a simple guide to semiconductor memory boom-bust cycles |
| a simple explanation of this often repeated
historic drama for newcomers to the SSD market, along with a cautionary
footnote about SSD's own future | |
| by Zsolt Kerekes,
editor - StorageSearch.com
- February 11, 2016 |
I recently discussed
the memory boom bust cycle with a non technical reader who asked me to explain
why high volume semiconductor memory makers get into the situation of
oversupply and lossy pricing.
Here's some of what I said.
The
semiconductor memory business has wavered between under supply and over supply
since the 1970s.
In the 1980s the best analagy was what economists
called the "hog
cycle".
But any study of memory market history for the
past 40 years shows that there are also other competitive factors involved.
Simplest example when suppliers choose to use price as a way to gain
advantages in the market.
But another factor is that memory is not a
usable product by itself. And memory technologies have to be designed to be
optimal for end markets.
Memory makers have to guess years in advance
what the mix of products in the market will be to ensure they will get a share
of the end market.
If they get those guesses wrong there is under /
over supply which impacts prices.
The investments which have to be made in manufacturing plants are huge
and take place years in advance of even knowing in detail what the detailed mix
of products will be.
Memory makers mitigate their risks by choosing manufacturing
processes which have flexibility. (For example the ability to switch between
making DRAM or nand flash using the same equipment). But even with long term
roadmaps such plans can go wrong.
For example - sometimes
technologies are harder to get working than expected.
In 2017 such
factors - discussed in my (July 2017) article
3D nand fab yield -
the nth layer tax - resulted in memory shortages which had widespread
consequences.
As a sanity check FYI the entire SSD market wasn't in any memory maker's
technology road map until recently. But it affects everything they now do.
In the long term the risk of being uncompetitive (no one buys the
memory and you certainly go bust) or the fear of being locked
out of the SSD
market counts as worse score than temporarily having too much product at
too low a price. Because being in the game means you will get another
opportunity to supply memory and adjust product lines in the next memory
boom cycle.
You wouldn't believe how much
data the memory people
suck in and analyze.
The problem is they are disconnected from the
end markets and because of that cannot make optimal judgements.
They're also biased by their
interpretation
of how they managed risks and investments in the past.
That's one reason why all the big companies will want to become
enterprise storage
box suppliers. (It's more than
branding, and
involves assumptions about being able to raise the revenue ceiling from
the same
raw flash while also racking up incremental design efficiency gains
which are only possible
at scale.)
But the memory makers don't understand enough at present to make it
work. And they may have the wrong cultures too.
If the market stabilizes as I predict in
this
article then the enterprise SSD market may become as predictable as the PC
and server market were in earlier decades.
But we're nowhere near that
yawn-inducing stability yet.
Instead we need to go through some more
rounds of extreme instability first.
(Because as I said
a year ago (in 2015) -
"I don't think we've reached stability in reference enterprise SSD designs
and use cases"
Later:- as you'd expect with such a vast topic - the boom and
bust business cycles in memory have been analyzed and dissected many times
before.
Here are some links I found later which provide helpful
guidance for those - like my reader - who want to get a better feel without
being buried by semiconductor concepts.
For historic reasons - most
such discussions focus on DRAM - because that was the dominant memory revenue
earner in earlier decades - but the memory type makes absolutely no difference
to the principles.
- The life cycle of memory generations is related to price and volume in -
DRAM Pricing - a
white paper (2002) on the web site of
Tezzaron - which says
- "this life-cycle has been repeated often enough to exhibit some
predictable patterns."
I would encourage anyone (even
experienced semiconductor veterans) to note the 8 patterns listed. This
list includes almost everything you need to know to understand what drives
the predictable aspects of memory market behavior.
What's
the relationship between the memory market and the SSD market?
The
relationship between the SSD market and raw memory market has changed over time
and will change significantly again.
Before the
enterprise flash
era you could go far in understanding the pricing models and
cost base of SSD
products based on an understanding of the raw memory market - because
memory was one of the main ingredients of cost. Also in those earlier days -
SSD market revenue was so insignificant in comparison to the raw memory market
- that SSD-specific requirements did not in any way influence the behavior of
memory companies.
From
2007 it
started to become clear that
controller
architecture could play a bigger part in the SSD pricing equation than memory -
and by 2011
the activity in the market (measured by product success) showed clearly that
SSD-aware software too had the potential to differentiate SSD products in
business ways which de-emphasized the role which memory played in
characterizing SSDs (in nearly all markets).
And in
2012 I commented
on this paradign shift in my end of year analysis in these words... "A
safe rule of thumb is that knowing the generic memory type characteristics
doesn't give you enough useful info about the SSD's characteristics and
limitations any more. Whereas only a few years ago - knowing the exact memory
type was an essential starting point in understanding the SSD."
Looking
ahead (from today in 2016) to the long term future for the memory market -
the same kind of conclusion I analyzed for the hard drive market in my 2012
article How will
the HDD market fare... in an SSD storage world? - which I summarized
as... "the SSD market will be bigger in revenue than the HDD market ever
was" is likely to hold true for the semiconductor memory market too. That
is to say - "SSD market revenue will be a much bigger (high multiple in
revenue) of the raw memory market.
But it is by no means inevitable
that the biggest memory companies will become the biggest SSD companies.
I
discussed the strategic business pressures (for memory, HDD, storage and server
makers - who all face similar disruptive challenges) in my 2013 article -
hostage to the
fortunes of SSD.
What relevance does this have to this article
about memory boom bust cycles?
It's not safe that they will be closely
tied to future SSD boom bust cycles.
I already discussed the potential
of next generation SSD
software adoption to result in segment-specific crashes in SSD revenue in
my 2013 article - meet
Ken - and the enterprise SSD software event horizon. We've already seen
this type of thing happening. This lesson is an extrapolation from real vendor
revenue experiences - not just a blue sky what-if?
So you can take the
"SSD software event horizon" as an
extreme market
boundary condition factor whereby future SSD busts could become even more
distantly connected to raw memory production costs.
And here's a
simpler example of memory to SSD business disconnect.
A big collapse
in memory prices could be seen as a good thing by independent SSD companies
(which don't have their own memory wafer fabs) because low memory prices may
help the SSD makers to open up new markets.
I'll just conclude this by
saying:- SSDs (with the right software and architecture and ecosystem context)
have roles which can sometimes be equivalated to storage, processors and memory.
But because SSDs enable a wider range of potentially competitive data
system implementations than were ever before possible before (the modern
era of SSDs) it's
unwise
to analyze and project SSD business models simply by parachuting in models
from other previous (and deceptively) similar markets. | | |
| .... |
 |
| .... |
and what about semiconductor memory pricing? |
This issue got an interesting airing on
linkedin in May 2017
following an initial post by the CEO of BeSang -
Is
Micron's 3D memory (really) cheaper than 2D? - ostensibly about the
comparative costs of 3D nand and planar and the respective SSD prices from
Micron.
At the time the industry was in a rare supply shortage with
high book to bill. Also the industry was undergoing revolutionary changes unlike
any seen since the 1970s with new memory types and new application roles for
memory architecture.
The note below is part of one of my comments in
these streams. | |
Manufacturing
costs are only a small part of the pricing decision process. More important are
the competitive environment and user value propositions.
So when it comes to nand flash then key pricing factors are whether
differences in characteristics can be used to advantage in higher value systems
products or not.
It also involves judgements about whether there will be customers in
the ecosystem who are willing to make the investments in software and
controller IP to deploy the memory in higher value products such as enterprise,
industrial and military SSDs and what their competitive supplier options are.
If 3D had worse endurance say than same market generation 2D it would
always be sold at a lower price regardless of how much the costs. Because the
market values endurance and can do useful things with it.
Other useful
characteristics in memories have traditionally been latency and power density.
Going back to the role of manufacturing costs...
Semi companies cross subsidise new products based on their analysis of
the projections of use and mix they see. This is a riskier business than those
outside the semi business appreciate and only works on a casino strategy.
Fab
independent SSD products are akin to card counting - and have many odds stacked
in their favor - because everyone is a
hostage to the
fortunes of SSD and in particular
rackmount
implementations which deliver the highest
efficiency (greatest
user value at lowest chip count - compared to solo drives).
But with
disruptive new memory such as 3DXpoint the casino can lose too.
Pricing
is a complex business.
My wife Janet Downes
used to run a 4 day workshop on pricing strategies and the disconnects between
cost and price for product managers in some of the world's largest tech
companies. (I only read the cover sheets).
I think Sang's (CEO of
BeSang) naïve in appearance but provocative
post
and the various comments (including those still to come) might warn readers
outside the hub of the industry that these issues are more complex than first
appear. And the better that people understand the semiconductor memory market
the more easily we can communicate with them and learn their often unstated
needs. |
related reading
what goes into the cost of
an SSD? - what goes in and what stays out
user value
propositions for buying SSDs - which led to early market adoption
exiting the
astrological age of enterprise SSD array pricing - how marketers unblocked
uncertainty with innovative pricing schemes
problems with
identifying segments in the enterprise SSD market during disruptive market
change - why gaps exist and why vendors risked over competing in too few
easily located segments | | |
| .... |
|
|
.... |
 |
.... |
|
|
.... |
| StorageSearch.com /
popular SSD
articles /
SSD market
history | |
.... |
 |
SSD education
who's who in the SSD
market in China?
terabyte
talliers and other storage market researchers
miscellaneous
consequences of the 2017 memory shortages
did leading DRAM makers
collude in 2017 to protect high prices?
are we ready for
infinitely faster RAM? (and what would it be worth) |
... |
| DRAMeXchange said the
supercycle of DRAM price growth - which had lasted for 9 consecutive quarters -
was over. |
| SSD news - October 2018 | | |
... |
| "Currently the average
Flash fab is just over 100,000 wafers per month... For DRAM 60,000 starts proves
to be an efficient volume." |
| Why are NAND
Flash Fabs so Huge? (September 2018) | | |
... |
"I would be the last
person to say this isn't a cyclical business because the one thing that
suppliers will never be able to control is the demand environment and as you
know we're in a very capital-intensive business.
So, once you have the capacity then almost no circumstances will you
choose not to produce because you have a very, very high fixed cost structure
and so the industry responded appropriately to the weakening PC demand
environment and first nobody added capacity, and secondly, we all shifted our
output to those more lucrative segments... " |
| Ernie Maddock,
CFO - Micron (January 10, 2017)
at
Micron
Presents at Needham Growth Conference (transcript - by SeekingAlpha.com) -
in reply to a question about the cyclical nature of the memory business and
similarities to the previous 2014 cycle. | | |
... |
 |
... |
Ironically, DRAM producers'
entrance into flash memory production actually contributed to defeating their
original purpose for entering: flash memory's average selling price dropped
below that of DRAM in 2006 due to oversupply.
Estimating proper supply
for the flash memory market is complicated by the unpredictable nature of flash
memory demand it is unclear what consumers will deem the next great gadget to
drive the market, and when it will appear. |
| The
Rise of the Flash Memory Market: Its Impact on Firm Behavior and Global
Semiconductor Trade Patterns - 1990 to 2007 (pdf) (USITC site). This document also includes flash
market revenue from 1990 to 2007. | | |
... |
As companies invest in new
fabrication facilities to produce new technologies, the risks inherent in these
business investments continue to increase.
In 1983, it cost $200
million to manufacture 1,200-nanometer chips in a leading-edge semiconductor
facility.
In 1997, it cost approximately $1.3 billion for a
350-nanometer facility, and in 2003, it cost as much as $3 billion for a 130-
nanometer facility. |
| Semiconductor
Manufacturing: Booms, Busts, and Globalization (2005) - article in
The Bridge, (NAE site ) | | |
... |
|
|
... |
"Although a weak
economy may contribute to poor results, fundamental structural factors have an
even greater influence on the memory sector. In fact, they may largely explain
why memory suffers more than other semiconductor segments during downturns and
why memory players were not able to create economic value between 1996 and 2012,
even though their technological innovations significantly contributed to the
semiconductor industrys growth.
First, competition was intense.
Second, both DRAM and NAND flash were commoditized and primarily
differentiated based on price per gigabyte" |
| Memory:-
Are challenges ahead? - a blog by McKinsey
(March 2016). | | |
... |
|
|
... |
|
|
... |
In the year 2000 no one
caught a cold from the Y2K bug - but 3 things did happen which would shape
enterprise server performance for the next 16 years.
1 - hard drives
reached the latency limit set by waiting for a 15K RPM rotating platter (they
never got faster)
2 - 64 bit processor clock speeds reached almost
their maximum clock speeds (they got more cores but the cores didn't get faster)
3
- in the RAM market - the fastest server motherboard memory latencies in 2000
were very similar to what they are today (in 2016)
We all know that
SSDs came to the rescue of latency constrained advances in computing which had
been stalled by (1) and (2) above.
With some help from software the
next target is (3).
Why's DRAM so bad?
Most of us thought
that DRAM was the gold standard for latency you can rely on (unlike that cheap
flipperty gibbet flash). |
| DRAM's
indeterminate latencies and the virtual memory slider mix - a new blog
(March 2016) from StorageSearch.com | | |
... |
| "Even though DRAM ASP
growth is forecast to slow in the second half of the year, the annual DRAM ASP
growth rate is still forecast to be 63%, which would be the largest annual rise
for DRAM ASPs dating back to 1993 when IC Insights first started tracking this
data. The previous record-high annual growth rate for DRAM ASP was 57% in 1997." |
| IC
Insights (July 18, 2017) in a
research
note about the memory market which also noted that unit shipments of DRAM
and nand were the same or less than in the previous year - reflecting that the
transition to more layers of 3D memory in the past year resulted in more bits
but not more throughput of working chips through legacy fabs. (Discussed in
SSD news.) |
 | | |
... |
|
|
... |
| As nand flash is still the
dominant memory technology used inside over 99.9% of SSDs and as the switch from
planar 2D to 3D was pitched as the lowest risk solution to continue supplying
the growing need for more flash storage - because memory geometry shrinks had
gone about as far as they dared go - there has been much uneasiness and disquiet
recently in markets which use these memory chips about the shortages which have
been the most obvious sign that progress has not been entirely plain sailing.
|
| are more dimensions of
analysis needed to get a clearer picture of future 3D nand successions | | |
... |
We've seen compelling
evidence in the SSD market that the companies which design memories are not the
best at extracting the value from what they make.
Despite many
acquisitions
the odds are stacked against the memory companies being able to fix these IP
deficits internally. They don't have the internal pressure and they dont have
external market knowledge they need.
The emergence of
NVMdurance for
example shows that a parallel universe exists of unsatisfied alternate routes
through the maze of endurance permutations with shipping and future flash
memory.
|
| StorageSearch.com' s editor (August 2016) in
reply to questions about the market. | | |
... |
So why does Micron want
more RAM capacity?
A semi-serious interpretation might be - it's a poison pill. Who would
want to buy a RAM company?
Another interpretation is that - by paying
a premium price for Inotera - Micron is saying it is worth more too. |
| Micron's $3
billion gamble on Inotera (December 17, 2015) | | |
... |
|
| |
