 
ok but technically for the streaming, not the loading, it would help, right? AFAIK, it takes multiple processes to start really maxing out the bandwidth of those NVMe drives.. for example, just streaming one patch from kontakt would never get close to using the possible speed. But what about if kontakt was set up to use little ram and stream as much as possible.. if there were like, 16 kontakt instances, with 16 patches each, a huge orchestration, all streaming off the one drive.. wouldn't NVME *then* make a difference? Or am i in la la land and sata would still have headroom? I am trying to work out for myself now, as i am finally getting an imac pro, whether to bother investing in something like an OWC thunderblade as an external TBolt3 drive, which is roughly 2-2.5gb/s transfer speeds..or whether to just get a couple of cheap samsung 2Tb T5's.. I will not write to the drives often, these will be purely for streaming samples so endurance should be fine...

This was always my assumptiom... I run Kontakt with a very low preload size and sure seems like I get better polyphony and performance than SATA III... Could be wishful thinking I suppose...
 
There is a huge sale on samsung T3 like I never seen at an aussie store.. he has already sold over 600..

the 2TB for 825 AU, minus 20%, so minus 165$.. therefore $660.. I can't even see them that cheap in the US.. that's about 450 USD for the 2TB.
I think i will get 2, simply connect one each to usb 3 ports on my caldigit thunderbolt 3 hub, and raid 0 them...
T3 sustains 450 read even over USB 3.. that should definitely take it to 700.. Believe it or not no one else has done this on video, anywhere. all people have done is raid some flash drives for fun, like a bunch of 8GB ones.

I think 4 terabytes of bus powered SSD for 1300$ AUD is dead cheap, and 4 terabytes is enough to hold every kontakt library i possess and then some.. So.. yeah.. Surely that will be fast enough..the write won't increase all that much with the soft raid but that is not relevant here.. even in all the flash drive tests, read speeds doubled in every case...

I really think i could get even 800 mb/s out of T3's!

Like that "Glyph" 2TB portable raid ssd which gets about 700 write and 800 read... but the 2TB is more in AUD than what i'll be paying for 4TB with the T3's.
Anyone with T3's and are happy with them, please let me know! Sale ends 15th, as in midnight tonight, 12 hours exactly!
 
This was always my assumptiom... I run Kontakt with a very low preload size and sure seems like I get better polyphony and performance than SATA III... Could be wishful thinking I suppose...

Sounds like transaction benefits of NVMe help out.
Prior to M.2 there were OCZ, HGST and Intel PCI-e Devices, and while they weren’t benching like NVMe speeds were seeing now, folks I know that used these said they had significant performance boosts.
Some of these devices were RAID 0 too, but transaction times could be an additional way to get higher polyphony, along with the CPU.
I noticed super low latency, which I dont really need, from above 4GHz CPU Speeds.
.07 msec. @ 96k/ 32 Samples.
I run my rigs @ 44.1k/64 samples and 4 msec.
Fine, considering 6 msec. is about the time a drummers ride cymbal reaches my ear.

Good that NVMe helps some of us.
I like it ever since Asus PS10S WS C236/Xeon 1U.
The NVMe M.2 sits right in front of my triple barrel 22k rpm fans.

B68231CE-8611-4AEC-A1FA-DE7CC278292A.jpeg
 
There is a huge sale on samsung T3 like I never seen at an aussie store.. he has already sold over 600..

the 2TB for 825 AU, minus 20%, so minus 165$.. therefore $660.. I can't even see them that cheap in the US.. that's about 450 USD for the 2TB.
I think i will get 2, simply connect one each to usb 3 ports on my caldigit thunderbolt 3 hub, and raid 0 them...
T3 sustains 450 read even over USB 3.. that should definitely take it to 700.. Believe it or not no one else has done this on video, anywhere. all people have done is raid some flash drives for fun, like a bunch of 8GB ones.

I think 4 terabytes of bus powered SSD for 1300$ AUD is dead cheap, and 4 terabytes is enough to hold every kontakt library i possess and then some.. So.. yeah.. Surely that will be fast enough..the write won't increase all that much with the soft raid but that is not relevant here.. even in all the flash drive tests, read speeds doubled in every case...

I really think i could get even 800 mb/s out of T3's!

Like that "Glyph" 2TB portable raid ssd which gets about 700 write and 800 read... but the 2TB is more in AUD than what i'll be paying for 4TB with the T3's.
Anyone with T3's and are happy with them, please let me know! Sale ends 15th, as in midnight tonight, 12 hours exactly!

Unless they're in the same enclosure I could see RAID 0-ing them being problematic. (If you one came lose (or you have a cat :P) seems like it'd be a little too easy to break the striping... Not a RAID ninja though so curious what others say...)
 
Thanks ED, hence why I asked.. However perhaps something like a 2950X overclocked to 4GHZ would be able to do it and make use of the speeds.. I plan to build such a system down the road so will find out then.. but that's 6 months away as I now ordered an imac pro.
 
Top Bottom