Wish I could afford this. It’s on sale and I still can’t afford it.
This is 1st gen Optane. NAND has caught up quite a bit in the last years…just check on the latest Enterprise SSD models and they often beat 1st gen Optane like the 905p. 6 years of progress and innovation is no joke in today’s storage market.
So you may pick some up cheap Enterprise SSDs in the coming years for probably less (and way more capacity).
I’ve found some Intel U.2 drives that are around 2 TB for 75 or so on eBay. They are used, so I recommend a warranty through eBay and making sure that the drive is healthy when it arrives. But just get a nvme to u.2, and you’re good.
It’s not Optane, but you can do storage tiering in Windows with PowerShell, I hope Wendell does a video about this someday as it is super helpful.
Github link, including scripts for PowerShell deployment:
You can do this with both regular sata SSDs but also nvme and PCIe SSD so you can even have tiers of SSD. I’m running a pair of 8TB hdd’s with a pair of SATA SSDs with plans to add PCIe/nvme storage.
I was shocked myself when I found a NAND-based SSD with 27 DWPD and 15 μs latency. Optane-level performance is definitely within reach without the Optane.
The big NAND manufacturers have been chipping away at Optane’s endurance supremacy and will eventually get around to latency too—but probably not for a few years and not at consumer-affordable prices.
For now, Optane is still an order magnitude better in terms of latency among the selection of SSDs that consumers can actually get their hands on.
The advantage of NAND is that the capacities have gone up. You can get 25TB drives with 3DWPD…that’s 75TB/day for 5 years. SLC and Optane may have bigger DWPD numbers, but if you can’t scale the capacities, PBW are the better approach there. Because 75TB/day is 25DWPD for a 3TB drive. And you can still overprovision your 25TB to get more if you like.
And while you can hit those limits with 24/7 sequential PCIe4 bandwidth, anything >10DWPD is of academic value at best because a day only has 24 hours. There is only so much data you can write at 7GB/s in 24 hours. First world problems
Micron lists 3-4x the PBW/DWPD values for sequential workload. These are marketing claims…but if true, we’re in >10DWPD territory.
Latency is still Optane territory. NAND isn’t anywhere near read latency, but 15μs on writes is getting damn close.
Just for science/math, I ran the numbers, and it doesn’t totally agree with what either of us have been saying…
- 86400 seconds in a day
- 7,500 MB/s cap for a four-laned PCIe 4.0 SSD
- 618 TB/day theoretically attainable by a four-laned PCIe 4.0 SSD
Let’s use the Intel Optane P5800X 3.2 TB SSD as a reference point. (It can actually achieve 7,500 MB/s. ) Being rated for 100 DWPD, that’s 320 TB/day or 584 PB total, and you’ll blast through the warranted write endurance in just 2½ years. (That is unrealistic of course; you’d have to be using Optane with a 100% sequential workload to reach that throughput, which is obviously not what one pays ten times more money for.)
The Intel Optane P4800 1.5 TB lasts a little longer with 60 DWPD, 90 TB/day, or 164 PB total, but only because its maximum throughput is 2,400 MB/s, so it can only hit its warranted endurance in a minimum of 4 years.
There are contemporary SLC NAND SSDs with endurance in the ballpark figure of 60 DWPD in the best case scenario (e.g., Micron XTR). And then there are capacity-oriented NAND SSDs which have low endurance, but a lot of potential for over-provisioning.
The Micron XTR 1.92 TB has 1.92 TB and a maximum endurance of 60 DWPD, assuming 100% sequential workloads with 128+ KiB writes. To wear it out, you’d have to be doing 115 TB/day or 210 PB total—far from Optane P5800X’s numbers and achievable even with paltry dual lane PCIe 3.0 and in just 7⅝ months.
The Solidigm D5-P5336 marketing material quotes 65.2 PB total endurance (secondary source), which works out to 36 TB/day—even less than the tiny Micron XTR 1.92 TB despite the massive capacity advantage. At its maximum write speed (3,300 MB/s), it’ll hit its warranted endurance in 7½ months.
Turns out I was wrong about NAND endurance catching up to Optane. It’s still far from it. But maybe it doesn’t really matter in real-world workloads, and NAND is already good enough.
Hi, can someone here link me a nand based product that will hit 15$\mu s$ of read and write latency for a sustained 4k random load? While you are at it do you have a mixed read write “bathtub curve” we can look at for comparison to gen 1 Optane? I don’t think a product exists that will hit that latency number in both read and write, and finding a nand device that has a straight line across a bathtub curve is going to be very unbelievable to me.
I got one of the “used” 6.4tb Intel u.2 drives, d7-5600 or 5700 (3dwpd, 7GB/sec 5years), anyways according to the drive tool it had zero on hours. Under $400.
The amount of information that people know about any given tech/hardware topic on here is nuts.
Solidigm has a new nand device that gets closeish
Nice Device Wendall. Would this be good for a “/playground” device? Where projects can utilize the disk for data integrity?
“/playground” = “/workspace”
so if you have lots of data moving in and out for projects, big or small?
Does it play well with Linux?
say
Linux Kernel 6.7.6-arch1-2 ???
Hey I see you guys really know your stuff…would an Intel 905p PCIE SSD be good for gamig in 2024. I have ai7-13700kf, RTX 4090 GPU, 32GB DDR4 RAM @ 4000 MHz, and my current SSD is the 4tb Corsair MP600 PRO XT
It honestly depends on the game. Some games have very unoptimized loading of assets, others really do stream massive files at 7GB/s. I would suggest using the Optane drive for your operating system and a few games and nand nvme for the rest.
At the end of the day gen 1 Optane cannot match the sustained read and write speeds of nand, but Optane is still Optane and nand cannot perform in mixed read/write workloads.
This, tbh i am not excited about optane for a game drive. Its much better as an os drive and mixed reads and writes for work. If you have way more ram than you need that can also feel like optane because everything gets cached in ram.
For me optane is cool because i can switch back a d forth between large projects all day which generates lots of random i/o. There is also a huge difference in the way cheap nand feels and a samsung 980 or 990 pro / solidigm p44 pro vs other nand for this kind of mixed optimizations on those specific products.
Hi Wendell,
I watched your video on the P5810 but noticed on techpowerup it is listed as using pSLC, I’ve never heard of it but my understanding it is QLC underneath but hides the other 3 cells and uses them for longevity
pseudo slc is pslc.
no one is making media that’s not certified for use as qlc, but that doesn’t mean you have to use the media in qlc mode. if that makes sense. this is also related to why its faster. reading back bits is technically initially an analog operation, and the a/d step is the most time consuming. only needing a single bit speeds that up quite a bit, and that’s what they’ve done here. so its not like they are running a normal 4 bit read and taking high/low in an “overclocked” media frequency but some actual different under the hood stuff happening. if I understand what they are sayinh.
obviously I haven’t looked at it under an electron microscope, but pSLC is normally 4-bit capable (technically analog) media plus machinery designed for 4 bit operations whereas this is 4-bit capable media plus different machinery designed specifically to operate in this way. is why its faster
Actually, I don’t think this is the case. At least, the fastest I’ve seen from a game was Waframe, briefly touching just under 2GB/s during startup, but mission loading is under 1GB/s.
Reading around about other peoples’ experiences, it’s pretty rare for a game to actually load assets at greater than 1GB/s, and when you think about it… That makes sense. Games aren’t loading a snapshot of a 14GB memory map state, so it’s a lot of looking at the file table, reading the header, and then reading out the chunks of data for a smaller file, typically only a few hundred megabytes at a time on the high side.
Optane actually seems to not improve load times in most games, too. Older games that already load pretty much instantly do benefit a bit, ie unreal tournament or quake, but the load times are already so fast it doesn’t matter, and newer games still have a fixed minimum load time, often of several seconds.
By the time videogame load times can genuinely meaningfully benefit from fast storage, they’re going to be optimized around high latency NAND flash in game consoles, and not for first gen optane storage, so it’ll actually likely perform worse more often than better.
Too bad CSGO is gone and dead, I’d love to bench the load times uncapped on that…
It’s good for two things, really: Boot/Root drive, and swap, both of which first gen optane does really well in. These use cases don’t often saturate the ~2.2GB/s peaks of first gen optane, but do benefit hugely from parallelization of read and write, and the very low latency, because the process is largely reading small chunks of data to figure out how and where to read the next chunks of small data.
OS boot is also shifting away from that model, too, leaning more into a hibernation style model, where the hardware spins up, and the OS checks if a hash of the current hardware config matches a hash stored in an image of the already-booted OS memory, and then loads that wholesale from the drive very quickly.
I remember that being a thing Microsoft was pushing towards and making progress on, at least.
Even the cheap 118G P1600x is shockingly snappy on Windows.
The Windows update progress bar doesn’t lie (or get stuck at 97%) anymore!
Thanks for this! Any idea if this works on win11 too, or jsut win10?