Sometimes, things are just cool.

> what do I get for it? Persistence?

Density. PCM is closer in bits-per-unit-area to Flash than to DRAM.

Yeah but that doesn't help if it's too slow to use as DRAM

It's another level in the hierarchy. DRAM is too slow to use as SRAM, but it's denser, so it has a place. Flash is too slow to use as DRAM (or PCM), but it's denser, so it has a place. Spinning rust. Tapes. Adding another layer between DRAM and Flash to take advantage of its unique trade-offs is not only reasonable, it's exactly what we always do.

It somewhat (maybe) had sense for HDD caching but I don't think HDDs in 2026 are used in cases where separate cache is needed.

Having a very low latency, durable write destination is a superpower for lots of database architectures (or things that share architectural details with databases, like some filesystems).

Optane had much lower latency than other types of durable storage, and also had very good endurance compared to NAND.

At a prior employer, we made extensive use of them as ZFS SLOG drives for large database servers. They made fsync very inexpensive.

Yep. I use a P4800X for ZFS special vdev. It's a monster. The thing has write endurance for many petabytes of write. The queue depth alone means sustained reads and writes are... sustained. Try that with a normal SSD/NVME.

And instead of upgrading latency from NVMe SSDs to Optane, we downgraded to EBS.

Yeah that's the fast persistent cache layer more or less

Huh. I wrote a whole thing here about serving a niche in serving local LLMs, but turns out, any niche it could have served is probably better served (per dollar) by either more RAM or modern SSDs.

There's an argument to be made for the persistent memory being a cheaper substitute for more RAM, but imo the gap isn't big enough to justify it when considered as a % of the rest of the system you're building around it and the performance tradeoff.

Important reminder to check your biases. I was sure Optane had missed its mark by dropping out just before LLMs became a primary use case for high $ systems.

There was a version of optane (optane DC?) that plugged into DIMM slots and cost less than RAM, with possible higher capacities. That would be a useful tradeoff had it kept up the advantage.

That's the persistent memory modules; they were roughly half the price per GB compared to DDR4 from my quick research, but substantially higher memory latency, which would significantly affect usage as a KV cache, which would probably be their primary use case since SSDs beat it in raw data throughput. Who knows how it would have developed over the years since it was discontinued, but my gut reaction to the numbers said you'd rather just invest in more RAM based on the numbers I saw.

Intel had multiple false starts on getting their platform ready for these devices to be used in a sensible manner. I have one of these false starts in a closet somewhere, and I did some kernel dev work for it. It was an incredible mess.

(I didn't actually have an Intel DIMM module -- I had a different vendor's. But fundamentally there needed to be a clean way to get data from software into the durable memory domain, and there wasn't.)

You can have very large MoE models run on small amounts of ram (or vram), if you can transfer the experts you need on the fly from ram to disk.

SSDs are good, but an improved Optane would've been better.

Optane was cheaper than DRAM per bit.

If Intel had volume production of Optane and had been able to quickly reconfigure it to produce a device that could stream out model weights fast enough for inference or could have gotten a read-write device that could hold KV caches effectively, they could have made an absolute killing in the current market.

HBF is slowly going after the former market. I'm not sure flash has a great answer for the latter due to endurance issues.

I can't think of any company that is as good as Intel at having a technology that's ready or almost ready at or before the time that the market is ready for and then completely failing to sell the product into the newly developed market.

Google?

>Optane was cheaper than DRAM per bit.

Optane was cheaper than DRAM per bit when DRAM was expensive. It price / performance didn't make sense as soon as the DRAM cycle took its dive and DRAM price became cheaper / normal / losing money depending on how or who you see and asked.

IIRC when it came out it was faster than the SSDs you could buy. It never really got that much better and then SSD mostly caught up.

More like they stopped investing in improving that tech so flash eventually caught up on throughput. It didn't catch up on latency and endurance has never been anywhere near a competition.

The other answers are hinting at it, but Optane was fundamentally much much lower latency than NAND flash used in SSDs.

SSDs use lots of tricks to hide this latency (using RAM). However, when the data is super critical, you can't just throw some data in RAM and hope that it gets written out.

Enterprise flash drives have battery backups and keep the system alive for a moment to flush out the writes, but Optane doesn't need special tricks to get this sort of durability, it just has it.

If you have a large stream of critical data flowing in that has to be stored and can't be lost- Optane suddenly looks wayyyy better than SSDs

It's mostly not competitive by virtue of being old, but for specific use cases it still stomps the latest Gen5 nvme drives. Specifically on things that need low latency and fast QD1 performance.

Think metadata in a ZFS array or database writes.

Also crazy endurance - often an order of magnitude beyond even enterprise nvmes.

...and because the things are basically un-killable thanks to endurance you can safely pick up 2nd hand used ones.

So they're a neat trick if used in precisely the right application. Hence:

>some fraction of technologists

>would've been awesome if only Intel hadn't killed it.

Maybe - it would presumably have improved further for niche cases, but doubt it would have conquered mainstream markets.

Very low latency, extreme durability, super high IOPS. It's SLC without the problems of SLC. The problem is it was a bit early, the interfaces never were fast enough (x4 3.0), and software that could have benefitted from massive amounts of it greatly, came out too late for it.

The actual problem is Intel was in such a bad position, they needed to shed weight, and Optane was not quick money.

In a DIMM slot though?

Optane had multiple form factors and could have adopted modern connections.

Optane was available as PCIe cards and NVME. I have one of the PCIE 1.2TB drives as my OS drive.

Seems halfway between SSD and RAM just like SSD is halfway between RAM and HDD. There's no fixed number of tiers in the storage hierarchy.