NAND flash technology may have peaked with the transition from planar to 3D, but there’s still a lot of room for iterative innovation.
That innovation won’t only come at the materials level, but also at the software and firmware level – controllers will continue to play a key role in boosting flash performance, as well as interconnects such as Peripheral Component Interconnect Express (PCIe) and Non-Volatile Memory Express (NVMe), which have been critical for exploiting the capabilities of NAND flash in solid-state drives (SSDs) and freeing them from limitations of hard drive architectures.
NAND flash can be found in a wide variety of devices, including smartphones, tablets, ultrathin laptops and even automobiles, but the biggest driver of innovation is arguably the enterprise data center, which is increasingly dominated by AI workloads requiring denser, faster SSDs.
Sebastien Jean, CTO of Phison Technology, said NAND flash SSDs are heading in two different directions – lower end, lower cost storage and higher end devices to meet the demands of AI. “Right now, everybody's driving sports cars to go pick up milk.”
This is because there’s a storage ecosystem designed around slower hard drive technology, which is now increasingly using SSDs as the cost of flash has come down. Jean said Phison is having serious conversations with hyperscalers to provide a low cost, NAND-based hard drive replacement. “What can we do to NAND to make it cheaper to essentially fill this niche in such a way that economically it makes sense?”
AI is guiding NAND flash development roadmaps
In the meantime, there’s a lot of focus on how to better optimize SSDs for AI workloads given the bandwidth limitations of PCIe while taking pressure off GPUs through controller capabilities, Jean said. “If you take that load off of a CPU and GPU, you free up a lot of cycles and they're easy tasks that are well-defined.”
He said it’s now possible for a controller to manage petabytes of storage. “You can't put a petabyte of NAND on the standard board size, but who says we have to keep the standard board size? There's a lot of room for innovation.”
When it comes to replacing hard drives with SSDs, it’s not a performance play, Jean said. “It's about being able to address a very large capacity of NAND while keeping the power down.” Such a replacement would leverage capabilities in NVMe that aren’t available in a SATA or SAS interface, such as accommodating large data movements.
One path for NAND innovation is Penta Level Cell (PLC) NAND, which Jean believes will have all the same issues that came with Quad Level Cell (QLC), and more, so it’s not a Holy Grail. “It will eventually happen, but that's not what really people are focused on.” Rather, the focus for SSD makers is how they can help the AI ecosystem grow, he said, and how best they can structure flash.
Software makes NAND flash SSDs smarter
Silicon Motion is addressing the needs of the AI ecosystem with QLC SSDs and controller technology, as well as its MonTitan Enterprise Development platform, which includes its proprietary PerformaShape multi-stage shaping algorithm configured in firmware that optimizes SSD performance on a per user defined Quality of Service (QoS) baseline. Algorithms like PerformaShape are an additional layer to NVMe capabilities such as Zone Named Spaces (ZNS) and Flexible Data Placement (FDP), which reduce latency, boost performance, and enhance endurance for AI data access.
Sandisk recently announced its UltraQLC 256TB NVMe SSD, which combines BiCS8 QLC CBA NAND with custom multi-core controllers. Aside from its high capacity, the SSD’s architecture also improves performance, power efficiency, and reliability. It also uses a BiCS8 2Tb QLC die to double storage density without increasing die size.
Companies that are both designers of the fundamental technology that goes into NAND, as well as controllers, packaging, firmware and storage devices are well-positioned to innovate around NAND flash.
They included Micron Technology, which recently announced its 2600 NVMe SSD, which leverages its ninth-generation QLC NAND in an SSD and features its Adaptive Write Technology (AWT), which improves sequential write speeds and performance of QLC NAND by delivering multi-tiered SLC, TLC and QLC dynamic caching architecture.
In an interview with Fierce Electronics, Mark Montierth VP and GM for Micron’s mobile and client business unit, said Micron’s six-plane architecture of its 2 Tb G9 QLC NAND allows for higher degrees of parallelism and increases read and write commands issued to the NAND simultaneously to improve performance.
While QLC is the latest and greatest NAND technology, he said there’s been many trade-offs, which has pushed Micron to look at how it can improve the user experience and optimize the capabilities of QLC while achieving TLC performance.
Overall, Montierth said Micron is focused on reducing power and improving performance across the form factors customer want – the company’s technology allows to offer up to two terabytes even in smallest form factor. He said innovation that end users see is just important innovation under the hood, such as seeing faster performance from their everyday productivity apps.
For SSDs, the industry wants to stick with the M.2 form factor, although UFS is also popular for mobile systems. Montierth said cost is a huge driver. “M.2 is pretty darn good and super established. It will take a very compelling value proposition to move away from that.”
NAND faces physics constraints
Going forward, Micron must accept the limitations of the NAND itself, Montierth said. “Physics is physics.” QLC is slower than TLC, which is slower than SLC, so the challenge is how to create SSD capabilities given that physic, which is where Micron’s AWT comes in. “It is a caching technology but done in a way and sized and proportioned to many, many workloads that we've looked at so that the actual end user performance is completely equivalent to TLC.”
Montierth said AWT is an example of how being a designer of the process technology as developer of the software allows for innovation despite certain realities and limitations that can’t be changed, and there are several innovation vectors Micron can employ, including the NAND process technology, CXL and controllers.
In the meantime, there’s emerging workloads in data centers and the cloud that are shaping how Micron thinks about NAND and how it approaches its next generation NAND architectures, Montierth said. “We live in an evolving world of requirements. If necessity is the mother invention, it is going to be the workloads and the demands.”
AI is a big driver, he said, and shaping any pressures on SSDs, NVMe, HBM or DRAM.
Micron has its own controllers while also employing third-party controllers from companies like Phison. Montierth said having a mix allows the company to find the right fit for customer requirements.
Jim Handy, principal analyst at Objective Analysis, said Micron has a lot of “smart stuff” available to it that can improve the performance of SSDs, even as adding planes increases die size, such as the adaptive write technology, which enables QLC to be used like SLC. “That's a clever thing that is managed by the controller.”
Clever controllers enhance “crummier” flash
In the same way that Moore’s Law has made flash in the SSD cheaper, it’s also makes it cheaper to put functionality in a controller, Handy said. “You can use crummier flash. There are all kinds of wonderful things that you can do now that you've got more transistors on the controller.”
While there’s lots of room for clever things that can be done with crummy flash, NAND will never see as big a move from planar to 3D again, Handy said. “I don't see anything lurking that's anywhere near as important as that.”
He said the biggest change going on right now is hybrid bonding, developed by YTMC several years ago, which is quite expensive at the time. Other vendors such as Western Digital are now using hybrid bonding with its BiCS 8 NAND.
Moving to hybrid bonding won’t be as radical change as the move to 3D NAND, and it won’t likely be as difficult, Handy said, adding it took Samsung three years to get to volume production of 3D from when it began shipping it in 2013.
“Other NAND flash manufacturers have not stated outright that they're going to do it, but they have indicated that hybrid bonding is a part of their plan,” Handy said. “That will get them a little bit more speed and it will keep costs in line.”