220 points by walterbell 4 days ago | 158 comments on HN
| Neutral High agreement (3 models)
Editorial · v3.7· 2026-03-15 22:20:32 0
Summary Technology & Commerce Neutral
This blog post evaluates Intel Optane memory technology and does not engage substantively with human rights themes or principles. The content is entirely technical and commercial in nature, focusing on product features and market positioning with no observable human rights advocacy, acknowledgment, or contradiction.
It stands out, because it didn't sell. Which is weird because there were some pretty big pros about using them. The latency for updating 1 byte was crazy good. Some databases or journals for something like zfs really benefited from this.
I feel sorry about the situation. From my perspective Optane was a godsend for databases. I was contemplating building a system. Could've been a pinnacle of vertical scalability for cheap.
Optane was crazy good tech, it way just too expensive at the time for mass adoption, but the benefits were so good.
Looking at those charts, besides the DWPD it feels like normal NVMe has mostly caught up. I occassionally wonder where a gen 7/8(?) optane would be today if it caught on, it'd probably be nuts.
My understanding is Optane is still unbeaten when it comes to latency. Has anyone examined its use as an OS volume, compared to today's leading SSD's? I know the throughput won't be as high, but in my experience that's not as important to how responsive your machine feels as latency.
Sure, they were expensive but they have great endurance and sustained read and write speeds. I use one in my car for camera recordings. I had gone through several other drives but this one has been going on 3 or 4 years now without issue. I have a couple more in use too. It's a shame this tech is going away because it's excellent.
Around the time of Optane's discontinuation, the rumor mill was saying that the real reason it got the axe was that it couldn't be shrunk any, so its costs would never go down. Does anyone know if that's true? I never heard anything solid, but it made a lot of sense given what we know about Optane's fab process.
And if no shrink was possible, is that because it was (a) possible but too hard; (b) known blocks to a die shrink; or (c) execs didn't want to pay to find out?
One potential application I briefly had hope for was really good power loss protection in front of a conventional Flash SSD. You only need a little compared to the overall SSD capacity to be able to correctly report the write was persisted, and it's always running, so there's less of a 'will PLP work when we really need it?' question. (Maybe there's some use as a read cache too? Host RAM's probably better for that, though.) It's going to be rewritten lots of times, but it's supposed to be ready for that.
It seems like there's a very small window, commercially, for new persistent memories. Flash throughput scales really cost-efficiently, and a lot is already built around dealing with the tens-of-microseconds latencies (or worse--networked block storage!). Read latencies you can cache your way out of, and writers can either accept commit latency or play it a little fast and loose (count a replicated write as safe enough or...just not be safe). You have to improve on Flash by enough to make it worth the leap while remaining cheaper than other approaches to the same problem, and you have to be confident enough in pulling it off to invest a ton up front. Not easy!
When most people are running databases on AWS RDS, or on ridiculous EBS drives with insanely low throughput and latency, it makes sense to me.
There are very few applications that benefit from such low latency, and if one has to go off the standard path of easy, but slow and expensive and automatically backup up, people will pick the ease.
Having the best technology performance is not enough to have product market fit. The execution required from the side of executives at Intel is far far beyond their capability. They developed a platform and wanted others to do the work of building all the applications. Without that starting killer app, there's not enough adoption to build an ecosystem.
It isn't weird at all. I would be surprised if it ever succeed in the first place.
Cost was way too high. Intel not sharing the tech with others other than Micron. Micron wasn't committed to it either, and since unused capacity at the Fab was paid by Intel regardless they dont care. No long term solution or strategy to bring cost down. Neither Intel or Micron have a vision on this. No one wanted another Intel only tech lock in. And despite the high price, it barely made any profits per unit compared to NAND and DRAM which was at the time making historic high profits. Once the NAND and DRAM cycle went down again cost / performance on Optane wasn't as attractive. Samsung even made some form of SLC NAND that performs similar to Optane but cheaper, and even they end up stopped developing for it due to lack of interest.
> besides the DWPD it feels like normal NVMe has mostly caught up.
So what you mean is that on the most important metric of them all for many workloads, Flash-based NVMe has not caught up at all. When you run a write heavy workload on storage with a limited DWPD (including heavy swapping from RAM) higher performance actually hurts your durability.
Intel did a spectacularly poor job with the ecosystem around the memory cells. They made two plays, and both were flops.
1. “Optane” in DIMM form factor. This targeted (I think) two markets. First, use as slower but cheaper and higher density volatile RAM. There was actual demand — various caching workloads, for example, wanted hundreds of GB or even multiple TB in one server, and Optane was a route to get there. But the machines and DIMMs never really became available. Then there was the idea of using Optane DIMMs as persistent storage. This was always tricky because the DDR interface wasn’t meant for this, and Intel also seems to have a lot of legacy tech in the way (their caching system and memory controller) and, for whatever reason, they seem to be barely capable of improving their own technology. They had multiple serious false starts in the space (a power-supply-early-warning scheme using NMI or MCE to idle the system, a horrible platform-specific register to poke to ask the memory controller to kindly flush itself, and the stillborn PCOMMIT instruction).
2. Very nice NVMe devices. I think this was more of a failure of marketing. If they had marketed a line of SSDs that, coupled with an appropriate filesystem, could give 99% fsync latency of 5 microseconds and they had marketed this, I bet people would have paid. But they did nothing of the sort — instead they just threw around the term “Optane” inconsistently.
These days one could build a PCM-backed CXL-connected memory mapped drive, and the performance might be awesome. Heck, I bet it wouldn’t be too hard to get a GPU to stream weights directly off such a device at NVLink-like speeds. Maybe Intel should try it.
Before people claim it doesn't matter due to OS write buffering, I should point out a) today's bloated software and the many-layered, abstracted I/O stack it's built on tends to issue lots of unnecessary flushes, b) read latency is just as important as write (if not moreso) to how responsive your OS feels, particularly if the whole thing doesn't fit in (or preload to) memory.
I configured a hetzner ax101 bare metal server with a 480GB 3d xpoint ssd some years ago. It’s used as the boot volume and it seems fast despite the server being heavily over provisioned, but I can’t really compare because I don’t have a baseline without.
The actual strength of Optane was on mixed workloads. It's hard to write a flash cell (read-erase-write cycle, higher program voltage, settling time, et cetera). Optane didn't have any of that baggage.
This showed up as amazing numbers on a 50%-read, 50%-write mix. Which, guess what, a lot of real workloads have, but benchmarks don't often cover well. This is why it's a great OS boot drive: there's so much cruddy logging going on (writes) at the same time as reads to actually load the OS. So Optane was king there.
That's at least physically half-plausible, but it would be a terrible reason if true. 3.5 in. format hard drives can't be shrunk any, and their costs are correspondingly high, but they still sell - newer versions of NVMe even provide support for them. Same for LTO tape cartridges. Perhaps they expected other persistent-memory technologies to ultimately do better, but we haven't really seen this.
Worth noting though that Optane is also power-hungry for writes compared to NAND. Even when it was current, people noticed this. It's a blocker for many otherwise-plausible use cases, especially re: modern large-scale AI where power is a key consideration.
> Has anyone examined its use as an OS volume, compared to today's leading SSD's?
Late last year I switched from a 1.5tb Optane 905P to a 4tb WD Blue SN5000 NVMe drive in a gaming machine and saw improved load times, which makes sense given the read and write speeds are ~double. No observable difference otherwise.
I'm sure that's not the use case you were looking for. I could probably tease out the difference in latency with benchmarks but that's not how I use the computer.
The 905P is now in service as an SSD cache for a large media server and that came with a big performance boost but the baseline I'm comparing to is just spinning drives.
That's data retention issues on the very first read-through of the media after sitting in cold storage for many years, with subsequent performance returning to normal. It's definitely something to be aware of (and kudos to the blog poster for running that experiment) but worn-out NAND will behave a lot worse than that.
> It seems like there's a very small window, commercially, for new persistent memories. Flash throughput scales really cost-efficiently
Flash is no bueno for write-heavy workloads, and the random-access R/W performance is meh compared to Optane. MLC and SLC have better durability and performance, but still very mid.
Any decent SSD has capacitor (enterprise) or battery backed (phones) DRAM. Therefore, a sync write is just “copy the data to an I/O buffer over PCIe”.
For databases, where you do lots of small scattered writes, and lots of small overwrites to the tail of the log, modern SSDs coalesce writes in that buffer, greatly reducing write wear, and allowing the effective write bandwidth to exceed the media write bandwidth.
These schemes are much less expensive than optane.
> One potential application I briefly had hope for was really good power loss protection in front of a conventional Flash SSD.
That was never going to work out. Adding an entirely new kind of memory to your storage stack was never going to be easier or cheaper than adding a few large capacitors to the drive so it could save the contents of the DRAM that the SSD still needed whether or not there was Optane in the picture.
I think it was killed primarily because the DIMM version had a terrible programming API. There was no way to pin a cache line, update it and flush, so no existing database buffer pool algorithms were compatible with it. Some academic work tried to address this, but I don’t know of any products.
The SSD form factor wasn’t any faster at writes than NAND + capacitor-backed power loss protection. The read path was faster, but only in time to first byte. NAND had comparable / better throughput. I forget where the cutoff was, but I think it was less than 4-16KB, which are typical database read sizes.
So, the DIMMs were unprogrammable, and the SSDs had a “sometimes faster, but it depends” performance story.
I never understood what they're meant to do. Intel seemed to picture some future where RAM is persistent; but they were never close to fast enough to replace RAM, and the option to reboot in order to fix some weird state your system has gotten itself into is a feature of computers, not a problem to work around.
I run two 1.5TB Optanes in raid-0 with XFS (I picked them up for $300 each on sale about two years ago). These are limited to PCIE 3.0 x4 (about 4GB/s max each). I also have a 64GB optane drive I use as my boot drive.
It's hard to tell you, because it's subjective, I don't swap back and forth between an SSD and the optane drives. I have my old system, which has a 2TB Samsung 980 Pro NVME drive (PCIE 4.0 x4, or 8GB/s max) as root, and a Sabrent rocket 4 plus 4TB drive secondary (also PCIE 4.0), so I ran sysbench on both systems, so I could share the differences. (Old system 5950X, new system 9950X3D).
It feels snappier, especially when doing compilations...
Sequential reads:
I started with a 150GB fileset, but it was being served by the kernel cache on my newer system (256GB RAM vs 128GB on the old), so I switched to use 300GB of data, and the optanes gave me 5000 MiB/s for sequential read as opposed to 2800 MiB/s for the 980 Pro, and 4340 MiB/s for the Rocket 4 Plus.
Random writes alone (no read workload)
The optane system gets 2184 MiB/s, the 980 Pro gets 32 MiB/s, and the Rocket 4 Plus gets 53 MiB/s.
Mixed workload (random read/write)
The optanes get 725/483 as opposed to 9/6 for the 980 Pro, and 42/28 for the Rocket 4 Plus.
2x1.5TB Optane Raid0:
Prep time:
`sysbench fileio --file-total-size=150G prepare`
161061273600 bytes written in 50.41 seconds (3047.27 MiB/sec).
Benchmark:
`sysbench fileio --file-total-size=150G --file-test-mode=rndrw --max-time=60 --max-requests=0 run`
WARNING: --max-time is deprecated, use --time instead
sysbench 1.0.20 (using system LuaJIT 2.1.1741730670)
Running the test with following options:
Number of threads: 1
Initializing random number generator from current time
Extra file open flags: (none)
128 files, 1.1719GiB each
150GiB total file size
Block size 16KiB
Number of IO requests: 0
Read/Write ratio for combined random IO test: 1.50
Periodic FSYNC enabled, calling fsync() each 100 requests.
Calling fsync() at the end of test, Enabled.
Using synchronous I/O mode
Doing random r/w test
Initializing worker threads...
Threads started!
File operations:
reads/s: 46421.95
writes/s: 30947.96
fsyncs/s: 99034.84
Throughput:
read, MiB/s: 725.34
written, MiB/s: 483.56
General statistics:
total time: 60.0005s
total number of events: 10584397
Latency (ms):
min: 0.00
avg: 0.01
max: 1.32
95th percentile: 0.03
sum: 58687.09
Threads fairness:
events (avg/stddev): 10584397.0000/0.00
execution time (avg/stddev): 58.6871/0.00
2TB Nand Samsung 980 Pro:
Prep time:
`sysbench fileio --file-total-size=150G prepare`
161061273600 bytes written in 87.15 seconds (1762.53 MiB/sec).
Benchmark:
`sysbench fileio --file-total-size=150G --file-test-mode=rndrw --max-time=60 --max-requests=0 run`
WARNING: --max-time is deprecated, use --time instead
sysbench 1.0.20 (using system LuaJIT 2.1.1741730670)
Running the test with following options:
Number of threads: 1
Initializing random number generator from current time
Extra file open flags: (none)
128 files, 1.1719GiB each
150GiB total file size
Block size 16KiB
Number of IO requests: 0
Read/Write ratio for combined random IO test: 1.50
Periodic FSYNC enabled, calling fsync() each 100 requests.
Calling fsync() at the end of test, Enabled.
Using synchronous I/O mode
Doing random r/w test
Initializing worker threads...
Threads started!
File operations:
reads/s: 594.34
writes/s: 396.23
fsyncs/s: 1268.87
Throughput:
read, MiB/s: 9.29
written, MiB/s: 6.19
General statistics:
total time: 60.0662s
total number of events: 135589
Latency (ms):
min: 0.00
avg: 0.44
max: 15.35
95th percentile: 1.73
sum: 59972.76
Threads fairness:
events (avg/stddev): 135589.0000/0.00
execution time (avg/stddev): 59.9728/0.00
4TB Sabrent Rocket 4 Plus:
Prep time:
`sysbench fileio --file-total-size=300G prepare`
322122547200 bytes written in 152.39 seconds (2015.92 MiB/sec).
Benchmark:
`sysbench fileio --file-total-size=300G --file-test-mode=rndrw --max-time=60 --max-requests=0 run`
WARNING: --max-time is deprecated, use --time instead
sysbench 1.0.20 (using system LuaJIT 2.1.1741730670)
Running the test with following options:
Number of threads: 1
Initializing random number generator from current time
Extra file open flags: (none)
128 files, 2.3438GiB each
300GiB total file size
Block size 16KiB
Number of IO requests: 0
Read/Write ratio for combined random IO test: 1.50
Periodic FSYNC enabled, calling fsync() each 100 requests.
Calling fsync() at the end of test, Enabled.
Using synchronous I/O mode
Doing random r/w test
Initializing worker threads...
Threads started!
File operations:
reads/s: 2690.28
writes/s: 1793.52
fsyncs/s: 5740.92
Throughput:
read, MiB/s: 42.04
written, MiB/s: 28.02
General statistics:
total time: 60.0155s
total number of events: 613520
Latency (ms):
min: 0.00
avg: 0.10
max: 8.22
95th percentile: 0.32
sum: 59887.69
Threads fairness:
events (avg/stddev): 613520.0000/0.00
execution time (avg/stddev): 59.8877/0.00
Title frames Intel Optane as 'stand out,' using positive framing language to promote product superiority without comparative evidence presented in visible content.