Intel's Optane Died Because 3D NAND Was Too Cheap, Not Because the Tech Was Bad
Phase-change memory was worked on from 1970 to 2022, and every time it was squeezed out by cheaper flash. 3D XPoint was stuck between NAND and DRAM, worse than both, and in the end even Intel gave up on it.
The argument · tap a timestamp to hear it
The physics of phase-change storage makes it slow
Phase-change storage uses a material switching between crystalline and amorphous states to store data: the crystalline state has low resistance, the amorphous state high resistance, and the two differ by three to four orders of magnitude. Melting the material and quenching it into the amorphous state is called RESET; the reverse is called SET. SET either melts the material into a liquid and then slowly cools it to crystallize, or heats it to a temperature high enough to convert it directly to crystal without melting. The author points out that SET is often the speed bottleneck for PCM. This physical mechanism has not changed since Stanford Ovshinsky discovered the phase-change behavior of chalcogenide compounds in the 1960s.
Intel built phase-change memory in 1970, then forgot
In 1970, RG Neale and DL Nelson of Energy Conversion Devices and Intel's Gordon Moore published an article in Electronics magazine announcing "Read-Mostly Memory," positioned between RAM and ROM, to store data that programs read often but rarely change. Each cell had a silicon diode plus a layer of phase-change material; Intel supplied the silicon substrate, ECD supplied the material. But in 1970 the minimum feature size was only 10 micrometers, power consumption was high, SET took milliseconds, and endurance was low. DRAM capacity was growing, EPROM arrived, and RMM disappeared. Years later, someone asked the 79-year-old Moore about the article, and he said he had no memory of it at all.
Numonyx treated PCM as its only shot at a comeback
NAND's rise sent NOR flash into long-term decline, and in 2007 global NOR sales fell 9% year over year. Intel merged its NOR business with STMicroelectronics' flash business, forming the Swiss company Numonyx in 2008, with investment from Francisco Partners and bank loans. NOR lacked NAND's growth prospects, so Numonyx could only bet on PCM—PCM had both NOR's fast random reads and NAND's high-capacity storage, and in theory could replace both. In February 2008 Intel announced delivery of its first PCM prototype, Alverstone, at 128 megabits and 90 nanometers.
Mobile PCM was swept away by a wave of 3D NAND
PCM's entry strategy was to attack the mobile market, replacing the NOR flash that stored boot code while also storing data. In 2009 Samsung released a 512-megabit PRAM at 60 nanometers, saying that paired with DRAM it could save more than 20% in power. Micron acquired Numonyx in 2010 for about $1.2 billion in stock, and in July 2012 finally shipped 1-gigabit PCM at 45 nanometers, which Nokia put into its Asha feature phones. But in August 2013 Samsung announced mass production of 24-layer 3D NAND, and 3D NAND then set off the steepest cost-decline curve in semiconductor history, and mobile PCM's most compelling reason was gone. In January 2014 Micron announced it was pulling out of mobile PCM.
Intel and Micron were split on Optane's positioning from the start
In July 2018 Intel and Micron agreed to a friendly split, each going independent after completing joint development of second-generation 3D XPoint. The disagreement was over product direction: Intel initially aimed Optane SSDs at consumer PCs and laptops, hoping to spread wafer costs through volume; Micron wanted to go straight into the data center. Jon Carter hinted in 2016 that consumer systems did not really need 3D XPoint's low latency, non-volatility, and endurance, or at least not enough to pay four to five times the price of NAND. In October 2018 Micron announced it would pay $1.5 billion to buy Intel's half of the IM Flash plant in Utah, and Intel shifted to producing it itself.
Optane in the data center required redoing motherboards and software
In April 2019 the first DIMM-form-factor Optane DC Persistent Memory went on sale, four years after 3D XPoint's launch, requiring matching Xeon processors, chipsets, and motherboards. It had two modes: Memory Mode combined Optane and DRAM into a single volatile memory pool, working out of the box, but its value depended on DRAM price fluctuations; App Direct Mode was the truly valuable mode, letting software write persistent data to Optane so it survived a reboot. Oracle or SAP databases typically take more than an hour to reload data from SSDs or hard drives after a restart, and Optane could eliminate that. One imprecise but directional test showed a 12.5x improvement. But App Direct Mode required programmers to rethink data consistency issues.
Only 40% of 582 proofs of concept used the truly useful mode
In 2020 Intel updated Optane adoption, saying there were 582 proofs of concept in the pipeline, with SAP HANA and Oracle Exadata as the main users, and the rest mostly in high-performance computing, infrastructure, and some AI analytics. But only 40% used the more valuable App Direct Mode. Intel said it believed the industry would eventually adopt this mode, but it might take years. In the same period Intel announced the sale of its 3D NAND business to SK hynix for $9 billion, with Optane not part of the deal, which looked like an endorsement but also let people infer that Optane's implied loss in 2020 exceeded $500 million.
After Micron's exit, Optane was left with a single supplier
In March 2021 Micron announced it was immediately stopping 3D XPoint development, saying the Utah plant had long suffered from underutilization and dragged the company down by $400 million a year; the plant was later sold to Texas Instruments. Micron attributed the cause to 3D NAND's continued price declines and to the difficult changes data centers and applications would have to make to fully exploit the product. Micron instead bet on the open-standard interface CXL, because machine-learning systems made Micron realize that not just CPUs needed more memory—GPUs and AI accelerator ASICs did too. Micron's public exit severely weakened Optane: customers could rely only on Intel as a single source, and Optane was proprietary technology. In July 2022 Intel quietly announced the shutdown of the Optane business on its earnings call.
In their own words · checked verbatim
Asked many years later about the article, a then-79 year old Moore replied that he did not recall it at all. Ouch.
This is the most significant non-volatile memory advancement in 40 years ... There have been plenty of attempts to find and develop new non-volatile memory technologies
Ed Doller9:25
Vendors had spoken, they preferred NAND and DRAM over NAND, DRAM, and PCM.
The demand is through the roof. There’s not enough. We’re trying to manage the capacity and the capital expense to build the capacity for years three, four and five.
Jon Carter16:36
It's not just the SSD and the HDD comparison ... It's non-volatile memory connected directly to the CPU, replacing DRAM, which is a completely different paradigm.
Shannon Poulin21:45
Optane was just not the thing for the time.
Jack of all trades. Winner of none.
Figures
| 3D XPoint read/write speed relative to NAND | 1000x faster | 0:02 |
| 3D XPoint endurance relative to NAND | 1000x better | 0:02 |
| Capacity of a single 3D XPoint die | 128 gigabits | 0:02 |
| 2007 global NOR flash sales, year over year | down 9% | 8:25 |
| Price Micron paid to acquire Numonyx | about $1.2 billion in stock | 13:33 |
| Micron's purchase of Intel's half of IM Flash | $1.5 billion | 17:38 |
| Total Intel and Micron investment in the IM Flash plant | $6 billion | 18:38 |
| Optane's implied 2020 loss | more than $500 million | 24:51 |
| Annual drag from Micron's Utah plant | $400 million | 25:52 |
| Price Intel sold its 3D NAND business to SK hynix | $9 billion | 24:51 |
Glossary
- 3D XPoint
- A non-volatile storage technology jointly announced by Intel and Micron in 2015, later made into Optane and QuantX respectively.
- Phase Change Memory
- A non-volatile memory that stores data using the resistance difference between a material's crystalline and amorphous states.
- Storage Class Memory
- A memory category between hard drives and solid-state storage, combining non-volatility with access speeds close to DRAM.
- App Direct Mode
- Optane's persistent memory mode, in which software can write persistent data directly to the Optane module.
- Compute Express Link
- An open-standard interface that connects compute, memory, and storage systems into pools that can be allocated on demand.
How to listen
Engineers and investors watching storage chips, data-center hardware, and semiconductor industry history, especially those who want to understand why a new technology gets squeezed out by a cheaper old one.
The first minute of technical background can be fast-forwarded; jump straight to the phase-change storage mechanism at 3:07.