One of the things which
demonstrates the extraordinary range of diversity in thinking about the SSD
market is the different answers you get to these 2 simple questions.
what's the best way to design a flash SSD?
where's the best place to put it?
in computer architecture will you get so many industry experts disagreeing on
such fundamental questions.
Try this thought experiment for
yourself replacing the concept of "flash SSD" with "microprocessor".
As in ...why can't we all agree about the future of microprocessors?
products like servers, notebooks and cell phones there's fairly uniform
agreement about how to design a microprocessor that's appropriate to the
performance and power/cost budget levels of these recognized standard products
- and whole families of micros which have been optimized for specific
categories of applications - so the designer of a dishwasher for example doesn't
have to waste time researching trends in 256 core 64 bit CPUs.
of us old enough to witness the excitement and confusion of the early
microprocessor market remember a similar state of confusion.
companies thought that micros might have an arm's length role in peripherals
like VDUs. Minicomputer companies initially thought micros were irrelevant
because they didn't have operating systems and compilers. First the hungry
micro companies ate the minis - then they ate the mainframes too. In other
product areas too - I remember having debates about whether it was excessive to
design products which had 2, 3 or more processors. It was easier to get the
job done with more - so that's what we did. And having crossed that line - and
proven it was a good idea later products used a lot more processors.
it go the same way with SSDs?
They're already eating the server
microprocessor market due to SSD-CPU equivalence - as I
predicted in 2003.
for what makes a good storage design in user environments are changing faster
now than at any time in history.
And in the past 3 years in the SSD
industry we've seen a lot of learning going on by designers of
SSD controller chips -
as successive products have demonstrated
(which make products work better for users), bad features (which made the
products look good in
benchmarks but didn't work so well in applications and
(important functions which are required for a hassle-free user life-cycle -
which never got into the original design).
That complex picture of
shifting sands of assumptions in the SSD market is the reason that
StorageSearch.com has thousands of
pages of content devoted to the subject of
SSDs. (It's still not
enough - I know.) And it's why thousands of
are writing on this subject. Just when it seems for a brief moment that
there's agreement about a topic - such as what's a good notebook or server SSD?
- Then a new technology generation - or article comes along raising a new
important issues - and it's back to having no safe answers any more.
is good - and where we end up with SSDs is going to look a lot different to
where we are today. The
SSD analysts agree
on some things - but disagree on others. Each SSD manufacturer would like to
think that their own way is best.
I have my own views about the future
of the SSD market and what it's going to look like. And they may be different
to yours. But maybe we can all agree and get accustomed to the idea that
what we are seeing now in SSD design and usage is still a long way from where
we will end up.
Tactically the SSD product decisions you make today
may be different to what you would do in 3, 6 or 12 months from now.
the present to get to the future and retaining flexibility by avoiding routes
which limit your future options seems like a good way to go. Translating that
into customer strategies - may mean don't worry if your notebook SSD choices and
suppliers are completely different to the SSDs you use in your servers. And the
tick lists will be different too. Server SSD choices? Just fix the problem you
have today - because your server assets may look completely different tomorrow.
Remember how email and the web transformed the server assets for
most companies in the early 1990s? That's the kind of seismic change SSDs will
have too. So having a 5 year SSD server plan may not be a realistic idea.
Thanks for reading the latest SSD squeaks
from the mouse site. I hope you find the articles and ideas you see here
helpful in stimulating your own thoughts - and helping you avoid obvious
pitfalls. If you like
what you see - why not tell your friends and colleagues to take a look too.
How can you replace all
enterprise hard drive capacity with SSDs starting from current projections
about flash wafer fab production capacity and HDD use?
SSD capacity - the iceberg syndrome - how does the amount of flash inside
a flash SSD compare to the capacity shown on the invoice? - There are huge
variations in different designs as vendors leverage capacity to tweak key
performance and reliability parameters.
What are the
application slots which suit SSDs? Can one type of SSD be good for more than
1 type of function? Are the SSD types we're seeing now all that there will ever
be in the market? Or will there be entirely new types of SSD still to come? If
so - when?
R/W and DSP ECC are market changing tools in SSD controller IP - but only a
small number of SSD vendors have this technology. Those who have it - say it's
a gamechanger. Those who don't have it - say they don't need it. This looks like
a replay of the flash versus RAM SSD wars - only this time the stakes are
Efficiency in SSD
design architecture (how many chips and how much cost are needed to get the
same SSD specification) could become a key business differentiator among leading
SSD companies. That's because there are many different techniques which work at
the question of what's RAM really? by which I mean
RAM in an SSD context
became increasingly complicated after the open declaration of the
SCM DIMM wars
era. This was due partly to the successes of enterprise flash which
seemingly flattened traditional latency gaps. However this was an illusion
because the SSD ecosystem would eventually prove to stretch latency bands into
more usable spectrums than had ever existed before. Lifting the fog from
virtual memory latencies which hadn't improved for over a decade unveiled
new opportunities for non DRAM memories.The next evolution may be that all
storage from the CPU to the cloud should just be regarded as context