124 comments

[ 2.9 ms ] story [ 180 ms ] thread
Wow. Now that's what you can call the next generation of storage.

Are there any others coming out soon or are Samsung far ahead of the competition?

It's really amazing that there's no Intel desktop that can run PCI Gen4 at the moment.

Will be interesting to see how stock will be. Especially if this is also one of the few SSDs certified to run in a PS5. Which has an M.2 slot.

>Are there any others coming out soon or are Samsung far ahead of the competition?

https://www.tomshardware.com/news/phisons-new-e18-ssd-contro...

Looking at 4k random reads at queue depth 1, Samsung is superior. But worse on writes. I'd go with faster reads, since that impacts load times.

The 7GB/s is meaningless unless what your disk does is move huge files around.

Or if you're running an analytics database, where table scans are the norm.
Ok, but on the other hand DB could be write heavy. It depends on use case.

In future SeqQD32 could be way more important if things like DirectStorage utilize it better.

5.5 GBps is needed to be an effective PS5 M.2 drive.

PCIe 3.0 caps at 4GBps.

PS5 uses a custom 12 lane flash controller, not an m.2 drive. I think they optimized for things other than peak bandwidth, too, like latency from disk to VRAM and random read. (And hardware lossy compression!)
https://www.gamesradar.com/ps5-ssd/

PS5's custom SSD will be 5.5 GBps. But it supports M.2 expansion.

> "We will be supporting certain M2 SSDs," Cerny confirmed. "These are internal drives that you can get on the open market and install in a bay on the PS5. They connect through the custom IO unit just like our SSD does, so they can take full advantage of the decompression, IO co-processors, and all the other features I was talking about. Here's the catch though: that commercial drive has to be at least as fast as ours. Games that rely on the speed of our SSD need to work flawlessly with any M2 drive."

-------

PCIe 4.0 is a very flexible and high-bandwidth protocol. GPUs have been doing PCIe -> GPU transfers for some time (without any intervention of CPU or DDR4 RAM). Its good to hear that the PS5 is fully taking advantage of these capabilities.

But it does mean that M.2 drives of a certain speed are needed to effectively play PS5 games, should you use the M.2 expansion port.

My mistake, haven't been keeping up. Thanks for the correction and info.
Are these the ones that Anandtech has commented use a modified controller instead of new controller? So scaling to PCIe 4.0 is using a lot more power than it should.
Samsung isn't "that far" ahead of the competition. They've been behind. PCIe gen4 drives have been out for most if not of all 2020. The 980 Pro is faster than the competition but the others will have drives this fast out soon too.
any particular reason why you say that? I haven't looked into the gen4 drives, but looking at the anandtech storage bench, their gen3 drives are enjoying a comfortable lead over the competition. the 970 evo plus matches or exceeds the transfer rate of its competitors in most tests, and has significantly better performance consistency when the drive gets full. the only problem is that the additional performance is not always worth the additional price per GB.
>the additional performance is not always worth the additional price per GB

There are factors besides performance. Samsung is not known for particularly high quality firmware (the 830 Evo performance fuck-up, multiple queued TRIM bugs, etc.), but after reading a disproportionate amount of horror stories of silent data corruption and data loss for most of their competition, I'd rather spend the extra bucks.

what I mean is, the typical consumer will probably never notice the performance difference between a 970 evo plus (or even 970 pro) versus any other middle-of-the-pack nvme drive. I personally went with the evo plus because I expect the drive to spend its life mostly full and I'm using it to compile c++ projects from. the extra $50 was worth the most consistent performance to me.

as for reliability, this is hard to judge from anecdotes on the internet. you really need access to something like the backblaze data to start drawing conclusions. I had an oem samsung drive die in my xps 13 a week after the one year warranty expired. I have a WD hard drive that's still going strong after eight years. unless you have a lot of data, I think it's best to just back everything up, accept that you're rolling the dice, and hit a $/perf point that's acceptable to you.

I say that because this is their first gen4 drive and other manufacturers have had gen4 drives out for many months.
The thing to remember is that most people's workloads do not tend to improve much from a slow but decent SSD to a fast SSD. Workloads just aren't often large and sequential enough. You can find lots of benchmarks of actual things on youtube/google, and things like boot times, game/app load times etc barely change.

For people who do need fast, large sequential operations, I'm sure it is amazing though.

There's still a nice jump from SATA SSD to NVMe SSD (doesn't have to be top of the line NVMe).
Depends. Many cheap NVMe drives that ship with budget laptops are actually slower than good SATA SSDs like Samsung 850.

The high speed bus interface is useless if your flash and controller are garbage to the point where NVMe is sometimes no indication of a laptop's storage performance and just a marketing name.

This thread is discussing only "decent" and better SSDs. Yes, I'm sure you can find absolute garbage with an NVMe interface.
Even "decent" brands make garbage NVMe drives for the budget segment to which most consumers are oblivious because they assume NVMe > SATA.
To clarify: For the purposes of my earlier comment, "decent" applies to individual models; not a brand's entire product line.
For many users it's not qualitatively faster, as most workloads (booting, web browsing, even most video game load times) are small-to-medium random reads dominated by processing overhead.

Image/video editing, bulk file transfer, and threaded IO are the key benefits of NVME.

> Workloads just aren't often large and sequential enough

Another important thing to mention is that even on sequential write workloads where NVMe shines they actually have limited benefit. Because after the SLC cache fills up seqential speed drops like a hot potato.

So if you're copying a smallish large-file which won't fill the cache then it'll be few seconds faster than a SATA drive. Or you'll be copying a large large-file, which'll fill the cache and drop to NAND speeds so it'll still end up being few seconds faster than a SATA drive.

Watching the space I wouldn't call this next gen storage. Well. The storage is OK but the way most systems work there is no perceptible difference between an NVM m.2 drive and Sata for most regular use cases (i.e. everyday use, not server or a simulation rig). This might change in future with stuff like DirectStorage.

The truly "next gen" performance was Intel's Optane 3D XPoint, but that is ludicrously expensive.

And even then it's like boot from SSD 12 seconds at $130 dollars vs Optane 10 seconds at $1200.

The most important difference between m.2 and sata is two less cables cluttering up your pc =)
And the much higher temperatures (if your M.2 is an NVMe). Could anyone knowledgeable explain why NVMe tend to get much hotter than SATA under similar conditions? I've been testing mine with not more than 400-500 MB/s and it quickly gets to 60°C+, while the old 840 Evo doesn't get above ~45°C (running a very similar synthetic workload). I would've guessed the opposite: the fully enclosed SATA drive should get much hotter.
Why do you care? The critical temperature on a retail 970 Pro is, IIRC, 84C, and is almost impossible to hit.
Mostly out of curiosity. Higher temperatures also put additional stress on flash memory, and I tend to run my hardware into the ground (I'm still using a 50GB Intel SSD from 2010), so there's that.
I'm not sure if I count, but I suspect two things. Heat dissipation and controller complexity. NVMe controller might be more computationally heavy and so they might emit more heat*.

But assuming that NVMe and SATA SSD controllers emit the same energy, it makes more sense that NVMe would be hotter.

Think about it. If you throw a ball of high heat meat would it cool faster if it was connected to a large body of water (SSD with its chasis) or a small body of water (NVMe small piece of plastic that is the NVMe)?

Another thing to check if your NVMe either has proper thermal pads placement and/or if the air from the front coolers is reaching it.

> NVMe controller might be more computationally heavy

Which is ironic because NVMe was supposed to be more efficient and doing less work because it's design wasn't bound with SATA&spinning assumptions and controller doesn't need to handle pcie-SATA conversions since it's directly connected to the bus etc.

Not sure about yours, but on my motherboard, it floats - other than the connector on one end and the screw on the other, there is no heat sink, and it just doesn't mass very much.
m.2 cards can sink a pretty surprising amount of heat through the connector and the standoff. But case makers can also provision for cooling m.2 cards. Components on m.2 modules are permitted to be 1.2-1.5mm high. My Intel NUC cools the m.2 device through a flexible thermal interface to the system's main cooler (that also cools the CPU). This works extremely well.

One problem with m.2 is the spec would permit a module with the controller on the bottom where it would be difficult to cool. But I think so far nobody has been dumb enough to ship one.

Well a 2.5" drive has a much larger surface area from which the heat can dissipate.

There's also the location of the temperature probe, it doesn't need to be much further away from the heat source to register drastically lower temperatures in either case.

Fully enclosed SATA has a much larger surface area to radiate heat.
It helps in these detailed areas of discussion to be precise.

m.2 is a form factor, or really sort of a family of form factors.

SATA and NVMe are protocols and interfaces. In m.2, there are 12 key indexes. Drives use indexes B and M. m.2 support both NVMe and SATA. NVMe drives for m.2 are keyed at index M. M.2 SATA drives are keyed at B or often both B and M.

m.2 supports SATA and NVME, sometimes but not always in the same slot. For NVMe speeds, you need an NVMe drive. Those come in PCIe x2 and PCIe x4 at PCIe 3.x and PCIe 4.0 revisions of the standard. The drive and slot must both support this. m.2 SATA 3 drives don't have two cables to them like a 2.5" SATA 3 drive usually does. It goes in an m.2 slot that supports SATA and is keyed appropriatly.

m.2 also has various lengths: 40mm, 60mm, and 80mm. Some drives are up to 110mm and many boards provide for that. The "2260", "2280" and so on are drive sizes, 22m by the length.

m.2 also supports non-drive cards like GPS, Bluetooth, WLAN, WWAN, or NFC.

Another, earlier than m.2, standard allowed drives with a SATA interface on a card with no cables. That was known as mSATA.

You don’t buy optane for faster Boot times, you buy it as a render cache.
My point wasn't that you would use or don't use it for render cache, but that it's what next-gen SSD would look like and that even it wasn't that impressive.

But yeah, you generally use Optane either as render cache or for non-volatile RAM. Or a ridiculously fast SSD drive.

I'm very surprised to see it often mentioned that NVMe M.2 vs SATA ain't noticeable. Because when I switched from a SATA to a Samsung 950 Pro (5 years ago) and latter on to a 970, it was a gigantic speed boost. The difference was, to me, very noticeable. Now I'm a software dev and do compilation and sometimes "work with files" but it's neither a server nor a simulation rig. Just a "normal software dev" setup.

Now I do stuff like doing grep (well, ripgrep really) on an entire home dir, compute often hashes of files and things which may not be considered "regular use cases" but which I think are still "regular" or at least semi-regular when doing software development.

To me PCIe 3.0 x4 (or now 4.0 x4 I guess) NVMe M.2 vs SATA is day and night.

> I'm very surprised to see it often mentioned that NVMe M.2 vs SATA ain't noticeable.

Here is loading times for windows and some blind tests:

https://youtu.be/V3AMz-xZ2VM

https://youtu.be/4DKLA7w9eeA

Your personal experience could be down to several factors.

For example - wear of your devices, how much your SSD was occupied, drivers, other components being updated as well, etc.

What SATA Drive were they testing?

Early SATA SSD has performance consistency problem, QD1 not as fast, etc.... they were way faster than HDD, but not perfect.

NVMe started off with controller that has already been fine tuned, hence comparing them to SATA SSD tends to be a lot faster. Now if you have similar controller on NVMe and SATA you might only notice the difference in the video. But the context still matters.

> The storage is OK but the way most systems work there is no perceptible difference between an NVM m.2 drive and Sata for most regular use cases

As a developer I would just have to disagree.

After using NVMe SSDs for a while, using SATA SSDs suddenly feel like I’m on spinning rust plates.

To me it’s a night and day difference.

> Are there any others coming out soon or are Samsung far ahead of the competition?

There have been a number of PCIe 4.0 drives coming out so far.

* Gigabyte AORUS (5GBps read) * XPG Gammix S50 Lite (5GBps read, lowest power consumption for laptops)

Both seem pretty close to the 5.5GBps PS5 speed, and probably would work. 7GBps sequential is the fastest SSD I've seen so far, though QD1 IOPS is more important to the typical user.

All of current consumer Gen4 drives are just rebadged Phison E16 drives.
Tiger Lake mobile cpus have pcie 4. There are already some reviews for these laptops.
I've been waiting for this to be release, along with Zen 3 CPUs before actually pulling the trigger on a new PC. I've been on macOS for 20 years now, so it should be a whole new world for me.
There's always something better on the horizon. If you're waiting for the next thing you'll always be waiting. PCIe gen4 SSDs are already out. They're not quite as fast as this but they still do 4-5GB/s.
But this just came out. And Zen 3 is going to be announced in 2 weeks.
I agree. Doesn't make sense to do a build right now. Wait for this fast SSD and Zen 3. Coming soon.
Also need to see if AMD has a response to Nvidia.
They haven't posed a real threat to Nvidia in a while, I wouldn't suspect any difference this generation. And while Ampere is a huge upgrade from the previous generation, getting an Ampere GPU seems to be difficult or impossible.
There are horizons and horizons, they aren't all just modest incremental improvements. Some of them are big leaps in price/performance ratios, for instance the latest generation of graphics cards that just came out. Threadripper too was something worth waiting for, something worth building a new machine to leverage. Depending on your use case there can be good/bad times to pull the trigger on a new machine.
Yeah there's always something better, but there's also inflection points. Right now with Zen3 and Ampere we are in the midst of a major transition the likes of which we haven't seen in a long time. Better to wait a few months and be on the other side of this one.
not necessarily for the better! synthetic benchmarks and specialized workflows not withstanding, the hardware is really not that different, but windows is absolute trash. stay on mac friend :)
I plan on using Linux, and maybe a virtualized macOS.

for legal reasons, I am joking.

It’s amazing to consider that these were transfer speeds between the CPU and RAM not that long ago.
That's because this data is literally streaming from a RAM cache. Without the cache, this wouldn't score nearly as high on sequential benchmarks.
What does cache have to do with sustained access speed?
Holy shit, it's ~$265. I didn't realize SSDs (M.2 ones no less) came down that fast. It was only a few years ago that the price was ~$1/GB.
Yeah its really great. Makes building the pc easy too, no cables. Just pop in in the motherboard.
I think it's been longer than that - I took a look at the receipts for a PC I about five years ago and the 512GB SSD was ~$220 at the time.
I think you can get 2TBs around 200$. Yup, confirmed on newegg, not even "white label".
Four bits per cell, though. I worry about that.
I bought some lower end SSDs (sata b/c my 10y old machine) more than a year ago that were below 10c/GB.

Even at a 50% premium for latest tech, prices haven't gone down that much since.

1TB SSDs $100 to $150 for the last year or two, initially just SATA ones, but M.2 (especially QLC) ones inched down to there (with 1TB SATA QLC regularly below $100 now).

Samsung 980 is a premium SSD with much faster performance, so $265 is pretty expensive, you can get a 2TB "normal" M.2 drive for that much.

So you'd only go for the Samsung 980 if you really needed the speed.

So are the Anvil tests the only ones that do substantial random read/write? These benchmarks seem opaque.

Last I really dug into SSD benchmarking, sequential was usually quite gaudy for a lot of SSDs that were seeking benchmark placement, but the real advantage of SSDs over HDDs was random access and IOPS.

IF this is such a massive improvement, is this mostly due to:

- PCI bandwidth utilization? - 8nm silicon? - controller algorithm/efficiency?

It's my impression that controller algorithms have peaked after the first couple iterations of consumer SSDs, so we'd be left with either bandwidth or silicon.

And how less durable is 8nm cells, since flash at a theoretical level becomes more fragile with each node shrink?

Darn. Just built a machine with 970 Pro :(
I mean, it's still a great drive.
Rats you could have paid more and got half the drive life.
I made this little comparison for myself, comparing 1 TB 970 Pro, 980 Pro and Intel Optane 905P 960 GB

                             970 Pro   980 Pro   Optane
    TBW, TB                    1,200       600   17,520
    QD1 random read, IOPS     15,000    22,000   73,000
    QD1 random write, IOPS    55,000    60,000   66,000
    QD32 random read, IOPS   500,000  1000,000  575,000
    QD32 random write, IOPS  500,000  1000,000  550,000
    Sequential read, MB/S      3,500     7,000    2,600
    Sequential write, MB/S     2,700     5,000    2,200
May be it'll be interesting for someone. Most numbers are from specs, some Optane numbers are from benchmarks. I guess that sustained write load will yield lower numbers for Samsung, because they're using their DRAM cache to achieve those numbers.
IOPS heavily depend on accessed block sizes though, the promise for optane, especially the DIMM kind is cacheline-addressability.
Wow, why was the TBW halved? New NAND process?

I remember SSD lifespan going down between the Samsung 470 (32nm flash) and its successor the 830 (20nm flash), and going down even further with TLC flash.

Doesn't seem to affect the average lifespan (it's more data than most people will write) but still, it would be nice if they lasted longer instead of being ultrafast.

970 Pro was MLC, 980 Pro is TLC (they call it 3 bit MLC)
>980 Pro is TLC

Any news on 980 Evo? Considering this downgrade, I'd assume Evo will run on QLC.

Ah, that explains it.
I checked Samsung server SSD SM863a MZ-7KM960N 960 GB. They claim 6,160TBW. It's using SATA and much slower than M.2 SSD though.
But server hardware is more expensive and can be harder to get.

They really should make a "high endurance" SSD model for consumers. I think many people would like that over benchmark speed. Personally, I can only saturate SATA 6 Gbps when doing backups.

In 2D MLC era, 50-75TBW for 500GB drive was common. Now in 3D TLC era, 300TBW for 500GB or more is common. It's enough endurance already and that's why industry moving torward 3D QLC.
(comment deleted)
These sequential write numbers are for small writes. For large writes both 970 Pro and 980 Pro have a performance drop (don't know how Optane behaves here). The 970 Pro dropped to around 2000 MB/s, for the new 980 Pro the performance drop will be even larger due to using TLC (3 bit MLC in marketing materials) instead of MLC.

This and the durability were the main reasons to pay the premium on the Pro line. Not sure what will be the value compared to the competition once the new Phison controller is out (or even Evo when there is a model for PCIe4.

I've not heard that before, NVMe SSDS are block devices, not file devices. So why would sequential writes care if it was one file or many? Maybe you are misunderstanding queue depth for large vs small files.
It appears that you have misread the comment you're replying to. The speed drop-off once a write is larger than the cache can hold is a well known issue with SSDs, especially those which use slower NAND like TLC.
Wow, I did misread it, thanks.
It's so fast, you can wear it out in just 33 hours of continuous writing!
Reminds me of when 4G came to my country. At the advertised speed, you would hit the data cap in minutes.
What can possibly sustain that rate (other than cloning another SSD of the same make over and over)?

The wear speed is actually the same as a 970 Evo @ 500GB, 3GB/s and 300TBW. There’s zero chance you’ll hit that during a normal computers lifetime even if you run data science tasks all day.

Imagine some buggy program which writes stuff and deletes it over and over again because of some weird condition. I could even imagine that some website could do that.

And you probably won't even know it! SSD is silent.

Hey maybe you or someone here can help me with this interesting problem I have. You see I wanted to put an Intel Optane SSD into an external enclosure to benchmark the absolute fastest Windows and Linux install and compare it to install times on other media(yeah I know i'm weird).

So I bought one of these SSDs off of ebay:

https://www.ebay.com/itm/Intel-Optane-Memory-M10-SSD-M-2-228...

And put it in this enclosure:

https://www.amazon.com/gp/product/B07MKCG5ZG/ref=ppx_yo_dt_b...

Instead of being recognized, one of the chips on the Optane ssd started to get very hot. Was wondering if anyone knowledgeable in Optane could explain where I went wrong here? Is it even possible to put Optane in an external enclosure and write to it like an SSD/external drive?

I am excited about Optane because it just seems like such an exotic technology given the interesting performance gains you get in terms of small file throughput.

The enclosure you linked says it is SATA only. The Optane drive interfaces as PCI-E. Note this photo here:

https://images-na.ssl-images-amazon.com/images/I/61vzuDqQjJL...

In the specs for the Optane drive you linked, it says its interface is a M.2 2280 (22mmx80mm) PCI-E 3.0 NVMe. The M.2 2280 is a physical port interface, the signaling is PCI-E, and the storage protocol is NVMe.

https://ark.intel.com/content/www/us/en/ark/products/135581/...

You'll want to get a different enclosure which mentions support for PCI-E NVMe drives. EDIT: I wonder what the impact of it going over USB and the additional USB to PCI-E controller. You should probably see about getting a Thunderbolt enclosure so theoretically the PCI-E interface doesn't have to go through yet another translation, otherwise you'll probably just kill the performance you'd hope to get out of the SSD.

Hmm interesting. So even though the physical connector with the two notches fits, the electrical signaling is wrong. Thats seems like a bad design no? All my desktops are on SATA so I am used to a situation where as long as the connector matches (ie, SATA or DDR notches) then I am good.

I wonder if I have burned my SSD by plugging it in...i'll have to find a adequate enclosure to find out. Although if I have burned it, then maybe it could then damage the new enclosure...hmm

>I wonder what the impact of it going over USB and the additional USB to PCI-E controller. You should probably see about getting a Thunderbolt enclosure so theoretically the PCI-E interface doesn't have to go through yet another translation, otherwise you'll probably just kill the performance you'd hope to get out of the SSD.

Hmm, the machines I am testing on don't have USB-C thunderbolt. I'll have to look into this. So I can see raw bulk performance being limited by the USB bus but would 4k performance be limited as well? I thought this was a limitation of the Flash memory and SSD controller.

> Thats seems like a bad design no?

Yes, that socket is not standards-compliant. It should have been B-keyed.

> Sold by SSK Direct > Business Name:Shenzhen Huajianli E-Commerce Co.,Ltd

No surprises here.

Is there a good brand of enclosures that I could rely on? I am not sure if this is a typical use case such that the big guys would manufacture an enclosure that's compatible.
NVMe is only primarily used for performance, attaching it to USB would negate the benefits. So you would have to look for a thunderbolt enclosure, assuming your PC has thunderbolt and those are rather pricy. USB to NVMe adapters do exist but I have no experience with them so I cannot recommend any.
When you're using the USB interface, there is some chip in between the actual SSD and your computer. USB doesn't directly handle reading and writing to SATA or NVMe drives, so software on your computer encodes it to go out USB, and then some chip takes that USB signaling and converts it to read/write to the SSD, and vice-versa. If the controller on the USB interface is poor quality, it may reduce the performance of the overall set up. I've had external regular SATA/USB enclosures have varying performance despite being the same underlying drives inside.

On a simplistic view: SSD <--> enclosure USB device <--> USB chipset on your PC <--> your processor

Compared to: SSD <--> PCI-e interface <--> your processor

Also note that USB is by definition a bus. Multiple devices can live on that bus and contend for the same bandwidth. If your mouse and webcam and keyboard and USB WiFi adapter and what not are all on the same bus, activity on multiple devices reduces the max throughput you have available to your USB device. This (usually) isn't true with PCI-e, which (usually) has dedicated bandwidth between the device and the CPU.

This isn't to discourage you from using this drive however you wish, just letting you know of potential pitfalls and challenges along the way. You may get the performance you wish from this, but the bottleneck of IOPS is far more likely to be limited by one of the chips in the USB path rather than the SSD itself.

>When you're using the USB interface, there is some chip in between the actual SSD and your computer. USB doesn't directly handle reading and writing to SATA or NVMe drives, so software on your computer encodes it to go out USB, and then some chip takes that USB signaling and converts it to read/write to the SSD, and vice-versa. If the controller on the USB interface is poor quality, it may reduce the performance of the overall set up. I've had external regular SATA/USB enclosures have varying performance despite being the same underlying drives inside.

Yes, I understand this. What I am thinking of is the overhead of reading and constructing each block from flash. Is this really limited by the USB controller capability(I guess theoretically yes since its in the chain?) or is it just limited by the Flash controller and typical USB controllers. I guess we can test a theoretical bottleneck by running an IOPS test and comparing it to official Benchmark numbers of that Optaine model. That way we can eliminate the USB controller if the numbers match up.

This is the structure from what I understand. PC Host Controller<--->USB Controller(in the enclosure)<--->Flash Controller(On the SSD)<--->Flash chip(On the SSD). Is this wrong?

So here with Optane we have a massively high throughput of small files but a lower throughput of large files.

Is this scenario something that a USB controller would noticeably slow down? I have never tested this before.

A regular SSD has lower throughput of small files in exchange for high throughput of large files(slower throughput but higher bulk transfer) which is noticeable if the USB controller was not up to the task (See most USB Drives which have this + flash controller on the same board)

Basically the whole reason this idea came up in my head is that I notice that despite using high bandwidth drives to install clean Windows and Linux copies, I notice certain bottlenecks in the install process that I suspect are a result of many small files being copied/unpacked(either copying package files in Linux or unpacking the WIM in Windows).

So that led to me wanting to benchmark a High IOPS drive and comparing it to low IOPS to see once and for all if the install process is choking on the drive vs something else (ie. the CPU capability for decompressing). Its just a weirdo side project that I wanted to try.

Eventually I even wanted to somehow try a ramdisk type scenario although I don't think that sort of hardware exists.

This is also motivated by the fact that I don't have a way to intercept the file reading process on the drive as they are being read to see if the drive really is the source of any bottlenecks or if its the installer code.

> A regular SSD has lower throughput of small files in exchange for high throughput of large files(slower throughput but higher bulk transfer) which is noticeable if the USB controller was not up to the task (See most USB Drives which have this + flash controller on the same board)

To a large extent that is a software problem. The installers/package managers are neither pipelining nor parallelizing their operations properly. If you're dealing with many small files then on the reader side you want to issue a window of readaheads for a certain size across multiple files rather than doing one file at a time and on the writer side you need to batch your fsyncs.

> So that led to me wanting to benchmark a High IOPS drive and comparing it to low IOPS to see once and for all if the install process is choking on the drive vs something else (ie. the CPU capability for decompressing). Its just a weirdo side project that I wanted to try.

There are PCIe slot to m.2 adapter cards, those are more efficient and cheaper than external adapters.

This seems to confirm what I'd read earlier, which is that they're cutting the TBW durability rating in half for this generation. I don't have PCIE 4 at this point, so I'm considering sticking with a 970 for an upcoming upgrade.
May I ask what workload(s) you have that cause you to write that much to your SSD? Looks like with this drive you could write an average of ~330GB per day over 5 years and stay within the 600TB write warranty (on the 1TB version, at least). So that's 0.3 drive writes per day, which isn't unheard of for non-enterprise drives.
Not the GP but I'm in a similar position. I run Hauptwerk[0], a software package that simulates pipe organs. It regenerates an on-disk cache every time an instrument's parameters are changed and this cache can be tens of gigabytes in size. 5 parameter changes a day for some of the larger instruments would put me over your 330 GB/day threshold.

[0]: https://en.m.wikipedia.org/wiki/Hauptwerk

Super interesting, thank you! I imagine there are other music software out there that have similar behavior. What kind of storage do you use to run Hauptwerk and take the beating from all those writes?
I've got a 2 TB Sabrent NVMe drive right now. Apart from budgeting for a replacement drive every few years I don't do anything special.
This is facinating - but would Optane or large RAM-disk work for you?
Reviewers never mention the real tech specs, although I appreciate that this article does mention the number of command queues the device supports. Three things I'd like to know before I buy are 1) how many namespaces are supported, 2) what logical block sizes are supported (the retail 970 supports only 512 bytes) and 3) what is the critical temperature threshold? The output of `nvme id-ctrl` would do the job.
interesting that it's a review for a "consumer" drive and contains purely synthetic results and no real world discussion whatsoever. just shut up and buy it kids! does it make any perceptible difference in your life at all? well you can tell the kids with 970s that youre cooler than them
Anandtech’s reviews are usually better and include at least a few compilation benchmarks (Linux, Chrome, etc). Maybe they’ll post one soon?
from their conclusion,

"the fundamental problem facing the 980 PRO and other high-end NVMe drives is... very few real-world consumer workloads... make good use of its full performance potential... 1M IOPS... sounds impressive, but that's only possible in fairly unrealistic conditions... the peak... performance of the drive simply does not matter to consumers today.

raid it up yo.

Does anyone remember Fusion-IO?

It used to have the fastest SSD on the market. What happened to it?

I used those with SQL Server back in the day. Sandisk bought them
SanDisk bought them in 2014, and after that, I have no idea what's happened to their technology...
In some ways Fusion-IO was the prototype for NVMe and as with many standards the innovator didn't end up winning the market.
Former Fusion-io employee. wmf is right in broad strokes: they got commoditized and couldn't figure out how to regain an edge before their fortunes dwindled.
As always, benchmarks depends on your application! It appears that for my workflow, this drive is a substantial _regression_ from the 970 Pro.

I have a firehose of data that needs to be stored without dropping samples for a scientific application. Thus the best benchmark is the “whole disk fill” benchmark. Samsung 970 Pro sustains 2500+MB/s The Samsung 980 Pro thrashes all over the place, going as low as 1200MB/s!

https://www.anandtech.com/show/16087/the-samsung-980-pro-pci...

It must be that 3-bit MLC always requires a write cache in the NAND. Once the write cache fills, the performance decreases until the cache is flushed. In this way, the new 980 PRO is really more like an EVO drive. The 970/960 PRO used 2-bit MLC which doesn't use a cache and therefore has consistent write performance across the drive. In addition, 2-bit MLC will have better endurance than 3-bit. In this way it does look like 980 PRO is a regression.
970 Pro uses DRAM cache.
We're talking about SLC caching in the flash itself in addition to any DRAM cache. The advantage of SLC cache is temporarily faster writes but the disadvantage is that write performance is variable.
My guess is this is the first PS5 rated / approved m.2.