|Choosing a DRAM
supplier for embedded server applications |
by Zsolt Kerekes, editor StorageSearch.com - November 2014
a DRAM supplier for new embedded server applications used to be much simpler
than it is today - involving factors like:-
- capacity - Although there was always a market for memory makers who could
pack more into the same sockets - either by qualifying new memory sooner than
traditional oems - or by mechanical packaging techniques which were able to
- speed - When you look into the details of each memory generation - it's
surprising how much variation there has been in products from different
suppliers - even if competing products started out with the idea of aiming at
the same standard generic speed.
- reliability - Aside from factors which the memory buyer can control - such
as testing and burn-in - the biggest determinator of memory (and most other
complex semiconductor devices) is how hot they get.
So a big part of
the reliability question is about power consumption.
But the actual
design of the chips (layout and details of the cells) also determine things like
the operating margins on refresh. And the only safe way to know these - is to
stress sample products and vary the timing parameters at different temperatures
- and see at what point they lead to failure. Another factor in how hot do the
memories get is airflow and conduction cooling. Are some modules better than
others by accident or by design?
- price and availability - Buyers will often say - these are the most
important factors. It may be so - depending when they enter the conversation.
But having cheap memory which fails - under condistions when your system is
still expeced to work - is an overriding technical consideration.
of supply is probably more important than spot market pricing. Will you still be
able to get devices when there are worldwide shortages of the raw memory chips?
And will you still be able to get devices in 2, 3 , or 7 years time - if you
still need them?
And that was - when memory was simple - and the applications were
- experience - Anyone who has been designing electronic systems for several
product generations learns things which are important to their business - which
are rarely discussed in the technical press.
Sometimes - even if you
can't identify exactly what it is that makes apparently identical products from
some suppliers more dependable than others - you know that it's true.
For example - in the earliest days of DRAM it was known (or strongly
suspected) that the designs from some Japanese companies had been based on
copying (or reverse engineering) the original designs of other competitors.
That introduced problems due to mask registrations errors and
different process tolerances.
It also used to be the case with leading US memory companies - such
as Intel - that the same product produced by different fabs would have different
failure characteristics - due to subtle undocumented differences in what was
happening in the production plants during different shifts.
variations are better managed nowadays - but you can still see differences in
the "same" product from batch to batch - depending on whether the
original chip maker has changed their production process - something which may
not matter to most of their customers - but which may make a difference to you.
For modern day embedded mamory designers there are
additional variables to consider.
- there's a much wider range of design variation in DRAM chips from different
suppliers and different product ranges. Obvious examples being 2D vs 3D.
- there's a much wider range of operational temperatures which embedded
servers can expect to experience. Not only at the cool end of the specturm (as
servers get out into places they wouldn't have fitted before) but at the hot end
too - as some users decide that a good way of saving money on electrical power
is to have less cooling in their racks. The high cost of rack space - also
means that servers are packed more tightly than they used to be. All of which
means that "standard" operating temperatures for memory devices - is a
different to what it was before.
- physical variations in DIMMs - in the past it used to be known that some
standard DIMMs occupied less height than others. Many of these differences have
now been standardized - so if you need low profile or ultra low profile modules
- you can now depend on these characteristics as part of the sepcification of
the product. (Much better than someone in manufacturing coming to tell you in a
panic that the memory you designed in last year doesn't fit any more.)
All of these variations mean that it really
pays to get to know your memory supplier, understand what drives their
thinking, try to assess in your own mind how much effort they have expended on
understanding and characterizing memory - and then see how well - all that
aligns with your own needs.
- value added tricky features:- in the market today we're seeing a lot of
application dependent value added features being offered in DRAM DIMM
footprints. The obvious examples are NVDIMMs. But some DRAM DIMMs have
integrated data integrity and ECC - while other DIMMS which look like RAM have
no internal RAM at all.
And if no one does exactly what you want
- then find a supplier whose product ideas you like the sound of - and don't be
afraid to ask about customization or BOM control so that you get the memory you
|SSD costs explained|
how fast can your SSD
articles on StorageSearch.com
is published by