 |
| leading the way to the
new storage frontier | |
... |
| ..... |
|
SSD history
Top SSD Companies in Q4
2017
after
AFAs - what's the next box?
3D nand fab
yield - the nth layer tax?
how fast can your SSD
run backwards?
who's
who in the SSD market in China?
capacitor
hold up times in 2.5" military SSDs
miscellaneous
consequences of the 2017 memory shortages
|
|
| ... |
|
which way next for SSD? |
| by
Zsolt Kerekes,
editor - December 1, 2017 |
It was the best of times.
It
was the worst of times.
2017
was a year like no other in
40 years of
SSD history.
The trajectory of SSD market's onward rebound
from memory shortages will be directed by existential questions.
For
the first time since
the modern era of
SSDs (the no-turning-back years since 2003) the approximate number of SSDs
shipped didn't rise substantially and instead remained essentially flat year on
year.
The most obvious cause for SSD shipments flatlining seemed to
be the much written about
memory
shortages.
Memory chip shipments in 2017 were reported by
market researchers to
have been about the same or less than in 2016 due to
worse than expected
manufacturing yield problems associated with next generation memories. This
was exacerbated by the impossibility of bringing new production capacity in
place fast enough due to long lead times of production equipment and
prohibitive investment risks.
While it's true that SSDs shipped in
2017 were on average bigger in capacity than in earlier years this was little
consolation to those SSD vendors whose growth ambitions - if they didn't own
wafer fabs - were shredded by circumstances seemingly outside anyone's control.
Stories
from SSD companies coupled with rising
memory and SSD prices
gave credence to the notion that user demand for SSDs would have increased
had supplies been available. Had user appetites for SSDs and memory been
satisfied then prices would have fallen rather than risen and in that case the
interpretation of shipment data would have led to a different market prognosis.
But
don't expect business to pick up where it left off when the next memory boom
bust correction kicks in. I think there are other factors already at work
which point towards the shape of future SSD shipments being materially
different in 2018 and 2019.
At the root of this are revisits to
fundamental questions:-
- what should memory systems should do?
For example should in
memory processing be part of the standard feature set?
- what they should memory systems look like?
In addition to the
obvious form factor and interface issues which attached to DIMM wars and
memory fabrics - another question is - should future memory arrays be optimized
as storage systems which can emulate memory? or as memory systems which
provide backwards emulation of legacy storage?
- to what extent should new memory systems be compatible with past
software investments?
Or - given the market's recent willingness to
engage with memory systems as the only way to advance affordable computer
boundaries - should the market aim higher? Should the opportunities (of
performance and cost) enabled by new memory system architectures change the
shape of the very processors and software they are intended to work with? You
might say - why worry? As we've seen with the great solid state storage
experiment it takes years to roll out new architectural dice and the
winning patterns can't exert a backwards influence.
I would argue
however that long before the truly revolutionary changes in memory systems
architecture are stabilized we are already seeing new influences coming from
pragmatic adaptations of currently shipping memory products (like NVDIMMs and
SoC compatible nvms) which - with the right software - have the ability to
change the ratios of other SSDs, memory and storage in the systems in which
they appear.
These incremental technologies will change the patterns
of use of memory in every kind of computing product.
Many SSD
designers in the past decade have been nibbling away at issues of
SSD efficiency -
answering the crude question - what's the best way to use any given number of
memory chips and if I can change the way they connect and the software. These
improvements have typically accumulated in chunks from as small as 5% to as big
as 50% in a single design feature (or patentable IP). As long ago as 2013 I
hinted at the tremendous gap between where we had got to compared to what
may be possible in my classic impacts of the SSD
software event horizon. Recently heralded companies like
Symbolic IO claim
that they do can even better.
The new reality is that
DRAM and
flash are no longer the
only defining memory types supported by useful software.
So called "emerging
memories" - some of which had gotten to be teenagers before they quit
their dark dens and emerged as data industry citizens - have this year
been at the heart of claims by systems oriented memoryfication startups that
they could change the world of storage and memory arrays arrays as much as
SSDs changed the landscape before.
In April 2017 on this very home
page - I asked the question - Are we there yet? - 23 years of SSD guides
later... I concluded at that time
the
memory systems market and this publisher are still "under construction".
Now with the benefit of hindsight it seems I was right.
Underlying the
shaping of the SSD market's future I think we must anticipate bigger changes to
come in the next few years. | | |
|
| ... |
|
 |
|
... |
|
| the need for new
processing/storage architectures
|
Editor:- December 12, 2017 - NGD Systems (formerly
called NxGn Data) today published the results of a survey on the need for
Intelligence storage for applications with large data sets.
The
purpose of the study was to gauge whether the movement of large data sets across
existing processing and storage architectures negatively impacts the cost and
usability of the data by applications.
The results of the survey -
Smart
Storage Survey Report (pdf) - conducted by G2M
Research for NGD - show that existing compute and storage architectures
adversely impact the performance and cost of these applications, and that new
architectures are needed if these applications are to continue to scale in size
and capabilities.
NGD has been a pioneer of in-situ processing and
their current approach is to leverage ARM processing cores within the
SSD controller
(although other types of implementation and tiering within the memory/ storage
assets have been reported in these news pages from other researchers and
developers).
Among other things the survey report says...
"Since
the advent of digital computers, the IT industry has regularly oscillated
between convergence and disaggregation, as well as how specific functionality
has been packaged and delivered to those who use it. The movement of processing
capabilities into storage media, as represented by in-situ processing in SSDs,
represents a new evolutionary path in IT that has been made possible by the
solid-state nature of SSDs." ...read
the article pdf)
Editor's comments:- 3 years ago I listed "in-situ
SSD processing" as #1 of the
12 key
SSD ideas which changed in 2014.
The implementations of this new
architectural idea has had to adapt pragmatically to changes in the SSD market -
notably the emergence of standards like NVMe and associated fabrics - but also
to other technologies which have been introduced to enable
memory
systems to work better such as
NVDIMMs, tiering
software and rethinking the relative size and roles of memory compared to local
storage (cloud
adapted memory architecture). | | |
|
| ... |
|
|
|
|
... |
|
|
|
|
... |
|
|
|
... |
|
|
| |
|
|
|
more pages like
this? |
| .. |
| Regulators will look at
competition in memory - says China Daily |
Editor:- December 28, 2017 - China's NDRC (the
business standards regulator similar in scope to the SEC in the US, and the
EU in Europe) will be looking at competition in the semiconductor memory
market prompted by complaints about shortages and high prices from mobile phone
makers according to
an
article in ChinaDaily.com.cn which says among other things...
"The
commission has spoken to Samsung, the world's largest storage chip manufacturer,
but has yet to launch an official antitrust review process targeting the
company..." ...read
the article
UMC offers 40nm SuperFlash from SST
Editor:- December
21, 2017 - UMC (a leading semiconductor
foundry) today announced
the availability of the company's 40nm process platform that incorporates
SST's
embedded SuperFlash non-volatile memory.
The newly available 40nm SST
process features a >20% reduction in eFlash cell size and 20-30% macro area
over UMC's mass production 55nm SST technology.
Editor's comments:- for
many of you who don't struggle with the temperature constraints of real
electronics systems you might glance at "40nm" and think - what's so
great about that? When 1Xnm has been the flag waving geometry for nand flash
for some time.
SuperFlash is NOR. And UMC says its robust SST process
performs according to JEDEC standards, with 100k endurance and more than 10
years of data retention at 85C and an operating-temperature range of -40C to
125C.
See also:-
military SSDs
optimizing
CPUs for use with SSDs
this
is not your Grandfather's industrial SSD market
IDC expects SSD revenue to reach $33 billion in 2021
Editor:-
December 20, 2017 - IDC
today
announced
- "The outlook for the SSD industry remains strong as units, revenue, and
total capacity shipped are all expected to see robust growth throughout the
2016-2021 forecast period. In a new SSD forecast update IDC now expects
worldwide SSD unit shipments to increase at a 5 year CAGR of 15.1%. SSD
industry revenue is expected to reach $33.6 billion in 2021."
See also:- more
market research
Samsung improves 10nm DRAM speed and yield
Editor:-
December 20, 2017 - Samsung
today
announced
today that it has begun mass producing the industry's first 2nd-generation
(faster) 10nm class 8Gb DDR4 DRAM. This has been accomplished using legacy
fab processing without needing yet the recourse of
next generation
EUV .
Re data integrity:- Samsung says a newly devised data
sensing system enables a more accurate determination of the data stored in each
cell, which leads to a significant increase in the level of circuit integration
and manufacturing productivity.
Re industry
memory
shortages:- Samsung's new 10nm DDR4 features an approximate 30%
productivity gain over the company's 1st–generation 10nm-class 8Gb DDR4.
Editor's
comments:- as always there are a wide range of memory geometries being used in
the market in various ages of legacy fabs.
But to set the context on
10nm DRAM here are 2 competitive comparisons from some smaller competitors in
the high capacity DRAM market.
- Taiwan based Nanya
- recently announced
production shipments of its own "independently-developed" 20nm
process 8GB DDR-4 DRAM.
Toshiba samples subtly different SAS hard drives
Editor:-
December 19, 2017 - I thought this was a joke at first. But it's for real. Toshiba today
announced
it is sampling the AL15SE - a 2.5" SAS HDD with 10,500 RPM spin
speed.
This is the first time there has been a new
spin
speed since the 1990s although it falls between the previous two fastest
bands.
Now I guess that rotating
storage reliability
experts can start worrying about whether the new frequency drives will create
subtle reliability reducing resonances if they are mixed in the cabinet with
other frequencies.
Seriously though - the combination of 12Gbs SAS in
the new hard drives and the almost imperceptible improvement in latency (hard to
notice when it's so slow compare to SSDs) shows that the storage industry which
has been desperately seeking more SSDs than it could get or afford in 2017 has
become receptive to any new nuanced device which can store data in an array
better than what came before.
See also:-
HDD articles & news
from the home page of SSDs...
in memoriam Diablo
Editor:- December 15, 2017 - Diablo Technologies has
filed for bankruptcy. More details can be seen in -
Ottawa-based
Diablo Technologies files for bankruptcy - in the Ottawa Business Journal.
There
had been earlier signs that the company had ceased operations.
A
story in the Register -
Lights
out at Diablo as plug pulled on website - noted a series of key personnel
departures and a recent statement by the company's legal counsel consistent
with the white space now on the company's home page and Diablo's non
appearance at FMS in
August.
Editor's
comments:- Here's what I said today to a reader who asked my views about the
fading away of Diablo Technologies.
My
interpretation
of Memory1 benchmarks (some of which were done by Diablo's own customers) is
that with some workloads you would get substantial performance benefits
(compared to assembling the same memory size with multiple DRAM based servers).
But – it was clear – even in Diablo's own published benchmarks that
with small data sets the performance was slightly worse (than without Memory1).
So as a user you had to be very clear to size the system in an appropriate
way. And clever enough to do it.
But more critically to the business case - I think investors must
have thought that if the company can't make buckets of money at a time of memory
shortage and high raw DRAM prices – then there would never be a better
opportunity.
Competitor Netlist
had shrewdly insulated itself from the cost of the next round of legal battles.
And the years of uncertainty in the earlier rounds must have deterred many
design wins.
The memoryfication market – with tiered enterprise
memory – now has so many competitors and so many form factors that no one
can be guaranteed to get a sizable chunk of it with any single product. It's
really a fragmented market in which there are many ways to get similar results
using entirely different mixes of technology at the server, box and
infrastructure levels. Sad outcome really.
Diablo's Memory1 was a
bold offering. And the industry is better for having been pushed towards
considering memory tiering at the sub microsecond level and the attendant needs
of software stacks and hardware - sooner - because of Diablo's efforts than it
might otherwise have done.
One way to interpret the ill preparedness
and rush to hype of Intel/Micron's unready Optane (née 3DXpoint) in
2015 was as a panic response to the outbreak of
DIMM wars
flashed over by rumors of Diablo's Memory1.
See also:-
Memory Channel
Storage SSDs - 2013 to 2017
Nanya presents overview of the memory market
Editor:-
December 14, 2017 - An overview of the $120B (in 2017) memory market - which
consolidates data from various market research sources appears in a
Presentation
to Analysts and Investors (pdf) - published today by Nanya Technology .
In
2017 worldwide revenue of DRAM was approx $69B - up 67% YoY.
In 2018
worldwide wafer starts for DRAM will increase moderately to 1,210K/month.
Toshiba and Western Digital bury the hatchet over flash
Editor:-
December 13, 2017 - Toshiba
and Western Digital
today
announced
a global settlement agreement to resolve their ongoing disputes in litigation
and arbitration, strengthen and extend their relationship, and enhance the
mutual commitment to their ongoing flash memory collaboration.
The parties' agreement to resolve all outstanding disputes ensures
that all parties are aligned on Toshiba's sale of TMC to K.K. Pangea, a special
purpose acquisition company formed and controlled by a consortium led by Bain
Capital Private Equity, LP ("Bain Capital"). The parties have agreed
on mutual protections for their assets and confidential information in
connection with the sale of TMC, and on collaborating to ensure the future
success of TMC as a public company following an eventual IPO.
salami
slicing Toshiba's SSD beauty pageant - timeline of stories
will increased fab capex secure future memory dominance?
Editor:-
December 12, 2017 - In his recent article -
Samsung's $26
billion bet (on ZDNet) - Robin Harris
says this capital expenditure on semiconductor production facilities by the
company which produces half the world's DRAM and flash memory ups the ante for
competitors. ...read
the article
AFA market revenue grew to $1.6B in 3Q17 - says Dell'Oro
Editor:-
December 6, 2017 - "AFA market revenue grew 33% yoy in 3Q17, reaching
$1.6 Billion" according to a new report -
Storage
Systems Quarterly - published by Dell'Oro
Group.
"All Flash Array is a very important technology
segment in external storage. In fact, as a percentage of external storage
revenue, it has been growing dramatically—from 22% in 3Q16 to 28% in 3Q17.
So every vendor is determined to expand their position in all-flash storage
systems" said Jimmy
Yu, VP at Dell'Oro Group.
Editor's comments:- Dell'Oro's press
release lists the 5 biggest vendors and the company can provide more analytical
data in their purchaseable reports.
HyperBus adopted by JEDEC
Editor:- December 5, 2017 -
Cypress Semiconductor
today
announced
that its HyperBus
low-pin-count memory interface has been included in the new eXpanded SPI
interface standard from the JEDEC Solid State Technology Association.
Cypress
provides a range of automotive qualified NOR flash "HyperFlash"
and self-refresh DRAM "HyperRAM" product families.
See
also:- interface glue chips,
storage ORGs
Samsung's future flash price trends discussed by Morgan
Stanley
Editor:- December 1, 2017 - A story on Bloomberg -
Morgan
Stanley Defends Downgrades of Asia Tech Titans after Rout follows the
reaction to an
earlier
analyst report by Morgan Stanley
which downgraded prospects for Samsung due to the
inevitability of corrections in the memory market and anticipated price
reductions.
Among other things the earlier report said - "We see
downside risk as NAND prices have started to reverse in 4Q17. Meanwhile,
visibility on DRAM supply-demand dynamics has reduced beyond 1Q18."
The
later Bloomberg article says - "Everyone knows NAND prices will fall..."
Editor's
comments:- comments from market
researchers are loved by the objects of their analysis when they result in
a positive market reaction. And are less well liked when the stories dent their
share price. In a 2012
spoof
article about one of the SSD heavyweights of that period I discussed this
syndrome and said - "If an SSD company is worth about the same on a
Friday as it was on the Monday - that's taken as a sign that everything's going
as planned." | |
| . |
|
|
| . |
 |
| . |
|
| |
| .. |
|
|
.. |
|
SLC / MLC /
TLC
tactical / permanent / real / virtual?
tradeoffs in the
design of mixed flash hybrid SSDs |
Editor:- December 20, 2017 - This month I
received a copy of a new (to me) paper -
a Survey
of Techniques for Architecting SLC/MLC/TLC Hybrid Flash Memory based SSDs (27
pages pdf) - from Sparsh Mittal,
Assistant Professor at Indian Institute of
Technology Hyderabad who is among the co-authors of this significant
reference document.
Re the scope - the authors say "For sake of a
concise presentation, we limit the scope of this paper as follows. We focus on
software-level management techniques for hybrid SSDs and not their circuit-level
design issues. We include techniques which use at least two types of Flash and
not those that merely use an SCM with a Flash cell-type.We focus on the key
ideas of each work and include only selected quantitative results, since
different works use disparate evaluation platforms and workloads. We hope that
this paper will be useful for computer architects, SSD designers and researchers
in the area of storage architectures."
Among other things the
paper discusses a wide range of externally referenced techniques including:-
- tradeoffs in using some portion of TLC or MLC as virtual SLC (to improve
latency and endurance)
- reliability and performance tradeoffs using volatile versus non volatile
RAM in buffers
- revitalizing worn MLC blocks as SLC
- varying the size of SLC designated buffers based on analyzing application
usage to optimize garbage collection
The authors note various
factors which are changing or need to change compared to previous generations
of SSD design.
- better runtime adaptation of control parameters
- the need for hybrid SSD specific simulators
- fairness and QoS (quality of service) joining the formula of design goals
in SSD design in addition to the traditional must-haves of performance and
reliability
If you've ever wondered about how to optimize SSD design by
using a mix of flash memory types in the same SSD then this paper is an
invaluable reference guide to the techniques which have been written about in
the public domain. ...read the
article (pdf)
hybrid drives,
SSD controllers | | |
|
.. |
|
 |
|
.. |
|
|
|
.. |
|
 |
|
.. |
|
|
|
... |
|
.. |
|
.. |
|
| Are we
there yet? |
| After more than 20 years of writing guides to
the SSD and memory systems market I admit in a new blog on
StorageSearch.com -
Are
we there yet? - that when I come to think about it candidly the SSD
industry and my publishing output are both still very much "under
construction". ...read
the article | | |
|
.. |
|
| If you're one of those who
has suffered from the memory shortages it may seem unfair that despite their
miscalculations and over optimimism the very companies which caused the
shortages of memory and higher prices - the major manufacturers of nand flash
and DRAM - have been among the greatest beneficiaries. |
| consequences
of the 2017 memory shortages | | |
|
.. |
|
|
|
.. |
|
|
|
.. |
|
| after AFAs?
- the next box |
Throughout the
history of
the data storage market we've always expected the capacity of enterprise user
memory systems to be much smaller than the capacity of all the other attached
storage in the same data processing environment.
 A
new blog on StorageSearch.com
- cloud
adapted memory systems - asks (among other things) if this will always be
true.
Like many of you - I've been thinking a lot about the
evolution of memory technologies and data architectures in the past year. I
wasn't sure when would be the best time to share my thoughts about this one.
But the timing seems right now. ...read the
article | | |
|
|
|
|
|
|
|
| |
