If the 673A’s like similar small box NASes I’ve worked with it’ll be not great for active temperatures at ~8 W/hard drive. Wouldn’t expect 25+ W/drive enterprisey stuff to go well (if at all), especially in deployments where noise is a concern.
In my experience damping often increases noise because higher temperatures and greater flow obstruction just end up ramping up fans. Probably best to plan and measure both ways.
Last I checked there weren’t any 20 dB(A) consumer NAS drives. Blue and Black aren’t rated for power on hours and, while Barracuda’s rated for 2400 hours/year, it’s not speced for noise. The 'cudas I’ve used combine noisier than Ironwolf with being 2-8x slower. ¯\_(ツ)_/¯
Only SSD failure I’ve hit so far was a Crucial BX200 that did the same thing. We’re down to couple SSD still deployed (all the other flash is NGFF drives or USB sticks), so it’s conceivable that could end up being the only one.
Can you articulate much of a value proposition for bus EC4 over DDR5 read+write and ODEC2? All I see is certain DED cases.
I would argue something like a define r5 has enough cooling for 6 drives and a cpu doing almost nothing all day while not ramping up the fans and being quiter then some more open encloures where there is only some plastic between you and the drives
My servers are all on DDR4 platforms at this point so I haven’t really read up on the particulars of DDR5 yet. The little I have read is that some of these particulars on DDR5 make non-ECC DDR5 less error prone than non-ECC pre-DDR5 RAM, but that it is no substitute for proper ECC.
One interesting exception might be the “in-band ECC” solution Intel has implemented on some of its embedded products which can provide ECC like function on standard non-ECC RAM, at the cost of a slight impact to performance and available memory.
I have considered using a system based on the Intel N97, N100 or N305 CPU’s for ZFS, and utilizing this in-band ECC setting. These could be pretty cool for a nice lightweight build.
there is an example of that here Real Ultimate Plex Server by dk9 - RTX A4500, Fractal Design Define 7 ATX Mid Tower - PCPartPicker though you may have different needs in regards to the software selection. Not sure I would call this build “low power” though since it idles around 150W, maybe there is some fat that could be trimmed off it. Though strangely enough I recently swapped the AM4 configuration there for an AM5 + Ryzen 9900X and the idle power draw is exactly the same. Which is something to think about if you have more budget and decide you also want compute power in your storage system.
I wonder how much difference there really is between Pro models and Non-pro models. They should all support ECC from what I understand.
And some of Microcenters AMD bundles are often priced very nicely. (f you happen to have a MicroCenter nearby, which I completely understand not everyone does)
They just don’t offer any Pro-model CPU’s in bundles.
Don’t get me wrong, I’d totally use an AM4 board and CPU if I already had them, but it just feels wrong to buy into an EOL socket at this point when you can get the current one at such an attractive price.
A bit late to the thread, but I second the WD NAS drives. I found their 16 GB variety to be reasonably quiet and performant. I also have a few Exos drives, they’re not bad, but their actuators are a bit more audible. I sit about 6 feet away from the box, so I notice drive noise, just like you do.
I also second @JeffreyW about exchanging the fans in your QNAS box if the current ones are bothering you. Well worth the couple of $. Noctuas are good, some Arctic fans can be just as quiet and cost a lot less. The fan in one of my older DAS boxes eventually got quite loud, probably after the bearings started wearing out.
Lastly, as your NAS is (from your OP) second hand, check that the rubber (PU) grommets that isolate your fans and your drives from the case are a. installed in the first place and b. still in reasonable shape. I personally find the noise resonance makes especially bothersome.
for AM4, the non-PRO APU’s do not support ECC I think. The rest of the Ryzen line should all support ECC. there are some other very minor differences described here, which I am not sure are particularly relevant to most use cases What is AMD Ryzen PRO: Advantages and Disadvantages - Profolus
You can approach the subject on the principle of buy the cheapest and use it until it breaks. If you don’t need super speed and it won’t be burdened with a million r/w then… You should always have a copy of the data from it anyway.
I’ve been using a regular 2.5 5400RPM Samsung in my home NAS for years, which currently has 2130.5 days of work time. Although when it’s not being used it goes to sleep. And its condition is such that it has 10 relocations and one pending. But as long as it works I won’t change it…
The moral of this is that the cheapest drive at the time working as a NAS still does the trick and didn’t cost $$$. Whether you have to buy a top shelf drive right away… depends on your needs, but if you don’t abuse the drive then a cheap one should come in handy for some time.
The point here is consider used enteprise ssds, if you have cash and case with good enough airflow to cool them.
Airflow being critical point here. U.2 drives can have 2x-5x power draw than nvme drives and they rely on passive cooling via metal chassis.
Affordable and large capacity consumer ssds are QLC by default, and when coupled with low cost controllers also by default they are completely unsuitable as you observe. Also insanely overpriced given what they are underneath.
Caveat is, solid state electronics in drive and controller might produce high pitch electronic whine thats more annoying that hdd rumble if you sensitive to it.
@twin_savage
Its weird how unpredictable this is. I just evacuated some irrelevant data from 1TB Samsung T5 portable ssd two years after last write and year since in was plugged in.
I expected some trouble, but there were no slowdowns or data corruption even year after drive being last powered on, and I transferred about 480 GB of data.
Yet few generations back samsung had the legendary 840 EVO that brought the performance implications of unrefreshed data to the front of public perception and required firmware upgrade to mitigate, not fix.
And that drive had issues for 6+month old data on drive that was powered on daily.
I would say data longevity is likely solved issue on normally used drives from reputable vendors.
Its just nobody test or reviews this, so anectada is only source of information we have.
If you’re ditching the 673A, yeah. Define 7 (5s are long done)'ll still be worse than an airflow case for temperature normalized fan noise, just with a good chance of remaining below your ambient noise floor, and the damping doesn’t do much at lower frequencies. Where it might help a little is in the higher tones off active actuators but the limited amounts of data mostly show different cases are just different, not really quieter unless specific components and workloads happen to line up with the specific properties of the case in its installed location for the listener positions of interest given the surround materials.
So best to plan and measure both ways. A potential advantage of small box NASes here is they can be pretty good about sleeping when not needed, which can end up with less active and idle noise as well as fewer drive hours, lower temperatures, and lower fan speeds. Even if a desktop build does spin down it’s unlikely to be competitive with ~2.5 W sleep.
Just some tradeoffs to think about.
Other caution I’d make is Fractal Aspects are mostly pretty average and the Dynamics and their older fans are mostly worse. Data I’m aware of for the larger Defines’ front drive array shows there’s probably a few dB(A), up to maybe like 4 °C noise-normalized, or combinations thereof to be had from fan changeouts. I’ve also found the same fan and same drive in different cases’ drive bays and front intakes can require substantially different RPMs for the same ΔT and that data is not favorable to Fractal’s drive sleds. So it looks to me like fan selection and bay loadout config can make more of a difference than the case. I don’t have enough data to suggest an optimum, though, and it’s not like you’ve many choices with 6+ 3.5 bays.
Nope, take a look at AMD’s connectivity specs. Or any of the numerous threads here on the topic. ECC’s crippled off Zen 3 non-Pro APUs and the 8000 series lacks support.
But 7400F, 9600, and up, they got it.
Mmm, any decently tested system should be having zero bus errors (we’ve got multiple threads here where folks have been unable to induce bus errors despite trying hard to overclock to the edge of stability, for example). Corollary to that is EC4 closing DDR4’s read CRC hole has a low probability of on bus correction. Similarly, bit flip probabilities are vanishingly low given decent DIMM cooling, non-stupid tRFC configuration, and non-aerospace operation. So EC4 mainly guards against DRAM degradation. But the majority failure mode there is chipkill, which requires EC8. I’ve also had a really hard time confirming an EC4 SEC triggers writeback to correct DRAM content. Presumably it does but the apparent lack of documentation actually saying IMCs do that is curious.
So, while DDR4 EC4 has high perceived value and brings peace of mind, I’ve consistently had difficulty articulating much of an engineering value proposition outside of high rel environments where it’s difficult to take an hour of downtime for a periodic memory test.
DDR5 closes DDR4’s read CRC hole, ODEC2 scrubs easily happen more often than EC4 reads, and there is documentation confirming scrubs perform correction. EC2 doesn’t DED but it looks to me like DDR5 DRAMs could easily end up SECing more reliably than bus EC4. The clearest DDR5 value proposition thus seems to be for EC8, but that only applies where you can find EC8 RDIMMs instead of EC4. And if you’re not doing datacenter kinds of things probably it’s ok to take some downtime to replace a chipkilled DIMM.
Another thing I get stuck on is, ok, suppose we’ve figured out how to have servers full of wonderfully proper ECC goodness. But none of our clients are ECC. If a client machine pulls some data, has a single-double-triple-whatever bit error, and pushes an update back to a server all of the ECC goodness is bypassed. Server centric workloads avoid this but bypass is usually the mainline use case for ECC NASes and file servers.
Also, suppose some SECs and some DEDs get logged off of an EC4 or EC8 installation. SECs, ok, cool, presumably correction’s complete and no action’s required other than maybe swapping hardware if the rate gets too high. But if there’s a DED, what can you actually do to identify which data is corrupted and repair it? If a zfs or a ReFS integrity stream can fix it you didn’t need ECC, and otherwise you’re probably just screwed. So it’s hard to see how having DED log entries makes much difference beyond providing a diagnostic indication for OOSing hardware and pulling DIMMs to limit further data corruption.
TL;DR I’d feel better if I could define what proper ECC is but it’s proven rather difficult to identify truly compelling use cases. The value proposition’s niche for DDR4, more niche for DDR5, and seems bound up with “well, we want to operate our hardware this way as opposed to that way” kinds of considerations.
I’d still argue anything SMR is very very bad for anything RAID. Rebuilds can take a week, and can fail in process due to the SMR drives being too slow at that sort of activity.
I consider those signs of impending failure and swap out the drive as soon as convenient.
Yes, I have redundancy (RAIDz2) and backups, but if you have several drives that are marginal like that, whats to prevent two or more from shitting the bed at the same time, or (more likely) one shitting the bed, and another one or two failing due to the sudden high load of a resilver?
And restoring things from backup is a pain i the butt. It’s an important option to have, but…
Sounds like you are way more well read on this subject than I am. I don’t claim to fully understand the inner workings of RAM. I just consider proper ECC with ECC error reporting to be a valuable safety measure on anything “server”. so I can tell when things go wrong, and do something about them.
Yes, the clients can corrupt files, and when that happens it sucks, but at least you can try to restore those files from snapshots.
My nightmare scenario is a pool scrub happening at the same time of a RAM failure, and the pool proceeding to rewrite all the data trashing it, due to bad RAM returns. The risk of this is low, but not zero, at least not without ECC.
Hmm. Could have sworn I heard of AM4 and AM5 user running ECC on regular Ryzens. Maybe I am conflating memories (Could use some of that ECC myself).
Yes, that’s all right. That’s why I clearly stated that it’s a slow solution.
And more in the context of a single disk and not several in a raid.
It’s simply the cheapest option if the OP needs to limit $$$.
It doesn’t mean that it’s a good solution, it’s just cheap and nothing more.
This particular small home NAS is a single disk and I have the backups I need. I don’t care about it, it just works, it’s been working for years. It’s normal that the disk is a bit tired after so many years of work and is slowly starting to fail… but at the moment I’m not in a big hurry to spend $ on a new one.
And no one is signing up to sponsor a new one.
The amount of TB that I’ve recovered many times from different copies has hardened me and doesn’t scare me.
If it was a large NAS with many disks and very important data, the situation would be completely different and it would have been replaced a long time ago.
OP didn’t define a budget and didn’t clearly state how many of these drives he wants. Only that it should be quiet and “enterprise reliability” and 8TB+ which sounds a bit funny in the context of a NAS found on the road.
But since we know what kind of box he has, we can assume that he wants 6 8TB and preferably more. But without a specific $ it’s a bit of a guess…
My take is that it is often a pain in the ass to restore from backups, so I’d rather just have my pool not degrade, but you sound like you know what you are doing and the risks you are taking.
Thing is, you never know when you are talking to people on forums.
It runs the gambit from people who have home labs with impressive redundant pools and multiple remote backups to people who didn’t realize it was risky to put all of their important data on standalone drives, run basic RAID5 pools not realizing they are not protected during rebuild if a disk fails, and at the same time thought RAID negated the need for a backup.
So whenever I say anything I err on the side of caution just in case the person I am replying to lacks some of this knowledge.
Yup! It’s all in the Ryzen and Ryzen Pro specs. ASRock has full line AM4 and AM5 desktop mobo support. Last I checked Asus and Gigabyte crippled EC4 off their low and midrange but offered it on some of their high end. MSI still wasn’t doing anything.
Pro 350, Pro 370, 9600, 9600X, 9700X, 9800X3D, 9900X, 9900X3D, 9950X, 9950X3D, …
If it’s software RAID, yeah. If Broadcom says anywhere whether they’re using ECC on their HBAs I’ve missed it, though it’s not something I’ve really dug into. There’s not much open datacenter data for DDR4 and DDR5 but what I’ve been able to find mostly suggests for failure modes that’d blow past SECDED. So that’s back to EC8 RDIMMs.
Flip of this is the current gen small box V3C14 NASes I can think of offhand run DDR5-4800 EC4 SODIMMs. Pricing is ridiculous on the ones I’ve looked at compared to previous gens but TCO’s still low compared to a server.
Personally I’d in principle be ok with non-ECC parity RAID given the right file system and journaling backups. ReFS integrity streams ain’t it, though, and zfs is hard to deploy at work because of how Windows centric the environment is. I also inherited a btrfs array that had lost a drive at one point and was not impressed, though it was hard to tell from the logs whether or not the rebuild failed due to user error.
Another hole I see is tools like rsync and robocopy have no way of checking file integrity, so will propagate corrupted updates to backup.
Yeah, we’ve got one RAID5 array I’m not all that happy about. It’s got hot and cold spares but is growing to enough TB/drive I’m getting to where I’d feel better with a RAID6 promotion. It’s backed up nightly to RAID1 shards that mostly spin down and sleep, so it’s not like our ability to tolerate correlated loss of 2-3 drives is poor, but rebuild time on a major fail’s going to be over two days.
Yeah I think Ryzen Pro is the way to go for ECC support and I am still wondering if I just sell the QNAP on ebay and go DYI what case I should use to get a good compromise between cooling and noise. The R5 is laying around in my basement so it would be the obvious choice.
Well, for what it is worth, I recemtly moved my Threadripper 3960x system away from an “everything, work, home productivity and games” system to a dedicated workstation.
When doing so I decided to upgrade to more RAM and go ECC.
I was going to order Crucial kits, but from the reviews out there it sounds like their quality may have been slipping. They also do not offer 4-way kits that have been tested together, but OWC does. There were several reviews of Crucial DDR4 ECC UDIMM’s suggesting they couldn’t get all 4 sticks to work together.
This OWC kit wound up costing just a little bit more, but it may have been worth it for the “it works out of the box” peace of mind.
So, I popped the ram in, and set everything to “Auto” except the RAM speed configuration which I set to DDR4-3200, and all of the RAM was detected in POST and just booted up.
Best way I have found to confirm whether or not ECC is indeed working on an AMD systems is to - under Linux - look for whether or not the kernel has automatically loaded the amd64_edac kernel module.
Success there too! (though I am a little confused about the numbers here. I definitely have 128GB of RAM available in system, but I only see 4x16GB reported here? Weird. )
Either way, as far as linux is concerned, it looks like I have functioning ECC.
Before calling this a win, I plan on running a full memtest. I loaded it, and the Passmark version of Memtest confirms functioning ECC, but I didn’t have enough time to leave it for the duration of the test. Maybe I’ll do it tonight.
All of that said, the Threadripper 3000 series is Zen2, so that does align with your statement above about it working on older Zen chips.
It is also a Threadripper. Maybe - what with Threadripper being closer to Epyc than to consumer Ryzen, they do things differently. But technically only Threadripper PRO is supposed to support ECC.
Maybe this is an ASUS thing, hacking something they weren’t supposed to into their BIOS and possibly pissing AMD off in the process?
Who knows. But I now have 128GB of working ECC RAM, and I am happy with that. The Threadripper 3960x is certainly not getting any younger. 5.5 years is a long time in the CPU world, but it has enough capability for my work needs for the foreseeable future.
I did the math, and it should be much cheaper to just have a dedicated game machine, and chase the latest and greatest on that, rather than continue to shoot for the “all in one” machine, and thus have to overspend on the latest and greatest workstation parts, only to still get disappointing performance in games.
But wow. I just realized I typed this reply in a thread with a completely different topic. Sorry about that OP.
Sorry about that OP.