• More News

    NAND Directs the Future of Memory Controllers


    The future of the memory controller is irrevocably tied to the memories they control. Similarly, they are governed by Moore’s Law. And while storage class memories (SCMs) may gain traction thanks to new architectures, the memory controller market is still very much dictated by NAND flash.

    “Everyone’s used to the flash’s density-doubling every couple of years or so, but they don’t consider the fact that the controller gets increasingly complex at a similar rate,” said Jim Handy, principal analyst at Objective Analysis. This has allowed them to go from SLC, to MLC, to TLC, to QLC while getting faster, wearing better, and adding features. “I often quip that controllers are getting so good so quickly that soon you will be able to hook a controller to a completely non-functional flash chip and still be able to pull out good data.”

    Marvell’s new Bravera SC5 controller family include performance security features that address the ever-growing workloads for SSDs in cloud data centers. (Courtesy Marvell)
    Even if emerging memories or SCMs can get traction, the market for memory controllers is very dependent on flash devices, and not even SSDs, said Handy. “From a unit volume standpoint, the largest number of controllers that ship is into flashcards — things like microSD cards.”

    In the meantime, he said, emerging memories generally require very little from a controller when compared with NAND when it comes to functions such as error correction code (ECC) and error leveling. 3D Xpoint memories such as Optane even have leveling built into it. “It’s a cakewalk to manage emerging memories compared to flash.”

    Any emerging memory controller is going to be rather simple to make, said Handy, while the controllers in flash card devices don’t nearly require anywhere near the sophistication of an SSD controller. Less sophistication means less cost.

    A low-end SSD controller might cost a dollar and a high-end SSD controller might cost $75 or a $100.” He said controllers are following the footsteps of pocket calculators that used to be hundreds of dollars. The features of a four-function calculator haven’t changed over the past several decades and because the chip inside them kept getting smaller and cheaper it got to the point a decade ago that people started giving away calculators for free.

    Handy said this is what’s in store for controllers if new features aren’t added or if customers are satisfied with the same functions that were available last year or for the last decade. “They’re either becoming more sophisticated, adding a lot of features, which they are, or they’re becoming much cheaper.” For companies that make their own controllers, there’s a benefit to keeping the cost of controller development down, but for companies that make controllers adding features such as increasingly sophisticated ECC keeps their products from becoming low-priced plumbing.

    NVM Express technical work group chair Peter Onufryk sees the controller as being a big part of an SSD’s value. Different vendors have different NAND, and some NAND is better than others, but the controller is what turns that NAND into something completely usable. It’s as integral to the SSD as the engine is to a car because it makes sure that NAND does it what it’s supposed for a given workload, whether the expectations are high or low, like a mid-range cellphone, he said. “It’s going to be in the trash two years from now.”

    Meeting those expectations with features and functions that differentiate the controller is key to the success of a merchant controller maker. Onufryk said there’s three primary users of these third-party controllers: hyperscalers who to do something special; SSD vendors who want to build out their portfolio; and, vendors that are selling unbranded SSDs.

    Ultimately, he said, any new feature in an SSD is made possible by the controller, whether it’s to balance power and performance, leverage the new capabilities of NVMe, or propriety functions. “Everybody’s still trying to still build their own stuff.”

    That can create a diverse set of feature and function requirements for controllers, but also challenges depending on whether it’s a fixed function controller or a programmable one, he said. “The programmability comes at the cost of power. On the flip side, these fixed function controllers aren’t able to evolve to handle new functions. There’s probably more people doing different things at this point than ever before.”

    However, said Onufryk, it’s a challenging business to be in because there’s only so many hyperscalers, and the roadmaps of NAND and SSD makers are in flux — where they might want to use a merchant controller can change. “The market is unstable.” It also requires a lot of expertise around firmware, media and ECC, he said, so it’s hard to imagine a startup wanting to enter the controller business or an established vendor wanting to develop controller products.

    Marvell is one of those merchant controller companies that’s focused on features. Its memory controller customers include Kioxia America and SK Hynix. Marvell’s new Bravera SC5 controller family include features that address the ever-growing workloads for SSDs in cloud data centers, including the need for security FIPS-compliant root of trust, AES 256-bit encryption and multi-key revocation capability. They also support PCIe 5.0 and NVMe 1.4b.

    Microchip offers its Flashtec NVMe controllers for high-reliability, high-performance PCIe Gen 4 SSDs, as well as supporting customers that have invested in both SATA and SAS. (Courtesy Microchip)
    Today’s memory controllers must also be responsive to form factor changes, NAND types, and meet rigorous physical, thermal and signal integrity requirements, said Thad Omura, vice president of marketing for Marvell’s flash business unit. The Bravera SC5 controller family is highly customizable as to address the variety of storage options needed for the cloud as workloads continue to evolve, which means cloud hyperscalers such as Microsoft Azure Facebook are just as much customers as SSD makers. “It’s all about meeting their requirements to fit in their demanding infrastructure for cloud and hyperscale.”

    Flash storage is growing at a very healthy pace in two main categories—storage and compute, said Omura. Marvell is support both with all the necessary protocols. However, one of the goals of hyperscalers is to reduce their inventory of drives so they can use essentially the same hardware for different applications. He said most products launched for cloud and data center SSDs are focused on storage or compute. “We can enable the same hardware to address both.”

    Another requirement is increasing the ECC capabilities because next-generation NAND in some ways less reliable as the NAND makers drive higher capacity. Other features that controllers must accommodate for are better power performance and thermal properties, as will as emerging form factors such EDSFF E1.S, which is popular with hyperscale customers as they move away from buying complete server from the Dells of the world in favor of building their own, said Omura. They’re even starting to build their own SSDs, which Marvell supports by selling them the chip directly and helps to simplify compatibility of all the pieces regardless of the form factor they choose.

    Some NAND vendors make all of their own controllers, while others do some of their own but turn to a third party. Omura said they might optimize parts of their roadmap and look to a company like Marvell when they don’t have the resources to develop solutions internally.

    Handy said there’s also a minimum cost to develop a controller and it costs more every time the controller gets increasingly sophisticated. A single company can’t afford the cost of developing its own controller for each different SSD its shipping so it opts to buy one from a vendor that’s selling controllers to two customers. “They can defray the cost of development over those two companies’ SSDs.”

    Microchip is another vendor who offers SSD controllers for NAND vendors who may have their own but also want to complement them with merchant controllers from Microchip. It’s also helping companies such as the hyperscalers who want to build more of their own data storage infrastructure for their custom requirements or to reduce costs, said Mark Orthodoxou, director of the company’s data center business unit. “We certainly have plenty of business with SSD suppliers because it turns out that building an SSD controller is actually really, really hard.” That has lead to some companies making missteps with their own implementations, he said. “In other cases, they just don’t have the ability to support all the features that are necessary to service their whole market.”

    Microchip is fully focused on enterprise and NVMe SSD controllers with PCIe Gen 4 products available now and a roadmap for beyond. What’s not in its future is controllers for emerging memories. Orthodoxou doesn’t believe there will be successful products with SCM, noting that Intel’s Optane-based SSDs have “flopped pretty hard.” He said it makes little sense to have storage class media on an NVMe interface. “What you’ll see is storage class media will move to CXL and that will be where the successes are.” Microchip is also focused on advancements on the hard drive side along with SSDs, including shingled magnetic recording (SMR) that efficiently increases areal density and capacity, as well as enabling an ecosystem over the long term for customers that have invested in both SATA and SAS.

    Omura said Marvell is keeping an eye on SCMs, but they have a long way to go before they might be a sizable enough market for a controller company. Intel is the lone player pushing hard with an SCM product with Micron having exited the 3D Xpoint market, and there needs to be more standardization around the interfaces for SCMs. “It doesn’t mean that we couldn’t have controllers that support them. We’re paying very close attention to it,” he said. In the meantime, CXL is coming to market, and Marvell is an active participant. “We do think initially CXL will focus on DRAM and then there is an ability to eventually integrate SCM.”

    Source: eetimes.com