Overclock.net banner

41 - 60 of 66 Posts

·
Registered
Joined
·
585 Posts
Agreed QLC is garbage, like the old seagate shingled recording drives kinda garbage. I wont even purchase them, for anything. There are plenty of good TLC drives out there with proven endurance and decent speeds at fairly similar prices. Even when TLC came out i was skeptical, but its track record hasn't been too bad. I'd take a TLC SATA SSD over even an NVME QLC SSD.
 

·
Registered
Joined
·
3,035 Posts
After being stuck with two QLC drives at work, because I bought them because they were insanely cheap compared to the older ones and me asking the rhetorical question, "Well, how bad can they be really?"... yeah, I'd take normal HDDs over QLC SSDs. Seriously. I can get sustained transfer rates of ~190MB/s on an AMD SATA controller (historically not quite as fast as Intel's) over hundreds of GB of data on a Seagate Ironwolf or Toshiba M07, while the QLC drive starts off nice and fast (roughly 500MB/s) but after about 100GB drops to <100MB/s speeds. Yes, double digit MB/s. Off an SSD. I have a USB memory stick about the same speed as that!

I had a bunch of the old Seagate SMR drives too. They looked great at the start but quickly I was left wondering, "Why did I buy these?".

...

I see that Turtle Rig posted again, but still hasn't provided links, despite the requests.
 

·
Vermin Supreme 2020
Joined
·
25,774 Posts
speaking of cheap nand... predictors nand @ +40% of current costs by Q4. Stories read just like every time this happens "one minute power outage = Y'ALL GONNA PAY!"
 

·
Registered
Joined
·
3 Posts
Discussion Starter #44 (Edited)
Thanks so much everyone for your replies, I should have also included more information about me, I'm currently finishing my studies of composition at the conservatory and most of the things I've worked on were handwritten or in Sibelius, so although I've been using (or better said testing) different DAWs (currently using Reaper) for the past four years, I couldn't input all the time I wanted towards them to learn how to use well. Besides this, last year I've tried doing some "bigger" projects and with my current PC (I5-4440 and 12GB of RAM) it was really slow, for some works with tape I ended having to bounce too many times to process the audio, and while trying to mock some of the works I've done that would require a small template (around 8 instruments) it used nearly all the RAM I have (and the load times were very slow, but that's because I still don't have SSD). Until I finish my studies, something less powerful would suffice my needs, but I want to build something that can be durable for the next years.

About the DAE Interface, it's something that I haven't put too much thought at the moment and I need to inform myself better, I'm not entirely sure about, but for the moment as I don't intend recording and working from samples, I thought about going for something that's not expensive and is recommended as Focusrite Scarlett Solo 3rd Gen or something similar.

I didn't know about QLC, and from time to time there are deals for Sabrent Rocket 1TB and 2TB, so that's nice! Maybe I could go for a cheaper SATA and afterward add those when I find a good deal.

I don't think I'll overclock it, and if I'm not mistaken, the x3900 benefits more of having a good cooling rather than overclocking. About the motherboards, if, in the end, I don't pick the Elite with the promotion, I think I'll go with Tomahawk Max at 114€, the extra features that have Pro Carbon (as having both m.2 slots) can be worth for the extra 30€ it cost?

And I have a few doubts regarding the RAM, if I go with a b450 motherboard, might I have problems trying to achieve 3600/3733 mhz?

And thanks, oreonutz and HybridCore, for the suggestion of case fan, it's something I didn't think about, I'll read more about this and decide within all of the ones you commented.
 

·
CS+EE Alumnus
Joined
·
10,321 Posts
Unless you're going for extreme speeds or timings, RAM is almost never limited by the motherboard. RAM is directly connected to the CPU and doesn't pass through the chipset. Thus how your RAM clocks is almost purely a matter of how strong your chip's IMC (integrated memory controller) is.

I'd read the short explanation of the Ryzen IMC linked in my sig.

Anyways, stay with the Micron Rev.E.

If you need wi-fi or Bluetooth, the pro carbon ac is a great deal at 30 euros more. For reference, most Intel 9260 + m.2 to pcie x1 + antenna kits are usually 35+ euros.

You also change from Realtek to Intel LAN if that matters to you. I don't really hear any complaints about Realtek drivers though so that shouldn't really be a decision point.
 

·
CS+EE Alumnus
Joined
·
10,321 Posts
The 2nd M.2 slot on the B450 Gaming Pro Carbon AC is limited to PCIe 2.0 x4 (~PCIe 3.0 x2 equivalent bandwidth) so you still need to go for the 2TB drive for best performance.

https://www.msi.com/Motherboard/B450-GAMING-PRO-CARBON-AC/Specification

I'm not sure what non-Threadripper boards, if any, can provide more than one M.2 slots capable of PCIe 3.0 x4 bandwidth.
 

·
Tech Enthusiast
Joined
·
1,340 Posts
The 2nd M.2 slot on the B450 Gaming Pro Carbon AC is limited to PCIe 2.0 x4 (~PCIe 3.0 x2 equivalent bandwidth) so you still need to go for the 2TB drive for best performance.

https://www.msi.com/Motherboard/B450-GAMING-PRO-CARBON-AC/Specification

I'm not sure what non-Threadripper boards, if any, can provide more than one M.2 slots capable of PCIe 3.0 x4 bandwidth.
The x470 Crosshair VII Hero is the only board I am aware of that gives you 2 M.2 Slots, that gives you full PCIE3.0 x4 Bandwidth to both M.2's at the same time, and will even allow you to RAID 0 them if you want, although I wouldn't recommend doing that unless you do nightly backups and can be totally fine loosing data if your drive fails before you had a chance to back something up.

But the Board would work great in a DAW Situation (I Know because several of my Builds Use it) and the only compromise you are really making is not having PCIE4, which as of yet isn't that big of deal. You can find deals on the board from time to time if you wait. Depends on how urgent it is, because when its not on sale they are still asking too much in my opinion, but when it does go on sale, its a complete bargain under $200.

As @HybridCore says, you won't have any issue, no matter the board you pick, getting your Ram up to 3600Mhz at the very least, as long as you are going with a 3000 Series Processor. After 3600Mhz its dependent on your Individual Chip as to how good the "Unified" Memory Controller is. (It doesn't matter much for this conversation, but on Zen2 Chips its a UMC, instead of IMC, because the Memory Controller is located in the IO Die instead of with the Cores, so they changed the name for it, but its otherwise behaves the same way as an IMC, and therefore will be used interchangeably by everyone, which is understandable)

Some Chips will hit 3800Mhz with ease, others will only allow you to hit 3733Mhz, some only 3600Mhz, but regardless 3600Mhz is the sweet spot and that is all but guaranteed! (Assuming the RAM can do it)

The Focusrite Solo that you had mentioned is a great interface! The old MBoxes that Digidesign used to make used Focusrite Preamps in them, and I believe its what set Focusrite on a path of making solid small Interface for home users for years to come. They have made some of my favorite products for home use. I infact still am using a Focusrite Scarlett 18i20 in my DAW, and It does everything I need. The Solo is a good choice especially if you only need one Preamp, and just need a good DAC for your Monitors and Headphones, it won't treat you badly at all.

If you wanted to start with a SATA SSD I would Highly recommend going with at the very least a Samsung 860 Evo, get the biggest capacity you can, right now you can get 1TB for under $150 which is a steal, and seeming as you are still using Mechanical Hard Drives (Or as I and many others like to refer to them, spinning rust) you will notice a HUGE Jump to the SSD, and if you grab something like the 860 Evo, they will last you, even with heavy use, the better part of a decade (Obviously I can't predict that with 100 Percent Certainty, but I have only had a total of 1 Fail on me in the 5 Years we have been using Samsung Drives Exclusively, and that is 1 out of the close to 400 Drives deployed in a 5 year span, so they are pretty reliable in my book). You would want to move yourself on up to an NVMe drive at some point, especially when moving Data around, but a good SATA SSD will get you off of Mechanical HD's and into a better spot, and ready for the world of NVMe's. Which to be honest the price of some of those TLC drives are getting pretty low too, so if you find one for a good price, might as well jump on it, but don't feel bad about going with something like an 860 Evo either.

Anyways, I think I have addressed everything. You got this man! Let us know if you need anything. I envy your task ahead! (I haven't worked in Sibelius since 2004, I hadn't realized it was even still a thing, thats awesome its still around!)
 

·
Registered
Joined
·
3 Posts
Discussion Starter #48
Ok, I'll go with Tomahawk Max then, I don't think I'll need the Gaming Pro Carbon AC features and, unfortunately, the Crosshair VII Hero costs over 300€ here and the cheapest it has been was 250€, so it would be too expensive, and I'm glad to hear that the Focusrite Solo is great!

About the SATA SSD, while waiting for a good NVMe deal, I'll go with 860 EVO or MX500, depending on the difference of price at the moment I'll purchase, and as the rig is not very urgent, I'll be buying each component when I find they're at a good price.

And yes, Sibelius and Finale are both still used! In fact, most people I know work with Sibelius except a few that use Finale.

This is how it looks in the end:
https://es.pcpartpicker.com/list/9tK38M
 

·
Registered
Joined
·
585 Posts
Looks like you got yourself a solid setup there. I'd grab a bigger HDD as the prices are cheap on em, but cant knock the good old 1tb blue. You should be fine with any ram settings above 2666 to be honest. Lower latency will help a decent amount, staying in the 16-17's does pretty well over say 3200 out in the 20's. Atleast for this type of usage. You hit the main nail on the head by going 32gb.

The Focusrite Solo will serve you well enough when you need to get in and out of the box from what it sounds like. Reaper... love it, Justin did a fine job on that piece of software! Sounds like you may have found your home.
 

·
Registered
Joined
·
313 Posts
After being stuck with two QLC drives at work, because I bought them because they were insanely cheap compared to the older ones and me asking the rhetorical question, "Well, how bad can they be really?"... yeah, I'd take normal HDDs over QLC SSDs. Seriously. I can get sustained transfer rates of ~190MB/s on an AMD SATA controller (historically not quite as fast as Intel's) over hundreds of GB of data on a Seagate Ironwolf or Toshiba M07, while the QLC drive starts off nice and fast (roughly 500MB/s) but after about 100GB drops to <100MB/s speeds. Yes, double digit MB/s. Off an SSD. I have a USB memory stick about the same speed as that!

I had a bunch of the old Seagate SMR drives too. They looked great at the start but quickly I was left wondering, "Why did I buy these?".

...

I see that Turtle Rig posted again, but still hasn't provided links, despite the requests.
How can one benchmark that?
 

·
Tech Enthusiast
Joined
·
1,340 Posts
How can one benchmark that?
So, with QLC Nand, the way to tell the ACTUAL performance of the QLC Flash itself, and not the SLC Cache they put on it to speed it up, is to simply do a long transfer, typically 50GB or more will find it, but I will usually use a tool like "HD Tune Pro". With that program if you have a fresh disk that has not been written to, then I recommend the "Benchmark" tab, set it to "write" (Note, this will Erase any data on the Drive, hence while I stated to use this only when the drive is fresh, if it isn't then the next test I will discuss is better). After setting it to Write, keep "Transfer Rate", "Access Time", and "Burst Rate" all Checked, while leaving "Short Stroke" unchecked, and hit "Start".

This will write to the ENTIRE drive, so if its a 1TB drive, then it Writes about 1TB in data, and then logs the Performance Data on a graph for the entire run, while giving you real time stats on the right. Using this test, with a higher Quality SLC, MLC, and sometimes even TLC, Drive, you will see a consistent speed throughout the entire drive, only tapering off as the drive fills up, which is natural for any flash, but usually this is only about a 10% to 15% drop in speed and is gradual. The only other reason you would see a drop on a higher quality drive is due to Thermal Throttling, which this test WILL cause on most drives if you have no active cooling on the Controller of the SSD. (Flash likes to be hot, so no need to cool the flash, but the Controller does not like going above a certain temperature, this is what I connect thermal pads to on a Heat sink, that I actively cool with a fan for this test.)

Even with Thermal Throttling, you will see a typically still see speeds in access of 800Mbps or more, depending on the drive it could be a lot more while the controller was Thermal Throttling.

With QLC Nand on the Otherhand, for the beginning of the test you will typically see excellent speeds on par with mid range TLC Drives, and then as your test gets into about 50GB to sometimes 100GB depending on the drive, you will see the Throughput drop off to typically 150Mbps or even less. Thats how you know you got a QLC Drive, if you didn't want to just look this up via google before hand.

The other way to test with The same Program, is do a "File Benchmark". This you don't need to actually Clear the drive, it will not erase anything, and instead just write a file. You dictate the Size and the Pattern. For a normal Sequential Workload you would Choose "Zero", which will simulate transferring a Huge Video File. I would choose about 150GB, just so you get to see the performance over a long enough period of time, and then hit start, and then same thing as before, as the test goes on you will see how bad performance gets.

Whether or not you run into this in the real world depends on the workload you run, but this is just the easiest way to dramatically show the difference.

For a lot of us, the real concern isn't even the Sequential Workload though, its the Random Workloads, and those are abysmal on these drives as well. Depending on the Situation, I still may use one, but to be honest, for like an OS Drive, I much rather use a Sata SSD, like a 860 Evo or a Crucial MX500, then a QLC NVMe Drive, as the Random Performance is much better on the Sata for most workloads. But I am sure QLC will get better as the year(s) go on, so maybe a drive will change my mind eventually.
 
  • Rep+
Reactions: Paradigm Shifter

·
Registered
Joined
·
3,035 Posts
How can one benchmark that?
oreonutz covered just about everything in his post.

Only thing I'd say is if doing a file test, perhaps make it 200GB.

<snip>

For a lot of us, the real concern isn't even the Sequential Workload though, its the Random Workloads, and those are abysmal on these drives as well. Depending on the Situation, I still may use one, but to be honest, for like an OS Drive, I much rather use a Sata SSD, like a 860 Evo or a Crucial MX500, then a QLC NVMe Drive, as the Random Performance is much better on the Sata for most workloads. But I am sure QLC will get better as the year(s) go on, so maybe a drive will change my mind eventually.
Great explanation.

Yes, I would hope QLC improves with time and/or some fancy controller tricks... but I do struggle to see how, short of either a) a much larger SLC cache or b) a multi-tier system, say 2GB DRAM, 32GB SLC, 256GB MLC/TLC and the rest QLC.
 
  • Rep+
Reactions: oreonutz

·
Tech Enthusiast
Joined
·
1,340 Posts
oreonutz covered just about everything in his post.

Only thing I'd say is if doing a file test, perhaps make it 200GB.


Great explanation.

Yes, I would hope QLC improves with time and/or some fancy controller tricks... but I do struggle to see how, short of either a) a much larger SLC cache or b) a multi-tier system, say 2GB DRAM, 32GB SLC, 256GB MLC/TLC and the rest QLC.
Appreciate it!

Yeah, I am in the same boat as you, and 100 Percent Agree. I also have no idea how they will improve performance enough to make them acceptable for high performance workloads, apart from making the Cache MUCH bigger, or as you said some kind of multi-tiered system, that would be interesting. But they have impressed us before. I remember I thought there was no way I would find TLC Drives acceptable, and while I still prefer MLC or SLC, there are a few TLC drives that have really impressed me on Random Performance, so perhaps, somehow, they will overcome it on QLC Drives to, but again I am like you, incredibly skeptical, but curious to see what they try next.
 

·
Registered
Joined
·
106 Posts
I made a test with a new Threadripper system:
3960x (water cooled)
64 GB Ram (B-Die) 3200 14 via XMP profile
3 x M.2 SSD + 1 x SATA SSD (all Samsung)
RTX 2080 Ti (water cooled)
RTX Quadro 4000 (air cooled)

Except from XMP for ram all is set to standard or auto (no OC).

I let Aida64 run a stress test (27 minutes) with all components active (CPU, GPUs, Disks).

Only issue is the sporadic dxg kernel latency.

(temp of 83 degree is the air cooled Quadro).
 

Attachments

·
Tech Enthusiast
Joined
·
1,340 Posts
I made a test with a new Threadripper system:
3960x (water cooled)
64 GB Ram (B-Die) 3200 14 via XMP profile
3 x M.2 SSD + 1 x SATA SSD (all Samsung)
RTX 2080 Ti (water cooled)
RTX Quadro 4000 (air cooled)

Except from XMP for ram all is set to standard or auto (no OC).

I let Aida64 run a stress test (27 minutes) with all components active (CPU, GPUs, Disks).

Only issue is the sporadic dxg kernel latency.

(temp of 83 degree is the air cooled Quadro).
I usually find Drivers to blame for this on AMD and Intel Systems. Typically Installing the Free Version of Process Lasso and keeping it default and running will take care of this, also running a realistic workload, like Blender tends to get closer to accurate results as to what you would get in the field.

That said, I am still yet to have a client order a Gen 3 ThreadRipper system from me, so I am un aware if the extra CCD's and different IO Die have more trouble with Latency then the AM4 Platform.

But this exact dxg Latency that you are seeing while running an AIDA Stress test, I have seen on the AM4 platform, and can confirm in my cases, that Installing Process Lasso, along with using Blender or some other Real Workload to stress the CPU (I always use Blender for this though because it tends to be the closest to the type of workload you would be doing in audio video work that is easy to script and repeatable) will solve the issue, if you consider it one. (Process Lasso tends to do a better job of Scheduling then Windows does, so I tend to use it anyway on Audio/Video Production Machines, but that said the Direct X Latency seems to not cause any issues when actually running a Recording/Mixing/Editing Session.)

I really appreciate you posting your results though, and would love to see if after you use it in a production environment if you are able to document a problem caused by this, if so would be great to know so we can start working on a solution to it before we see issues pop up in studios using the Zen2 Architecture.
 

·
OG AMD
Joined
·
8,803 Posts
I made a test with a new Threadripper system:
3960x (water cooled)
64 GB Ram (B-Die) 3200 14 via XMP profile
3 x M.2 SSD + 1 x SATA SSD (all Samsung)
RTX 2080 Ti (water cooled)
RTX Quadro 4000 (air cooled)

Except from XMP for ram all is set to standard or auto (no OC).

I let Aida64 run a stress test (27 minutes) with all components active (CPU, GPUs, Disks).

Only issue is the sporadic dxg kernel latency.

(temp of 83 degree is the air cooled Quadro).

XMP is not a guarantee. Looks like a lot of page faults, try running RAM at 2400MHz and retest.
 

·
Registered
Joined
·
106 Posts
Thanks for chiming in.
I am not an audio pro, but like to understand what is going on in my system. As it is newly build by me I am now testing.

From my old PC (pre build) I know that latency is an issue. Had attached a keyboard and tinkered with cubase (only playing around, never had the time to really learn). In that machine an Intel network driver caused audio issues. Out of curiosity I am now checking the new build.

Your hint about hard page faults:
Had to read about them again. They happen if part of data is swapped to disk or if files are not yet fully in memory (e.g. parts of dlls are loaded on request). They high page fault number shown in the last screenshot came from loading applications the first time.

To cross check I have disabled the swap file, loaded applications, closed them, loaded them again. When loading them the first time hard page faults go up. When loading them again nearly no additional page faults appear. To cross check I let MemTestPro run for eight hours (zero errors).

One new screenshot is from the idle system after a Bios flash - memory is at default 2133.
Over time page faults increase slowly (virus scanner and indexing service are running in background).

The other screenshot is after the blender open benchmark finished.

I have not yet tried process lasso because I don't have a starting point (e.g. interestingly Cinebench20 shows better results with the ram at 2133 default than with XMP 3200 and the latencies of dxg seem to better as well - something to investigate).

To me that looks ok so far but what do you think? If you are interested I happy to make more tests (on holiday - so, some spare time :)

[EDIT]:
While I was writing and loading the screenshots into this page I got a new reading of latencymon - third screenshot.
 

Attachments

·
Registered
Joined
·
585 Posts
You hit the nail on the head... Looks like your XMP settings weren't stable and were actually causing latency. Get the ram properly stable at a decent speed and things will only get better.
 

·
Registered
Joined
·
106 Posts
I have a starting point now:


Screenshot is from the system after running a 15 min. render with 100% cpu load and then letting it run idle + plus screenshot and web browsing (page faults are from starting software - no issue here).



What's interesting:


Mouse:

I had a wired mouse attached but no special driver (3d Connextion Cad Mouse).
Now I have a Space Navigator Pro and a new Cad Mouse attached (both wireless). I had expected latency issues with that, but it seems to be ok.


Sound / Bluetooth:
I had disabled all sound and Bluetooth devices in BIOS and disabled them in services, but still got peaks in ntoskrnl.
Now devices are still disabled but a RME Babyface is attached via USB and enabled with all drivers (system driver and TotalMix FX are running). This seems to be fine as well.



Graphics:
Whatever I did there were always nvidia driver peaks.
Now I have disabled the Nvidia Display container LS service and it seems the spikes are gone (can't use Nvidia control panel now though).


Windows timer:

HPET is disabled via bcdedit and replaced by 0.5 ms platformtick. But I am not sure if this realy makes a difference (maybe I test that).



Tried Process Lasso:
It did not help to avoid the latency peaks.



But:
Once the system is more or less idle for some time there is always a sporadic ntoskrnl peak.


Running Aida64 stresstest with GPU, CPU, RAM, Disks active is a no go. In that case everything peaks like hell (but I think that is to be expected).
 

Attachments

·
Registered
Joined
·
106 Posts
Latency base configuration is done now.

With respect to this thread my first questions was, can this Threadripper platform provide low latency in idle?

For me I got the answer: yes.

But I also wanted to know what causes the spike once the system is idle for some time.

Answer: Core parking as part of the power saving.

I used this exercise to get an idea of how the "Windows Performance Toolkit" works. Now I can say those tools eliminate the guesswork. LatencyMon tells you that you have an issue, but the tool kit gives you hints where and why.
 

Attachments

41 - 60 of 66 Posts
Top