Chinese chipmaker unveils Optane-like storage class memory

Numemory
(Image credit: Numemory)

Officially, the development of mainstream storage class memory (SCM) like Intel Optane has been suspended at big companies like Intel and Micron, but it looks like China-based Numemory, also known as Xincun Technology, has quietly introduced a new Optane-like memory. The company's first SCM memory devices wed the performance of dynamic random access memory (DRAM) and the non-volatile capabilities of NAND flash memory. But it looks like at a high price.

Good performance, not enough capacity (yet?)

As for general performance, South China Morning Post claimed last fall, citing the company, that Numemory's NM101 chips can be used for SSDs that can load a 10GB high-definition video in just one second. This is in line with what modern performance-mainstream SSDs with a PCIe 5.0 x4 interface can do. However, there is no word about the new memory's data retention rates or possible endurance limitations regarding the number of writes. 

Because Numemory's NM101 seems to use an industry-standard NAND interface with a standard 1.2 volts I/O (typical for internal voltages of low-power M.2 SSDs), it is possible to use these devices for SSDs, provided that there is a controller (and firmware) that supports this memory, it is possible to build SSDs that will be drop-in compatible with existing PCs. 

However, 64Gb (8GB) and 128Gb (16GB) memory devices are not exactly economically viable for SSDs these days. One needs 128/64 ICs to build a 1TB SSD. This may not be a problem for an enterprise-grade U.2, ruler, or add-in card SSD that can accommodate loads of NAND packages. But M.2-2280 modules can only house up to four memory packages on one side and four on another side, so it is probably impossible to build 1TB M.2-2280 drives based on Numemory's NM101 and NM102. However, judging how rapidly Numemory has managed to double the capacity of its devices, the company can indeed become a contender for high-performance SSDs. 

Also, if some clients in China would like to use SCM for some performance and relatively capacity-hungry applications that do not necessarily need an M.2-2280 form factor, they can certainly do so, unlike their peers elsewhere now that Intel's Optane (3D XPoint) is gone. Then again, Intel's 2nd Generation Optane memory also had two decks and offered 128 Gb per device, so building truly high-capacity storage devices using such memory was hard, too. Yet, Intel sold boatloads of Optane-branded products.

64Gb and 128Gb devices with a 3200 MT/s interface

Xincun Technology's 1st Generation Numemory device emerged back in September last year. The NM101 is a 64Gb (8 GB) single-level cell (SLC) 3D-stacked memory IC with an industry-standard 3200 MT/s NAND interface. The operating voltage of the memory array is 5.3/4.5 volts (which points to a mature process technology used to make the IC), whereas the I/O voltage is 1.2 volts, an industry standard. 

Since 64Gb (8 GB) capacity is hardly enough for more or less high-capacity storage devices, the company quickly followed up with its 2nd Generation product earlier this year. The NM102 memory chip features a 128Gb (16 GB) capacity yet retains an SLC architecture and a 1.2V I/O voltage. Numemory does not disclose the operating voltages for its NM102, so we can only make guesses. 128Gb is the capacity of Intel's 2nd Generation Optane (3D XPoint) memory devices. 

Xincun does not disclose further specifications of its devices, only briefly mentioning 'ultra-fast microsecond-level response times,' which is a very vague description of latency, to put it mildly. To put the number into context, Intel's Optane promised 10 ~ 15 microseconds read latency and over 200-microsecond write latencies, whereas modern 3D TLC NAND can boast 80 microseconds read latency and write latencies that are in the ballpark of hundreds of microseconds. By contrast, the latency of DDR5 SDRAM is 10 to 20 nanoseconds, depending on the exact memory subsystems.

No details about architecture

Unlike the vast majority of SCM producers, Xincum does not reveal the underlying architecture of its devices (impedance memory, phase-change memory, MRAM, FeRAM), so we do not know how close this SCM is to DRAM and how far away it is from SLC 3D NAND. Yet, the manufacturer states that its memory is a 'a promising next-generation non-volatile storage technology for large-scale mass production.' Speaking of mass production, it is also unclear whether Xincun has its own production capacity or has outsourced manufacturing to a foundry. As the company has 220 employees and 80% of them are R&D personnel, we doubt the firm has its own fab. 

Xincun Technology was established in Wuhan in July 2022 with the aim of developing storage class memory, which was then missing from the portfolio of commercially available memory technologies developed in China, which strives to become self-sufficient in terms of critical chip supply. Production of Xincun's chips can be considered a notable achievement in China's ongoing efforts to strengthen its semiconductor capabilities in general and develop unique technologies in particular.

Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

Read more
YMTC
Chinese chipmaker ships record-breakers: YMTC quietly begins shipping 5th Gen 3D TLC NAND
Western Digital
Sandisk puts petabyte SSDs on the roadmap, has yet to reveal release date
Nvidia CEO Jensen Huang holding Project Digits
Nvidia, SK Hynix, Samsung and Micron reportedly working on new SOCAMM memory standard for AI PCs
Crucial P510
Micron's new PCIe 5.0 SSD promises longer battery life — DDR5 portfolio also gains higher capacity CUDIMMs and CSODIMMs
Kioxia
Kioxia's new 10th gen 332-layer 4.8 GB/s 3D NAND flash is 33% faster than its 8th gen ICs
Micron
Micron and SK hynix unveil new LPDDR5X SOCAMM memory up to 128GB, debuts on Nvidia GB300 systems
Latest in SSDs
Tech Deals
Crucial's 2TB T705 - the fastest available consumer Gen 5 SSD - has sold out at Amazon, but you can get the same drive with a heatsink for an extra $20
4TB SSD Deals
4TB SSDs now 5 cents per GB, Samsung drives on Sale
Seagate SSD
Newegg is selling this speedy 1TB PCIe 4.0 SSD at 6 cents per GB
Kioxia LC9 123TB SSD
Kioxia shows off new 122.88 TB SSD — PCIe 5.0 LC9 packs a whole lot of QLC NAND
Crucial T705 SSD
Crucial's superfast 2TB T705 Gen 5 SSD is back down to its lowest-ever price — Grab one of these drives for the ultimate gaming PC build
Samsung 9100 Pro SSD
Samsung 9100 Pro SSD Review: Samsung’s Capable Answer to Phison’s Storage Gauntlet
Latest in News
Despite external similarities, the RTX 3090 is not at all the same hardware as the RTX 4090 — even if you lap the GPU and apply AD102 branding.
GPU scam resells RTX 3090 as a 4090 — complete with a fake 'AD102' label on a lapped GPU
Inspur
US expands China trade blacklist, closes susidiary loopholes
WireView Pro 90 degrees
Thermal Grizzly's WireView Pro GPU power measuring utility gets a 90-degree adapter revision
Qualcomm
Qualcomm launches global antitrust campaign against Arm — accuses Arm of restricting access to technology
Nvidia Ada Lovelace and GeForce RTX 40-Series
Analyst claims Nvidia's gaming GPUs could use Intel Foundry's 18A node in the future
Core Ultra 200S CPU
An Arrow Lake refresh may still be in the cards with only K and KF models, claims leaker
  • usertests
    Although it wasn't the smash hit that Intel and Micron had hoped, some companies definitely had use cases well suited to the characteristics of Optane/3D XPoint. New tiers of memory/storage are destined to reappear time and time again, unless we eventually get universal memory that can consolidate some of them.

    With the relatively low capacities described in the article, the first stop may be AI products that need a lot of memory. If the AI model isn't changing often, the endurance hit won't be too bad.
    Reply
  • jkhoward
    My main computers all use an Optane OS drive to this day. I bought a spare 280 GB PCIe one recently on sale, and a 128(?) spare NVMe one for a laptop. This technology should have never died.
    Reply
  • JRStern
    Intel Optane never officially admitted what their technology was, either.
    And it turned out to use too much power to write, generated too much heat, so couldn't stack like flash.
    So it could never produce price/performance that fit between NAND and DRAM.
    The idea that it could easily back up main DRAM in some useful way was a pipe dream.
    So where is this stuff going?

    >Numemory's NM101 chips can be used for SSDs that can load a 10GB high-definition video in just one second.

    At what power level? With what kind of cooling? Or do we just let it melt after the demonstration.
    Reply
  • Findecanor
    When Optane ("3D cross point") was being sold, there did appear third-party software that did take advantage of it. There did appear datacenter solutions that took advantage of it. I've also read about several research papers about persistent data structures, file systems and databases that had been enabled by the existence of Optane.

    If this Chinese firm has developed a successor for Intel's Optane, I do think there could be a niche ready for it.
    But there need to be the right politics, the right marketing and similar speed and durability as Optane to make it attractive, or that niche will remain small.

    Non-volatile memory technology is not a panacea: it is not going to change the world. At best, it is a middle-ground between DRAM and NAND Flash in both speed and longevity.
    It still needs wear-levelling, overprovisioning and caching in DRAM protect it from wearing out prematurely, just like NAND Flash does.
    Reply
  • bit_user
    The article said:
    Xincun does not disclose further specifications of its devices, only briefly mentioning 'ultra-fast microsecond-level response times,' which is a very vague description of latency, to put it mildly. To put the number into context, Intel's Optane promised 10 ~ 15 microseconds read latency and over 200-microsecond write latencies, whereas modern 3D TLC NAND can boast 80 microseconds read latency and write latencies that are in the ballpark of hundreds of microseconds. By contrast, the latency of DDR5 SDRAM is 10 to 20 nanoseconds, depending on the exact memory subsystems.
    I have a lot of respect for Anton, which makes it all the sadder when he gets something so basic, so wrong.
    : (
    The figure he quoted is for NVMe-based Optane SSDs, like the P4800X:

    Source: https://www.anandtech.com/show/11930/intel-optane-ssd-dc-p4800x-750gb-handson-review/5
    Even with the P5800X (second-gen, PCIe 4.0), this figure dropped to < 6 usec for QD1 4kB random reads (99th percentile):

    Source: https://blocksandfiles.com/2020/12/16/intel-launches-three-new-optane-drives-one-is-worlds-fastest-ssd/However, as you can also see from the above, the overwhelming amount of that 6 usec is spent in hardware, software, and firmware overhead. The actual 2nd gen Optane chips, themselves, had a latency of just 100 nsec (writes?) to 340 nsec (reads?) per 64 byte transaction.

    FWIW, the time to transfer 4 kB of data at the nominal speed of PCIe 4.0 x4 is more than half of a microsecond, if my math is right. So, if you cut the read size down to just 64 B, to match how the PMem DIMM is being measured, it'd shave off only about 10% of that 6 usec figure and the SSD would still be a couple orders of magnitude worse than the DIMM.
    Reply
  • bit_user
    usertests said:
    Although it wasn't the smash hit that Intel and Micron had hoped, some companies definitely had use cases well suited to the characteristics of Optane/3D XPoint. New tiers of memory/storage are destined to reappear time and time again, unless we eventually get universal memory that can consolidate some of them.
    3D XPoint isn't the only solution to this problem.
    "Micron’s nonvolatile DIMMs (NVDIMMs) combine the speed of DRAM with the persistent storage of NAND flash to remove I/O bottlenecks and deliver big performance"

    https://my.micron.com/content/dam/micron/global/public/products/product-flyer/nvdimm-flyer.pdfThat's an old, DDR4-based product. I don't know if they've updated it for DDR5. I think this sort of thing is likely to live on CXL, in the future (if not already). That opens up a much more wider and practical array of form factors, especially if you have some form of battery backup. I think CXL only adds a couple hundred ns of latency, which should easily match what Intel achieved with their PMem 200 or 300 DIMMs being on the CPU's local memory channel.

    Edit: here's a CXL-based memory module that combines NAND and DRAM to support persistence or tiering (which is called out as a separate use case):
    https://semiconductor.samsung.com/news-events/tech-blog/samsung-cxl-solutions-cmm-h/
    Reply
  • bit_user
    Findecanor said:
    When Optane ("3D cross point") was being sold, there did appear third-party software that did take advantage of it. There did appear datacenter solutions that took advantage of it. I've also read about several research papers about persistent data structures, file systems and databases that had been enabled by the existence of Optane.
    The key thing is to avoid syscall overhead (i.e. accessing it via the kernel), because that more than doubles the native latency of PMem DIMMs.

    If you merely have a memory tiering solution, you can naively run large-memory apps and let the tiering system try to migrate pages (i.e. promoting "hot" pages and demoting "cold" ones). For some applications, being able to go straight to persistent storage from userspace is probably still necessary, especially if the data access pattern is sufficiently random and exhibits too little data reuse. Basically, database workloads. I can't think of anything other than server apps that fall into this category.
    Reply
  • bit_user
    jkhoward said:
    My main computers all use an Optane OS drive to this day. ... This technology should have never died.
    Here's the thing about that. Intel sold off all their SSD assets to SK Hynix, except for Optane, which they just killed instead. SK Hynix is a NAND maker with its own fabs. If 3D XPoint were truly a viable technology, going forward, I'm sure they would've taken those assets as well. I expect Intel would've sold them for a song, given the alternative was simply taking a total loss on it.

    According to what I read, Optane was never profitable for Intel. More worryingly, it seemed unable to keep pace with density and cost reductions happening in both NAND and DRAM. While Optane performs much better than NAND, it's a couple order of magnitudes slower than DRAM and many orders worse than DRAM on endurance. So, it was caught in this weird niche and basically being squeezed from all sides.

    I think @JRStern is also right about it having a problem with write efficiency, which could be a concern for datacenters and mobile applications.

    I also looked into cold storage data retention and found that it seems even worse at that than NAND, which is quite the opposite of what I was hoping or expecting.

    P.S. I'm not a hater, as I did snap up a P5800X when Optane was end-of-life'd. Turns out, I should've waited to buy it. Ebay currently has some new, Dell rebadged drives selling for less than half what I paid (including from well-rated US sellers). Anything above 400 GB is way overpriced, though.
    Reply
  • davidjkay
    being able to random access and write at byte/word level rather than large block and persistence in power failure can be a big boost to some database type applications and potentially ai... can mean saving on lots of io and complexity to deal with the limitations of ssd and dram...

    eg a small change may require lots of io to log and be able to do commit or rollback... and it has to be done immediately to follow rules about not losing data if power failure...
    Reply