Intel roadmap includes faster Optane SSDs, 144-layer NAND

Optane vs. NAND flash performance

By contrast, Intel's testing with its DC P4610 NAND SSD showed latency starting at about 80 μs to 100 μs and gradually rising as the IOPS increases, until contention for shared resources saturates the drive at about 300,000 IOPS to 350,000 IOPS. The first-generation DC P4800X Optane SSD starts at about 10 us of latency and reaches saturation at about 500,000 IOPS, at which point the latency shoots up steeply.

"With Alder Stream, they're going to be able to put a whole lot more traffic into the SSD before the latency hits a brick wall," said Jim Handy, general director and semiconductor analyst at Objective Analysis.

Although Handy views Alder Stream as an important development, he predicted that Intel'sOptane DC persistent memory  modules will ultimately prove more popular than the Optane SSDs. He said the memory modules deliver considerably higher performance at roughly the same cost, so users comparing the two options will flock to Optane DCPM.

The next upgrade to Optane DC persistent memory, code named "Barlow Pass," will fall in line with the next generation of Intel's Xeon processors, which are code named "Cooper Lake" and "Ice Lake." The roadmap that Intel laid out also shows third- and fourth-generation Optane DC persistent memory and SSDs coinciding with its Sapphire Rapids and future Xeon processors.

Optane persistent memory workloads

Workloads that early adopters target with the Optane DC persistent memory include databases, artificial intelligence, analytics, virtual infrastructure and storage, according to Kristie Mann, senior director of product management for Intel's NVM solutions. 

More than 200 proofs of concept are in progress with enterprise, cloud and communications service providers, including Argonne National Laboratory, T-Systems and Verizon, according to an Intel spokesperson.

Intel claims major OEMs, software vendors and cloud service providers are developing products and services using Optane DC persistent memory. They include Alibaba, Cisco, Dell EMC, Google, Hewlett Packard, Microsoft, NetApp, Oracle, SAP and SAS.

"The large number of OEMs give credibility that Optane will live on as a key NVM component into the future," said Jeff Kato, a senior analyst and consultant at Taneja Group. "It's important, because IT can start to purchase Optane and know they will not be abandoned in the future."

Intel recently opened a new technology development facility in Rio Rancho, New Mexico but has yet to disclose its high volume manufacturing plans. Intel and Micron ended their joint development of 3D XPoint, but Intel continues to receive supply from Micron.

Optane DC persistent memory is primarily used in servers now, but Intel claimed it would soon be coming to workstations for content creators and others who deal with large data sets. Optane DCPM could deliver up to 3 TB per CPU and 6 TB per client workstation, according to David Lundell, senior director of client SSD strategic planning in Intel's NVM solutions group.

144-layer 3D NAND

Flash technologies on the horizon include 144-layer 3D NAND SSDs, code named "Arbordale +", that Intel plans to ship as quadruple-level cell (QLC) before following with the less dense triple-level cell (TLC) NAND. Intel claimed its floating gate technology provides an advantage in terms of read accuracy and data retention over the charge trap technology that other manufacturers use.

Intel's Hady said the floating gate technology would be extensible to five bits per cell, also known as penta-level cell (PLC) NAND, to further increase density and lower costs over QLC NAND.

During the Flash Memory Summit in August, Toshiba and Western Digital also discussed their exploration of PLC NAND and the challenges associated with developing it. Toshiba indicated its development work includes both charge trap and floating gate technologies.

"The main challenge in multi-bit designs is generating and maintaining narrow voltage threshold distributions. If the distributions overlap, then you cannot differentiate the original value, leading to bit errors," said Doug Wong, a member of the technical staff of Toshiba Memory America's Memory Business Unit.

Toshiba's Wong said lower endurance and/or data retention time would make PLC best suited to read-intensive workloads in colder storage applications. Toshiba, Western Digital and Intel have not laid out time frames for PLC.

"I'm not going to hold my breath," Handy said. "QLC is just starting to become big now, and it was first introduced in 2006. So maybe in another 13 years, PLC will become big."

Greg Wong, founder and principal analyst at Forward Insights, said cloud providers will drive QLC adoption to increase significantly next year, and enterprises will follow in 2021 and 2022. Enterprise QLC use cases include content delivery networks, analytics, databases and machine learning, he said.

Kato expects vendors to push QLC as a flash-based capacity tier that could replace slower hard disk drives. He envisions uses cases in high-performance archive applications or entry-level storage.