 |
|
| . |
|
| sizing the
scale of market for ULLtraDIMM SSDs |
Editor:- January 21, 2014 - On the face of it the
raw text of SanDisk's
announcement
yesterday - confirming what most of us already suspected (IBM's new eXFlash
DIMMs are indeed rebranded SanDisk
ULLtraDIMM
SSDs) didn't add materially to the sum of human knowledge about SSDs - but
in a conversation yesterday afternoon with Esther Spanjer
Director of Marketing at SanDisk
I got some interesting new insights into how the
memory channel
SSD Trinity (IBM,
SanDisk and
Diablo Technologies)
publicly think about this new proposition.
cost - Esther
didn't really want to discuss the detailed market price per terabyte of
ULLtraDIMMs - but it's clear that SanDisk is comfortable with a top level
analysis which for simplicity starts with assuming that a terabte of fast
PCIe SSD (from say Virident
or Fusion-io) costs
about the same as a terabyte of fast flash DIMM SSD.
Commencing
from that assumption IBM/SanDisk/Diablo are saying that users can get lower
system costs - in high end virtual desktops - using flash DIMMs - because they
support about 2x as many users in each server as industry standard
PCIe SSDs. Contributory factors are that ULLtraDIMMs have lower latency, and
also incur a lower DRAM
footprint.
The stinger in the tail is that the ULLtraDIMM solution uses
19nm
flash (due to
adaptive R/W)
- whereas all the ultrafast PCIe SSDs use more expensive traditionally
managed flash memory - which tilts the balance in cost comparison roadmaps.
Esther
also drew my attention to 2 interlinked performance aspects of the
ULLtraDIMM architecture (which I would describe as
SSD symmetry
characteristics) which intrinsically lend themselves better to some types of
applications. These are:-
- deterministic latency:- SanDisk has some benchmark results which show
that its ULLtraDIMM latency has low jitter - compared to some competing PCIe
SSDs.
SanDisk's SSD isn't unique having this characteristic. But what
is unique is being able to maintain this at a latency which is 2x to 3x better
than PCIe SSDs. This opens up new market applications - for motherboard based
SSD acceleration which previously had been restricted to more expensive FC
RAM SSDs.
- scalability:- Esther said that the
IOPS and
aggregate throughput of ULLtraDIMM SSDs scales almost linearly as additional
modules are added to a server. (This had been one of the original architecture
intentions stated by Diablo a year earlier.)
Esther said the 2
things which make it easier for ULLtraDIMM SSDs to do this are:- the higher
performance ceiling of the DRAM bus in servers compared to PCIe, and the fact
that the memory channel doesn't get interrupted and throttled in the same way as
PCIe lanes by other I/O devices in the same system.
(SanDisk's
ULLtraDIMM SSDs aren't unique in satisfying this desirable trait of performance
scalability in motherboard acceleration however, despite the fact that their
selected benchmark
comparison product suggests otherwise.) Another thing we touched
on was flash memory supply issues.
As SanDisk's ULLtraDIMM
SSDs are single sourced - it's reasonable to ask what would happen if IBM
ramped up its demand for these products?
That looks to me like a
situation where being part of a large
flash memory maker is an
advantage - compared say to being a small SSD startup at the end of a supply
chain. My guess is that if IBM demand for flash DIMMs ramped in the forseeable
market environment (in which fab capacity for all flash is finite) - then
SanDisk would probably prioritise this higher value enterprise SSD product
compared to allocating flash chip capacity to lower margin
consumer outlets.
That
leads on to the question of - how much revenue will this ULLtraDIMM SSD
bring to SanDisk in the next 18 months?
I didn't discuss that
particular question with Esther. - But it's an interesting thought experiment.
My thinking goes something like this.
I begin my recipe
by adding together the combined historic revenue of some comparable high
performance PCIe SSD companies (for example Fusion-io plus Virident). Let's
unitize this raw ingredient and call it "one".
Next come
some scaling factors. (Yours may be
different.)
We need to take into account and interpret the market
success that
IBM
disclosed recently in its rackmount SSD announcement and also factor in an
adjustment for market-receptiveness for new SSD server solutions which
promise to cost less than previous solutions, while being offered from a "safe
supplier" (and which don't force users to learn new software).
Toss
these factors into your market model blender and whirr gently for about 30
seconds - remembering to screw the lid on tightly.
Then I think a
modest ramp up revenue frothing model for this type of product would look
like "three" (3x the number you started with) but "five" or
higher might not be too wild either.
Remember this is only a
starting point. Because if you assume instead that IBM isn't the only channel
for these products - and may not even prove to the biggest route to market
then you need to go back to the
assumptions cupboard
and get a bigger blender.
But let's stay with the simple IBM case for
now.
How would all that impact competitors?
server
competitors?
Esther said IBM thinks this technology makes IBM
more competitive than other server companies in high end markets while also
being aware of the inevitably looming reality that SSD enhanced servers
reduce the number of servers that users need. (To do the same things they did
before.)
SSD competitors? - (This is my initial assessment.)
The
total market for SSD enhanced servers is growing more
complex as new
SSD software vendors
are enabling new architectural uses of servers - software defined storage being
just one example. As
I said in
2011 - the SSD market is changing from a world in which few servers had
SSDs inside to a world in which all new servers are enhanced by SSDs. As users
have diverse business needs - these trends can only be satisfied by an equally
diverse range of systems - in which
SATA SSDs,
PCIe SSDs and
memory channel
SSDs all play a part.
In the short term (1 year from general
availability) memory channel SSDs - like SanDisk's ULLtraDIMM SSDs - will take
part of the market previously occupied by fast PCIe SSDs and some of the
market previously satisfied by very fast rackmount SSDs. But they cannot replace
all these functions at the present time (due to capacity limitations, physical
proximity / distance constraints and high availability considerations - just
to mention a few factors).
On the other hand - by enabling 10TB class
memory as a commodity in servers - the new SSDs will also enable new
applications which were previously impossible or uneconmic.
Overall - I
think that in the next few years (while being vulnerable to the same
disruptive revenue shrinkage effects of improved SSD utilization ratios -
which I discussed
the SSD software
horizon and which can affect all vendors) the aggregate impact of the
ULLtraDIMM SSD market should be viewed as being mostly additive to market
revenue in a growing enterprise SSD space rather than being cannibalistic.
And
I'm confident that will attract new competitors in the memory channel SSD
space - either by alternative competing designs, new licensing options for
Diablo's memory IP, or new server designs in which the flash controller
management is designed into the stahdard motherboard. | | |
|
. |
|
 |
|
. |
|
| Micron doesn't currently
have a low latency, high capacity, flash DIMM SSD technology platform in its
product line yet. But I think it's inevitable that Micron will have to publicly
address this product gap / opportunity / threat to maintain confidence in its
server customer base. Clearing the patent and IP decks in readiness for this
may have been one of the factors in the December 2013 patent deal with Rambus. |
| the Top SSD Companies in
Q4 2013 | | |
|
. |
|
|
|
|
|
|
|
..... |
|
| |
 |
| .... |
| IBM shows off what's it's
been doing with the RamSan rackmount SSD product line it acquired from TMS -
and also launches first memory channel SSD based servers |
Editor:- January 16, 2014 - For most of the
previous decade (2000 to 2009) Texas Memory Systems
was THE company which competitors aspired to match in market position when it
came to fast
rackmount SSDs.
In the early part of this decade (2010 to
2012) TMS lost
its monopoly on rackmounts as it inevitably had to share the expanding market
with a lot of other companies - starting with
Violin (which overtook
TMS in
brand strength
in 2011) and then
other companies like WhipTail,
Kaminario,
Pure Storage,
Nimbus and
Skyera which had all
established strong market recognition by the end of
2012.
But
in those latter years (from
2009 onwards)
not only was TMS competing against all those newbie rackmount vendors - but
it was also engaged in another hotly contested
part of the enterprise
SSD market in fast PCIe
SSDs - where its product line was trying to find a place somewhere in the
narrowing gaps between Fusion-io
and Virident.
Then a year ago - in January 2013 -
IBM completed the
acquisition of TMS (which had been announced in
August 2012) and
since then we haven't heard much about these products apart from a few
glimpses - which enabled us to observe that TMS's rackmount products had been
retained and renamed - while their PCIe products were quietly end of lifed.
This
week - among other things - IBM has launched a new fast rackmount SSD family -
whose controller
architecture is effectively an enhanced adaptation of TMS's 8th
generation RamSan with some tweaks to incorporate newer memory, iron
out some RAS wrinkles (you can now change everything inside from the front or
back - without sliding the rack out) and a big investment to present a
software friendly face.
The new software capabilities are being
done by products which are being offered as external-to-the-box unbundled
subsystems (control enclosures) for those who want them. This means that the
performance and efficiency of the raw flash array isn't compromised in any way.
IBM's
new SSD box (a 2U
HA 16GB
FC fast rackmount SSD with
upto 48TB usable capacity,
priced at $683K
approx list) is called the
IBM
FlashSystem 840.
Earlier this week I spent an hour talking about
this new product with Woody Hutsell
and Levi Norman -
who are both now back in the IBM branded TMS fold having both sampled the
delights of some other leading SSD companies in recent years. Woody wrote about
his experiences
in a recent blog.
As I've known both of them for many years - I
couldn't help but start by saying - "This feels like one of those movies -
where they decide to make a sequel many years after - but all the actors look
much older. It's lucky for us this conversation isn't going out on YouTube."
You
can get a flavor of what IBM thinks it's doing with this new product - and more
details in its
briefing
document (pdf) - and I won't repeat much of that detail here.
Woody
said "It's interesting to me how much attention the flash operation is
getting within IBM's storage organization."
He went on to say
that IBM's big commitments to flash such as the $1 billion investment
announced last April
are seen within IBM as popular actions "which are important as we need to
compete." As a result - many competent people (in IBM) want to be a part
of the flashsystems effort.
Anther change in scale since TMS became
part of IBM is that the size of the development team for the flash systems
rackmount has quadrupled.
Sales are good too. IBM has shipped
over 1,500 of these flashsystems. In effect Woody said this was
limited by the fact that for 3 quarters IBM shipped everything they had planned
to make.
Woody said he thought that this alone - even without all the
other SSDs which IBM was selling into the enterprise market meant that IBM was
probably on its way to be one of the biggest vendors in the market.
I
said - a dominant market share in enterprise flash in 2014 might look like 5
or 10 per cent as there are hundreds of companies in the market. - We'll have
to see how things work
out.
But my guess is that with a few assumptions about density,
channels etc this means this rackmount IBM product line has possibly been
generating about $500 million of revenue in the past year - which explains where
some of the revenue missing from competitors' reports may have gone to.
Something
else which appeared in the briefing paper singing the praises of IBM's
expanding universe of enterprise flash product offerings - eXFlash DIMMs -
sounded to me like just another name for
SanDisk's
memory channel
SSDs (later
confirmed
to be the case) which appeared in another announcement
IBM
server announcements today - see footnotes for more.
What's my
final take on this? (FlashSystem 840 announcement)
IBM is now the
company to make comparisons with if you're looking for fast rackmount SSDs with
some high availability options. Particularly if you're working in a complex
environment - are a big customer and think you will be reassured by the
availability of compatible products and pre sales technical sales support.
IBM's
density - in terms of rack units needed to build a
petabyte SSD - is
better than some other fast systems - but remains an order of magnitude less
efficient than
Skyera - due to the
difference between IBM's use of eMLC compared to Skyera's claimed ability to
use TLC due to
adaptive
controller architecture - which is 2 generations (4 years) ahead of what's
used in this particular IBM box. (Having said that - IBM does already use some
degree of adaptive flash SSD technology in other systems - by virtue of the
SSDs it designed in from
SMART.)
Going
back to scary Skyera
- "On the other hand" - I said to Woody - "Skyera doesn't have
the same HA
or software in place
yet. But not everyone needs all these features."
Overall - for
competitors in the same high performance and reliability class as this new IBM
box (which includes companies like Violin,
Fusion-io etc) - IBM
can still be beaten on
price. It was ever
thus.
Footnotes - IBM's first memory channel SSD servers
In
another
IBM
SSD announcement today (alluded to above) about its new server
architecture which leverages
memory channel
SSDs - and making a comparison with
PCIe SSDs - IBM said -
"Our evaluators are seeing 5-10 microseconds write latency for eXFlash
DIMMs in preliminary testing vs. 15-19 microseconds latency for PCIe-based flash
storage from Fusion-io,
Micron, and
Virident, and 65
microseconds latency for
Intel S3500 and S3700
SSDs."
We've seen increasing granularity of detail emerge about
the system characteristics of memory channel SSDs emerging in a trickle of
announcements, and experimental user reports in the past year. Now that the new
flash DIMM SSD products are becoming generally available - there will soon be
better clarity on real world costs and performance. | | |
| . |
|
|
|
top SSD companies
what
changed in SSD year 2013?
How fast can your SSD
run backwards?
the enterprise SSD
software event horizon
my flash care scheme is
100x better than yours
|
| . |
| "I had the benefit of
the brick wall of ignorance. Not knowing what couldn't be done." |
| Skyera's CEO, in
the article - scary
Skyera | | |
| . |
 |
| . |
| We've all seen new
companies launching SSD software and pitching for the enterprise with products
which are little more than spruced up versions of "hello SSD world!" |
| (interview with
Maxta) | | | |
