myne wrote on 2026-08-09, 07:23:

Trashbytes wrote on 2026-08-09, 06:18:

myne wrote on 2026-08-09, 06:13:

They're already using HBM for laptop SOCs. First, battery life. That'll bleed into desktop at the ultra low end non-upgradable. […]
Show full quote

They're already using HBM for laptop SOCs.
First, battery life.
That'll bleed into desktop at the ultra low end non-upgradable.
Then server... possibly low end, possibly high. Not sure.
Server will be the first with HBM+DDR. HBM will be the first level ram, and have DDRx on the side.
Which will move into HEDT as some supermegaultracore.
And finally midrange desktops.

Why?
Cache is mad expensive.
Ram isn't fast enough.
DDR isn't close enough or low power enough.
HBM will allow 90% of the effective performance and power of an L3, and relatively slower DDR will compliment it for pure size.

Agree mostly with the exception of HBM on SOC for servers . .that wont happen, itll be HBM on CAMM2 mostly for scalability and maintenance reasons.

Eh. I just can't see 8 sticks of CAMM2 per socket. Which is what the server guys always want/expect regardless of whether they need it.
Plus, DDR is likely to remain cheaper in bulk forever since you can validate each chip individually. HBM has to be assembled and then tested.

Well it wouldn't be 8 sticks since CAMM2 is a LGA type socket, itll be a totally different beast than what you would normally have in a server. HBM memory dies can also be tested prior and during assembly, they have been able to do that for a while now with the newer versions of it, its still more prone to failures with the TSVs and final packaging but even that is far better than previously.

As for DDR remaining cheaper ...they are mass producing HBM then DDR isn't being produced at all or in verry small quantities since you can only fabricate one type of memory on the machines at a time (This is hard reality not some fact pulled out of my behind). So they would be using the remaining stock of DDR they have currently which is already hitting stupid prices for ECC server grade memory, 20k+ USD for 256GB isn't odd to see right now.

I know you have a thing for DDR because of how good it is for DIY stuff but the reality is that right now they are producing very little of it and HBM is sold out till 2028 at this point so the DDR production situation is not changing any time soon and by the time it does itll be moving to DDR6 CAMM2 and not DDR4/5 so itll still remain expensive. (Consumers wont see DDR6 till 2030 and enterprise till late 2029 if the forecast is right)

I don't have a thing for ddr.
I personally don't care that much.
But big fat server people will be wanting a terabyte minimum by now.
8 sticks of 128gb is definitely going to be easier to produce than one single perfect 1tb camm2.
What they sacrifice in raw unit speed, they more than gain back with 4/8 channel.
It's just economics.
The more you trash, the more each unit costs. One dud ddr chip in that scenario is a loss of 16gb (assuming 8/stick). One dud hbm chip (assuming 1chip =1 stick worth) is 128gb. Plus the much more involved packaging process.

Yields always improve, but the complexity certainly means that it will always be worse.

That said, we might see a return to the days of "known bad sectors" where either the chips are over provisioned and remapped or the spd records the factory dead addresses. It's not a showstopper for most people to lose a few mb if it's cheaper.

I'm not aware of the issues with the lines. I understood hbm to fundamentally be ddr with a controller and package change.
Ie, the silicon is fundamentally the same.
Running 2 packaging lines doesn't seem like an impossible problem to solve.

myne wrote on 2026-08-09, 10:27:

I don't have a thing for ddr. I personally don't care that much. But big fat server people will be wanting a terabyte minimum by […]
Show full quote

I don't have a thing for ddr.
I personally don't care that much.
But big fat server people will be wanting a terabyte minimum by now.
8 sticks of 128gb is definitely going to be easier to produce than one single perfect 1tb camm2.
What they sacrifice in raw unit speed, they more than gain back with 4/8 channel.
It's just economics.
The more you trash, the more each unit costs. One dud ddr chip in that scenario is a loss of 16gb (assuming 8/stick). One dud hbm chip (assuming 1chip =1 stick worth) is 128gb. Plus the much more involved packaging process.

Yields always improve, but the complexity certainly means that it will always be worse.

That said, we might see a return to the days of "known bad sectors" where either the chips are over provisioned and remapped or the spd records the factory dead addresses. It's not a showstopper for most people to lose a few mb if it's cheaper.

I'm not aware of the issues with the lines. I understood hbm to fundamentally be ddr with a controller and package change.
Ie, the silicon is fundamentally the same.
Running 2 packaging lines doesn't seem like an impossible problem to solve.

HBM and DDR both use the same silicon wafer, TSMC makes that so supply is limited to what they can produce and allocate, TSMC is also the company that does CoWoS which no other company is able to do this also restricts HBM output. So yes they are both DRAM but the core processes used by both are different enough the fabs need to dedicate capacity to one or the other. (Yes I'm sure profitability is also involved here along with market manipulation and fixing to keep HBM prices high. I don't believe them at face value either, I'm sure they could find a way to handle both)

I too don't understand why the fabs wont throw billions at a new fabrication line for DDR ..but I suspect the raw wafer bottle neck at TSMC may be the reason. (We really do need another TSMC like company, we have all our eggs in their basket and its worrying)

There is more to this server equation than just capacity .. just looking at HBM4 ..its 32 channel memory with 2 pseudo channels per channel on a 2048bit bus with the capability of operating at 0.7 of a volt so what you want really depends on what you are using that server for. (At scale I suspect memory capacity per server is far less of an issue than operating voltages and heat would be)

DDR has a place I just don't think its as big as it used to be, HBM has so many more advantages over DDR I can understand why they want to move the industry away from DDR. DDR is also hitting the same Thermodynamic wall SSDs have hit, DDR5 already gets very hot which isn't helping memory reliability or data integrity and DDR6 is looking to be even faster and hotter than DDR5 .. so how long before memory requires its own heatsink and fan or AIO to prevent it throttling? (I can see the AIBs already salivating at the thought of being able to sell another bit of RGB cruft to the consumers)

Just so you know I'm enjoying this discussion !

Tsmc?
Didn't think they made ram.
Are you sure about that.
I think you might be thinking the controller.
Either way, micron, hynix and Samsung will be making what's selling.
Transitions are rarely without overlap.
There are clear benefits to both HBM and SOC but there are also drawbacks and it is relatively trivial to address the drawbacks by keeping Ddrx alongside it.

Plus, it gives Intel/amd another level of complexity to baffle customers with more SKUs than your average 711.
Was that the I7 hfdjjgdthb56uo8hb? Or the I7 hfdjjgdthb46uo8ht?
Because if it is the I7 hfdjjgdthb56uo8hb then you only have 4gb hbm at half speed but if it's the hfdjjgdthb56uo8ht then you have 16gb at full speed. You definitely want the hfdjjgdthb46uo8ht.

TSMC only makes specialized integrated memory for the products that it is contracted to fab.

myne wrote on 2026-08-09, 13:02:

Tsmc? Didn't think they made ram. Are you sure about that. I think you might be thinking the controller. Either way, micron, hyn […]
Show full quote

Tsmc?
Didn't think they made ram.
Are you sure about that.
I think you might be thinking the controller.
Either way, micron, hynix and Samsung will be making what's selling.
Transitions are rarely without overlap.
There are clear benefits to both HBM and SOC but there are also drawbacks and it is relatively trivial to address the drawbacks by keeping Ddrx alongside it.

Plus, it gives Intel/amd another level of complexity to baffle customers with more SKUs than your average 711.
Was that the I7 hfdjjgdthb56uo8hb? Or the I7 hfdjjgdthb46uo8ht?
Because if it is the I7 hfdjjgdthb56uo8hb then you only have 4gb hbm at half speed but if it's the hfdjjgdthb56uo8ht then you have 16gb at full speed. You definitely want the hfdjjgdthb46uo8ht.

I think you are confusing a blank silicon wafer with an actual finished one which are two different things.

TSMC makes the blank base die for HBM, they don't do any fabrication of DRAM themselves, you can look this up with little issue. So the main bottle neck is how many wafers TSMC can spit out to be allocated to the fabricators who then turn it into memory at their own foundries and EUV labs. I'm sure they source from other wafer fabs also, but I doubt any have the capacity that TSMC does or the quality.

TSMC does far more than just make chips ..they are one of the largest makers of silicon wafers for fabricators.

Looking up Optane drives and they are extremely expensive.

Summary vs good NVMe:
Optane has:
1. Higher sustained sequential writes AFTER thr NVMe drive has filled the cache
2. Much higher 4k random access and latency.
3. Way better endurance when compared to a lot of consumer based drives. When compared to other enterprise SSDs of the same size, you are looking at up to double the endurance (10wpd vs 5wpd).
4. Way higher price

NVMe
1. Way higher sequential writes before cache is filled.
2. Much lower price
3. Endurance on good drives is generally going to be way higher than what a normal user could use up in 10+years. This is dependent on the amount of spare space the drive has.
4. Much cheaper.
5. Cacheless drives are trash.

No matter the drive you are using, you can get lower latency by using a ram cache software like Primocache. Some board makers, such as ASUS have specific solutions for their motherboards. ASUS is RAMCache/RAMCache II

I've never felt the need to spend the insane amount for an Optane drive.

It would have gotten more market share had it not been priced to the moon in th first place.

Trashbytes wrote on 2026-08-09, 13:32:

I think you are confusing a blank silicon wafer with an actual finished one which are two different things. […]
Show full quote

myne wrote on 2026-08-09, 13:02:

Tsmc? Didn't think they made ram. Are you sure about that. I think you might be thinking the controller. Either way, micron, hyn […]
Show full quote

Tsmc?
Didn't think they made ram.
Are you sure about that.
I think you might be thinking the controller.
Either way, micron, hynix and Samsung will be making what's selling.
Transitions are rarely without overlap.
There are clear benefits to both HBM and SOC but there are also drawbacks and it is relatively trivial to address the drawbacks by keeping Ddrx alongside it.

Plus, it gives Intel/amd another level of complexity to baffle customers with more SKUs than your average 711.
Was that the I7 hfdjjgdthb56uo8hb? Or the I7 hfdjjgdthb46uo8ht?
Because if it is the I7 hfdjjgdthb56uo8hb then you only have 4gb hbm at half speed but if it's the hfdjjgdthb56uo8ht then you have 16gb at full speed. You definitely want the hfdjjgdthb46uo8ht.

I think you are confusing a blank silicon wafer with an actual finished one which are two different things.

TSMC makes the blank base die for HBM, they don't do any fabrication of DRAM themselves, you can look this up with little issue. So the main bottle neck is how many wafers TSMC can spit out to be allocated to the fabricators who then turn it into memory at their own foundries and EUV labs. I'm sure they source from other wafer fabs also, but I doubt any have the capacity that TSMC does or the quality.

TSMC does far more than just make chips ..they are one of the largest makers of silicon wafers for fabricators.

Plenty of companies make wafers. What's unique about theirs?

You guys seem to be mixing up the M.2 drives and the U.2 drives into the same answers and thinking they cross over, which they dont, not in the way the answer given implies.

ElectroSoldier wrote on 2026-08-10, 07:30:

You guys seem to be mixing up the M.2 drives and the U.2 drives into the same answers and thinking they cross over, which they dont, not in the way the answer given implies.

Are you suggesting they are not both NVME based NAND drives using slightly different socket formats? Or that their sockets are not cross compatible ? perhaps its that they both use different Nand technologies or different controllers, what is it you think we are implying here ?

myne wrote on 2026-08-10, 00:27:

Trashbytes wrote on 2026-08-09, 13:32:

I think you are confusing a blank silicon wafer with an actual finished one which are two different things. […]
Show full quote

myne wrote on 2026-08-09, 13:02:

Tsmc? Didn't think they made ram. Are you sure about that. I think you might be thinking the controller. Either way, micron, hyn […]
Show full quote

Tsmc?
Didn't think they made ram.
Are you sure about that.
I think you might be thinking the controller.
Either way, micron, hynix and Samsung will be making what's selling.
Transitions are rarely without overlap.
There are clear benefits to both HBM and SOC but there are also drawbacks and it is relatively trivial to address the drawbacks by keeping Ddrx alongside it.

Plus, it gives Intel/amd another level of complexity to baffle customers with more SKUs than your average 711.
Was that the I7 hfdjjgdthb56uo8hb? Or the I7 hfdjjgdthb46uo8ht?
Because if it is the I7 hfdjjgdthb56uo8hb then you only have 4gb hbm at half speed but if it's the hfdjjgdthb56uo8ht then you have 16gb at full speed. You definitely want the hfdjjgdthb46uo8ht.

I think you are confusing a blank silicon wafer with an actual finished one which are two different things.

TSMC makes the blank base die for HBM, they don't do any fabrication of DRAM themselves, you can look this up with little issue. So the main bottle neck is how many wafers TSMC can spit out to be allocated to the fabricators who then turn it into memory at their own foundries and EUV labs. I'm sure they source from other wafer fabs also, but I doubt any have the capacity that TSMC does or the quality.

TSMC does far more than just make chips ..they are one of the largest makers of silicon wafers for fabricators.

Plenty of companies make wafers. What's unique about theirs?

Not all silicon wafer blanks are the same, the ones I'm referring to are made specifically for HBM fabrication at Samsung, SK Hynix and Micron. Rather than taking my word for it you can go look it up.

As for what's unique, likely not much outside of quality and certification, but my feeling is that the Big three would want to use the best quality wafers they can source for HBM fabrication since as you suggest its a difficult process and using any old wafer will likely not produce the output they require.

Again don't take my word for it, go look it up I may have gotten my wires crossed when I was digging around for information on HBM4e and DDR6.

cyclone3d wrote on 2026-08-09, 15:36:

Looking up Optane drives and they are extremely expensive. […]
Show full quote

Looking up Optane drives and they are extremely expensive.

Summary vs good NVMe:
Optane has:
1. Higher sustained sequential writes AFTER thr NVMe drive has filled the cache
2. Much higher 4k random access and latency.
3. Way better endurance when compared to a lot of consumer based drives. When compared to other enterprise SSDs of the same size, you are looking at up to double the endurance (10wpd vs 5wpd).
4. Way higher price

NVMe
1. Way higher sequential writes before cache is filled.
2. Much lower price
3. Endurance on good drives is generally going to be way higher than what a normal user could use up in 10+years. This is dependent on the amount of spare space the drive has.
4. Much cheaper.
5. Cacheless drives are trash.

No matter the drive you are using, you can get lower latency by using a ram cache software like Primocache. Some board makers, such as ASUS have specific solutions for their motherboards. ASUS is RAMCache/RAMCache II

I've never felt the need to spend the insane amount for an Optane drive.

It would have gotten more market share had it not been priced to the moon in th first place.

Cacheless drives are trash for PCIe 3 yes, PCE 4 and 5 cacheless drives using HMB on DDR5 systems are nearly as performant as cached drives. You would be very hard pressed to tell the difference between them in normal usage.

I have a few PCIe 5 cacheless drives with HMB and they are ok in my book.

Trashbytes wrote on 2026-08-10, 08:15:

ElectroSoldier wrote on 2026-08-10, 07:30:

You guys seem to be mixing up the M.2 drives and the U.2 drives into the same answers and thinking they cross over, which they dont, not in the way the answer given implies.

Are you suggesting they are not both NVME based NAND drives using slightly different socket formats? Or that their sockets are not cross compatible ? perhaps its that they both use different Nand technologies or different controllers, what is it you think we are implying here ?

I love these kinds of answers. They gaslight your last answer and shut down any kind of replay you might want to make in one go by claiming authority status before you even start...
Ill bet you a "Linux Guru" too right?

Intel Optane was a rabbit hole I went down a while ago, back when it was relevant and there were a lot of snare traps along the way mostly from people who come out with answers like yours. Or read all the papers on the internet but had no practical experience of using it and so kept leading me down dead ends.

But to give a clean answer that doesnt get me told off.

No, Im not suggesting the sockets arent NVME or that they arent electrically compatible.
The point is that Optane M.2 Memory modules and Optane U.2/PCIe SSDs are completely different products with different requirements, different controllers, and completely different underlying storage technologies, so answers about one do not apply to the other.
This is where my previous answer is built from, because the post before it makes several assumptions which might not be either technically correct or possible to do in the real world.
The post before states a load of speeds that look great especially in the world where SSDs are very expensive and an H10 or H20 Optane SSD can be had for £100 (I mention them because people who read these threads see theyre Optane and assume there will be some speed benefit to them). Which it true, but what it leaves out is that a system after Intel gen 10 CPUs will only really use the NAND portion of the drive. The Optane portion simply isnt supported anymore. So yes, the numbers look good on paper, but only on platforms that actually support Optane Memory acceleration. On newer systems, youre effectively buying a QLC SSD with a chunk of Optane you cant use.

For instance an i7 7700 might support Optane if in a 300 series chipset board, but if its in a 100 series chipset board it doesnt. But then you can take that 300 series chipset board and put a i7 6700 CPU in it and it doesnt support Optane any more. Even though its the exact same board, in fact its the same system.
I have both. An HP EliteDesk 800 Mini G4. Both identical apart from one has an I7 6700 and the other has an i7 7700. One cant use Optane the other can. OR more accuratly based on what people think, and going by your post I include you. They can both use the NVME M.2 ssd. They just cant both use the Optane portion of it.

To break it down clearly:

1. Optane M.2 “Memory” modules (16/32/64GB)
These are not NAND drives at all.

They are 3D XPoint caching devices that rely on Intel RST and require:
7th, 8th or 9th gen Intel Core CPUs (Note Core, Pentium and Celeron is not supported even if the right gen)
200‑series / 300‑series chipsets (B250, H270, Z270, B360, H370, Z370, Z390, Q2370 etc... However not all motherboards with those chipsets have it supported in the BIOS, OEMs cut features)
BIOS support for Optane Memory mode (See above, this is important, you might have a supporting chipset that doesnt have Optane support or has Optane support but only in certain BIOS revisions)

They do not behave like normal NVMe SSDs. They cannot be used as standalone storage devices. This is the product people often confuse with “Optane SSDs”.

They are NOT NAND they are 3D XPoint.
Some, H10, H20, come with NAND glued to the Optane yes but the Optane isnt NAND its 3D XPoint.

2. Optane U.2 / PCIe SSDs (900p, 905p, P4800X, P5800X)
These are NVMe SSDs, but again not NAND — they use pure 3D XPoint media.

They do not require any special CPU or chipset.

They work on any NVME capable system, Intel or AMD.

They have completely different controllers, firmware, endurance characteristics, and latency behaviour compared to NAND NVMe drives.
Going back to your gaslighting quoted post.
im not "suggesting they are not both NVME based NAND drives" Im out n out saying they are not NAND they are 3D XPoint.

3. Why the distinction matters
When people say “Optane only works on certain CPUs/chipsets”, they are talking only about the M.2 Optane Memory modules.

When people talk about Optanes performance advantages (latency, endurance, SLOG use, etc.), they are talking about the U.2 / PCIe Optane SSDs, which do not have the same platform restrictions.

So yes — both use NVMe as the protocol, but:

They are not both NAND

They do not share controllers

They do not share platform requirements

They do not behave the same way

They were designed for completely different purposes

That’s the crossover confusion I was pointing out. Because when a thread like you mentions Optane, even when there is a picture of the drive people assume all Optane operates the same but it doesnt.

I was merely referring to the management of that particular layer in software.
OS support for optane implies that it is trivial to tweak the same architectural considerations for multiple performance levels of RAM.
The kernel knows how to shuffle things around for decent performance.

Loosely speaking the architecture is this:
Core> L1 > L2 > L3 (if applicable) > RAM > Optane > disk.

The hypothesised future is this:
Core> L1 > L2 > L3 (if applicable) > HBM > DDR > disk.

From a high level they're practically identical. Existing OS kernels know how to deal with one, so they will have no problems dealing with the other.
That said, I'm more referring to the server side where Optane was on DIMMs and was used as slower RAM.
The consumer level "big disk cache/swap" is... well, not the same, but not that different.
It would splitting hairs to say they're that different.
It doesn't really matter that much whether the kernel sees it as a swap file on a fast drive, or as slower ram.
Pohtayto, P'tahtoh

Again you seem to be mixing things up on the fly. I think.

My comment was regarding a post that showed a CrystalDiskMark score for Optane SSDs.

It looks like you’re talking about Optane DC Persistent Memory. The server class DIMMs that sit on the memory bus and behave like RAM. Those did require OS support, NUMA awareness, DAX mode, PMEM APIs, etc.

What I’m talking about here is Optane Memory (M10 or M15) and the Optane portion of H10 or H20 hybrid SSDs. These don’t behave like RAM at all. They’re just storage side caching layers managed entirely by Intel RST + BIOS, and the OS never sees or manages the Optane portion directly.

So the memory tiering hierarchy you’re describing applies to Optane DIMMs, not to the consumer Optane Memory products in M.2/U.2 form. They’re completely different technologies with different requirements and different behaviour.

The reason why I said anything is that people read this and dont say anything, they go off buy this stuff because its Optane and think it will work not realising they bought something that will never work in their system.

As to what the future might hold I dont know, it wasnt a part of the conversation I was getting into, Ill let you get on with that one.
I just couldnt read what I did and say nothing because I know with prices being what they are people see the Optane drives out there in M.2 format and think it will work, when in actual fact they will get a QLC NAND disk and no access to the Optane glued to it because the OS and BIOS no longer supports it.

A really really bad one to talk about without explanation is Optane DIMMs or Optane Persistent memory, because its DDR4 and really stupidly cheap now... a 256Gb stick for £199!!!

No. I'm not mixing it up.
I'm noting that the architectural changes required for OS compatibility with 2 tier RAM are already complete.
That's it.
Until the last post I never mentioned the connection type. It's not really relevant.
Both Windows and Linux are already aware enough of the general concept that at worst it would be configuration tweaks.

ElectroSoldier wrote on 2026-08-10, 12:33:

I love these kinds of answers. They gaslight your last answer and shut down any kind of replay you might want to make in one go […]
Show full quote

Trashbytes wrote on 2026-08-10, 08:15:

ElectroSoldier wrote on 2026-08-10, 07:30:

You guys seem to be mixing up the M.2 drives and the U.2 drives into the same answers and thinking they cross over, which they dont, not in the way the answer given implies.

Are you suggesting they are not both NVME based NAND drives using slightly different socket formats? Or that their sockets are not cross compatible ? perhaps its that they both use different Nand technologies or different controllers, what is it you think we are implying here ?

I love these kinds of answers. They gaslight your last answer and shut down any kind of replay you might want to make in one go by claiming authority status before you even start...
Ill bet you a "Linux Guru" too right?

Intel Optane was a rabbit hole I went down a while ago, back when it was relevant and there were a lot of snare traps along the way mostly from people who come out with answers like yours. Or read all the papers on the internet but had no practical experience of using it and so kept leading me down dead ends.

But to give a clean answer that doesnt get me told off.

No, Im not suggesting the sockets arent NVME or that they arent electrically compatible.
The point is that Optane M.2 Memory modules and Optane U.2/PCIe SSDs are completely different products with different requirements, different controllers, and completely different underlying storage technologies, so answers about one do not apply to the other.
This is where my previous answer is built from, because the post before it makes several assumptions which might not be either technically correct or possible to do in the real world.
The post before states a load of speeds that look great especially in the world where SSDs are very expensive and an H10 or H20 Optane SSD can be had for £100 (I mention them because people who read these threads see theyre Optane and assume there will be some speed benefit to them). Which it true, but what it leaves out is that a system after Intel gen 10 CPUs will only really use the NAND portion of the drive. The Optane portion simply isnt supported anymore. So yes, the numbers look good on paper, but only on platforms that actually support Optane Memory acceleration. On newer systems, youre effectively buying a QLC SSD with a chunk of Optane you cant use.

For instance an i7 7700 might support Optane if in a 300 series chipset board, but if its in a 100 series chipset board it doesnt. But then you can take that 300 series chipset board and put a i7 6700 CPU in it and it doesnt support Optane any more. Even though its the exact same board, in fact its the same system.
I have both. An HP EliteDesk 800 Mini G4. Both identical apart from one has an I7 6700 and the other has an i7 7700. One cant use Optane the other can. OR more accuratly based on what people think, and going by your post I include you. They can both use the NVME M.2 ssd. They just cant both use the Optane portion of it.

To break it down clearly:

1. Optane M.2 “Memory” modules (16/32/64GB)
These are not NAND drives at all.

They are 3D XPoint caching devices that rely on Intel RST and require:
7th, 8th or 9th gen Intel Core CPUs (Note Core, Pentium and Celeron is not supported even if the right gen)
200‑series / 300‑series chipsets (B250, H270, Z270, B360, H370, Z370, Z390, Q2370 etc... However not all motherboards with those chipsets have it supported in the BIOS, OEMs cut features)
BIOS support for Optane Memory mode (See above, this is important, you might have a supporting chipset that doesnt have Optane support or has Optane support but only in certain BIOS revisions)

They do not behave like normal NVMe SSDs. They cannot be used as standalone storage devices. This is the product people often confuse with “Optane SSDs”.

They are NOT NAND they are 3D XPoint.
Some, H10, H20, come with NAND glued to the Optane yes but the Optane isnt NAND its 3D XPoint.

2. Optane U.2 / PCIe SSDs (900p, 905p, P4800X, P5800X)
These are NVMe SSDs, but again not NAND — they use pure 3D XPoint media.

They do not require any special CPU or chipset.

They work on any NVME capable system, Intel or AMD.

They have completely different controllers, firmware, endurance characteristics, and latency behaviour compared to NAND NVMe drives.
Going back to your gaslighting quoted post.
im not "suggesting they are not both NVME based NAND drives" Im out n out saying they are not NAND they are 3D XPoint.

3. Why the distinction matters
When people say “Optane only works on certain CPUs/chipsets”, they are talking only about the M.2 Optane Memory modules.

When people talk about Optanes performance advantages (latency, endurance, SLOG use, etc.), they are talking about the U.2 / PCIe Optane SSDs, which do not have the same platform restrictions.

So yes — both use NVMe as the protocol, but:

They are not both NAND

They do not share controllers

They do not share platform requirements

They do not behave the same way

They were designed for completely different purposes

That’s the crossover confusion I was pointing out. Because when a thread like you mentions Optane, even when there is a picture of the drive people assume all Optane operates the same but it doesnt.

Well that clears that up !

But we were not discussing the memory modules/cache modules or even the M.2 Optane drives, just discussing the lovely U.2 Optane drives that were blazingly fast and were indeed storage drives and comparing them to current M.2 NVME drives which while exceptional fast still fall behind in 4k reads and writes compared to the Optane U.2 drives.

Just saying you may have jumped the gun here a bit and decided to throw us into that rabbit hole with you.

For the record I have used Optane is all its forms and can happily say that Optane memory locked to Intel only platforms can burn happily in hell where it belongs. (Seriously it was a mess)
The NVME Optane drives however were really good.

Trashbytes wrote on 2026-08-11, 05:20:

For the record I have used Optane is all its forms and can happily say that Optane memory locked to Intel only platforms can burn happily in hell where it belongs. (Seriously it was a mess)
The NVME Optane drives however were really good.

Optane had clear advantages from a performance standpoint and projected life span.
But....
Optane also had some very noticeable limitations when it came to $/ GB and storage density that drastically limited market penetration and ultimately market acceptance.

It's worth noting that Optane had some very interesting uses on database servers as a high speed write ahead log & Pem storage .
But ....
Those configurations were exotic enough that getting them deployed at your average IT department, while technically possible, was practically impossible. "What's octane?" "That's outside our VM standards". "Just buy more RAM" "Our hyper converged infrastructure doesn't support it" "Can we upgrade the SAN instead?"

douglar wrote on 2026-08-18, 15:36:

Optane had clear advantages from a performance standpoint and projected life span. But.... Optane also had some very noticea […]
Show full quote

Trashbytes wrote on 2026-08-11, 05:20:

For the record I have used Optane is all its forms and can happily say that Optane memory locked to Intel only platforms can burn happily in hell where it belongs. (Seriously it was a mess)
The NVME Optane drives however were really good.

Optane had clear advantages from a performance standpoint and projected life span.
But....
Optane also had some very noticeable limitations when it came to $/ GB and storage density that drastically limited market penetration and ultimately market acceptance.

It's worth noting that Optane had some very interesting uses on database servers as a high speed write ahead log & Pem storage .
But ....
Those configurations were exotic enough that getting them deployed at your average IT department, while technically possible, was practically impossible. "What's octane?" "That's outside our VM standards". "Just buy more RAM" "Our hyper converged infrastructure doesn't support it" "Can we upgrade the SAN instead?"

It was a thing that was very niech even in its own day and that day was back in the 7th 8th and 9th gen days.
But it was a thing and even now people get confused with what it was. With this being a retro forum it wont be to man years in the future before somebody asks if Kaby lake, Coffee lake and Coffee lake refresh CPUs are old enough to be considered retro. And when they do what might want to know about Optane because for those generations it was a thing.

Looking back at RAM prices, starting out in 1990 and 30-pin SiMMs, and we are still cheaper than around 2011 when adjusted for inflation. Go back to the earlier days and the prices were insane.