IEEE Solid-State Circuits Magazine - Spring 2016 - 74

Use of NVMe drives can reduce storage latency by as much as 60% and
the protocol is an ideal candidate
for accessing the storage over the
network with remote direct memory access (RDMA), which adds only
an additional 10 μs of latency compared with local drives by bypassing the TCP/IP stack. This concept
is extended to interconnected network "fabrics," which are being
standardized as the "NVMe Over
Fabrics" protocol. The deployment
of these low-latency storage fabrics will greatly expand the realm
of data center storage possibilities.
For example, it enables a hyper-converged infrastructure model, where
each SSD, directly attached to a
server, is shared by servers connected through fabrics. Alternately,
it may prompt adoption of rackscale architectures, where SSDs are
collected in a single large sled and
shared across multiple servers in
the rack, such as the JBOF (just a
bunch of flash) sled recently introduced to the Open Compute Project.
In traditional data centers, all
flash arrays (AFAs), which combine
a controller with multiple SSDs in
a chassis, started to be deployed
around 2012. AFAs appeal to users
through the very high input/output operations per second and low
latencies that can be achieved by
leveraging flash over traditional
hard-disk drives, but, initially, that
performance came at a steep price
premium that could only be justified in special cases. However, over
time, several developments have
spurred rapid growth of AFAs in
traditional data centers. The rapid
decline in NAND cost, coupled with
AFA vendors' growing expertise in
flash management software, has
allowed the latest AFAs to benefit
from wider market adoption. Additionally, unlike spinning disks, flash
media is well suited to data reduction

S P R I N G 2 0 16

techniques, such as inline deduplication, which has driven the AFA cost
per gigabyte much closer to that of
traditional storage arrays. With AFA's
US$/GB within total cost of ownership
[10] striking distance of traditional
and hybrid arrays, AFA vendors have
taken advantage of the speed and
flexibility of solid-state storage to
write new storage management software from the ground up. The resulting systems offer a degree ease of
use and automation far greater than
that from any traditional enterprise
storage system, such as "one-click"
provisioning and allocating performance per volume. Latest systems
have also pulled even with high-end
traditional storage systems in terms
of data services offered, from mirroring and snapshots to synchronous and asynchronous replication.
Thus, AFAs are now an increasingly attractive option in traditional
enterprise data centers, especially
for the use cases involving virtual
desktop infrastructure, server virtualization, or databases. In the future,
as the price of NAND continues its
decline, AFA penetration is expected
to extend deeper into the data center
to target more applications and lower
cost storage tiers.

Summary
The history of semiconductors memory is marked by continuous innovation and advancement. Dependent
on advancing to smaller nodes to
maintain growth and profitability,
the industry is approaching a turning point. The speed of scaling is
slowing down, but the industry must
remain focused on a path to provide
the IT industry with reduced costs,
higher performance, and lower
power memories for the future. New
memories, new techniques, and new
market approaches are changing the
rules, but the objective of advancing
technology remains the same.

IEEE SOLID-STATE CIRCUITS MAGAZINE

References

[1] K. Kim, "From the future Si technology
perspective: Challenge and opportunities," in Proc. 2010 Electron Device Meeting, 2010, pp. 1.1.1-1.1.9.
[2] K. Kim, "Silicon technologies and solutions for the data-driven world," in Proc.
IEEE Int. Solid-State Circuits Conf., 2015,
pp. 1.1.1-1.1.9.
[3] U. Kang, et al., "8Gb 3-D DDR3 DRAM using through-silicon-via technology," IEEE
J. Solid-State Circuits, vol. 45, no. 1, pp.
115-119, Jan. 2010.
[4] What would you do without NAND flash
technology? [Online]. Available: http://
www.flash25.toshiba.com
[5] [5] J. Elliott, and E. S. Jung, "Ushering in
the 3D memory era with V-NAND," in Proc.
Flash Memory Summit, Aug., 2013
[6] Y. Cai, et al. "Program interference in MLC
NAND flash memory: Characterization,
modeling, and mitigation," in Proc. ICCD,
2013, pp. 123-130.
[7] J.-B. Lee, "Revolutionary Era of Flash Storage," in Proc. Samsung SSD Forum Japan,
Feb. 2016.
[8] [Online]. Available: http://www.samsung.
com/global/ business/semiconductor/
minisite/SSD/global/html/whitepaper/
whitepaper04.html
[9] J.-U. Kang, J. Hyun, H. Maeng, and S. Cho,
"The multi-streamed solid-state drive,"
in Proc. Hot Storage Conf., Philadelphia,
PA, June 2014, pp. 13-17.
[10] [Online]. Available: http://www.samsung.
com /s e m icon duc tor/supp or t /to ols utilities/tco-calculator/

About the Author
Jung-Bae Lee is the senior vice president of memory product planning,
Samsung Memory Division. He has led
the memory product planning (NAND
and DRAM) and application engineering team since 2013, where he plans
new memory products, defines mid-/
long-term product strategy, and builds
industry ecosystems. His responsibilities include enabling of all the company's current and future products such
as NAND, SSD, UFS, DRAM, Mobile
DRAM, and NVDIMM. Previously, he
led DRAM design team, where he
had been working on various types
of high-speed, low-power DRAMs
and circuit design technologies. He
has focused his passions on DRAM
design since he joined Samsung in
1995 and has been involved in designing many of the world's first DRAM
products, such as DDR/DDR2/DDR3/
DDR4/LPDDR3/LPDDR4 and WideIO DRAM.

http://http:// http://www.flash25.toshiba.com http://www.samsung http://www.samsung

Table of Contents for the Digital Edition of IEEE Solid-State Circuits Magazine - Spring 2016