editor - StorageSearch.com
RAM used to be components:- SRAM,
DRAM etc. Now after 10 years of the
PCIe SSD market
learning curve experience we have to rethink the RAM concept. In the
virtualized world of SSDs everywhere - RAM is whatever the software thinks is
RAM - always provided that the typical average performance (at the apps level)
meets user QoS needs. And just as there have always been more than one type of
real physical RAM - there are now compelling ecosystem arguments (and user value
propositions) for supporting different multiple types and tiers of of SSD
virtualized RAM too. So in effect - RAM has changed from being tied to a
physical component to being a virtualized systems
software idea - and
the concept of RAM even stretches to a multi-cabinet memory fabric. Maybe it was
always so. There was more to RAM than the simple idealized notion which most of
us carried in our minds. Shattering that cozy RAM-in-a-chip picture forever
is a new truth now that the economic impact of the
software and the
systems architecture is nearly as important to the RAM (which you think
you're working with) as the raw nanometer layout of the memory cells.
can memory do
can memory chips be
made in the wrong place?
memory systems - the new box?
an SSD view of
semiconductor memory boom bust cycles
RAM news - ain't what it used to be
| data integrity in DRAM -
new paper lists and describes known techniques|
19, 2018 -
Survey of Techniques for Improving Error-Resilience of DRAM is a new
research paper by Sparsh Mittal
- Assistant Professor at IIT Hyderabad and Maruthi S Inukonda
published in the
of Systems Architecture.
The authors say - "Aggressive
process scaling and increasing demands of performance/cost efficiency have
exacerbated the incidences and impact of errors in DRAM systems. Due to this,
improvements in DRAM reliability has received significant attention in recent
years from both academia and industry. In this paper, we present a survey of
techniques for improving reliability of DRAM-based main memory. We classify the
works based on key parameters to emphasize their similarities and differences.
This paper is expected to be useful for computer architects, chip-designers and
researchers in the area of memory/system-reliability. ...read
Editor's comments:- Among other things this
paper has a detailed analysis of multi-dimensional ECC schemes, multi pin
correction codes (discussion of fault coverage and overheads), the efficacy
of retirement schemes for pins and chips,
RAID and various schemes
for stacked DRAM arrays.
rethinking memory systems design - 2018 horizons
July 27, 2018 - I saw a succint summary of the deep question - "move
data or compute locally?" - yesterday on pages 183 to 187 of
Memory System Design Robustness, Energy, Performance (290 pages - pdf) - by
Professor Onur Mutlu
in a keynote he presented July 3, 2018.
"A memory access
consumes about 1,000x the energy of a complex addition."
movement is a major system energy bottleneck."
Need a Paradigm Shift to... make computing architectures more data-centric."
the 2nd half of this (long) paper Onur describes the state of advanced research
and thinking into proposing and evaluating design solutions which intersect
with the ideas of optimizing data movements and processing inside memory chips
and memory arrays. ...read
the article (290 pages - pdf),
more papers like this
by Onur Mutlu
new article - re the utility of very much faster RAM
May 14, 2018 -
are we ready
for infinitely faster RAM? (and what would it be worth) - is a thought
provoking new article on StorageSearch.com
If someone could offer you a memory system which had the same
storage density (bits per chip / module / box) as mainstream RAM - but which had
latency and bandwidth (as measured by what the application sees) which was
infinitely faster - could we use it? - how much would that be worth? and how
would it change markets? ...read the
improving the latency and energy of commodity DRAM using adaptive
Editor:- March 13, 2018 - The enterprise flash SSD
market has a long history of design advances which came from the cumulative
understanding gained by the independent characterization of memories - this
mostly having been done by independent SSD and
rather than the original manufacturers of the
themselves. But I haven't heard much in the past 10 years about similar
activities related to DRAM - and part of the reason may well be that the
companies which used to do such in depth RAM characterizations in earlier phases
history - the RAM SSD
companies like Texas Memory Systems and Solid Data Systems - had mostly stopped
design work on new high capacity RAM SSDs by about 2008 due to the competitive
advantages (in a storage array context) of
I was surprised and delighted to come across a new
Understanding and Improving the
Latency of DRAM-Based Memory Systems (pdf) - by Kevin K. Chang
- Carnegie Mellon University (submitted December 2017 as part his PhD) which
document (in 200 pages approx) describes his ongoing work and insights into
DRAM characterization and system optimization opportunities.
research measured and analyzed the relationships between supply voltage and
latency in commodity DRAM and explored ways to optimize latency while still
maintaining data integrity and reducing power consumption. Among several schemes
also described in this paper:-
- an adaptive latency scheme he calls "Flexible-LatencY DRAM (FLY-DRAM)"
which leverages the variation of latency that occurs within different
locailities of DRAM chips.
...read the article (pdf)
- Voltron - a new mechanism that improves system energy by dynamically
adjusting the DRAM supply voltage using a new performance model which is based
on a better understanding of the relationships between cell retention, refresh
rate, temperature and other system factors.
is FRAM museumware?
Editor:- January 7, 2018 - It's
been a long time since I heard anyone
update from Fujitsu
Semiconductor says that its FRAM devices are displayed in an exhibit
called "New Electro Hall (Link to Cyberspace)" in the Science Museum,
FRAM products have very low capacity (kb upto 1Mb) and low speed (1 to
3.4MHz) but they have very low power operation too and can be used in "batteryless"
systems which harvest power from non traditional power sources such as vibration
- using piezoelectric transducers.
Editor's comments:- Having said
that - don't dismiss FRAM as this may be a type of museumware whose glory days
are yet to come.
A new article this week -
A New Memory
Contender? - in SemiEngineering.com surveys the history and technology
trends in ferroelectric memories and describes potential successors to FRAM
Samsung improves 10nm DRAM speed and yield
December 20, 2017 - Samsung
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
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.
shortages:- Samsung's new 10nm DDR4 features an approximate 30%
productivity gain over the company's 1stgeneration 10nm-class 8Gb DDR4.
Nanya presents overview of the memory market
December 14, 2017 - An overview of the $120B (in 2017) memory market - which
consolidates data from various market research sources appears in a
to Analysts and Investors (pdf) - published today by Nanya Technology .
2017 worldwide revenue of DRAM was approx $69B - up 67% YoY.
worldwide wafer starts for DRAM will increase moderately to 1,210K/month.
Series A funding for RISC CPUs in DRAM
September 8, 2017 -
UPMEM - a fabless semiconductor
startup - today announced
3 million Euros series A funding for its Processing In-Memory technology.
integrates user-API accessible RISC processors as SoCs in DRAM. The company has
in eeNews (Oct 2016) as saying...
"The fundamental benefit of
processing-in-memory is the combination of DRAM and CPU. We attach 1 DPU per
DRAM bank. It means 16 cores per 8Gbit DRAM chip. On a 16Gbyte DIMM, we deliver
256 cores, and 8 of them can be added to a standard CPU socket. We end up with a
co-processing system of 2048 cores together with 128Gbytes of DRAM per socket."
The PIM chip, integrating UPMEM's proprietary RISC processors (DRAM
Processing Units, DPUs) and main memory (DRAM), is the building block of the
first efficient, scalable and programmable acceleration solution for big data
applications. Associated with its Software Development Kit, the UPMEM PIM
solution can accelerate data-intensive applications in the datacenter servers 20
times, with close to zero additional energy premium.
"We are no
longer in an era were CPUs and other hardware getting continuously faster would
mask the slow speed of inefficient software," said Reza Malekzadeh, General
Partner at Partech Ventures (among the investors). "UPMEM's solution
addresses the performance needs of modern scale-out applications while
preserving datacenter and infrastructure hardware investments."
Editor's comments:- As a fan of
in assessing new technologies - on
"A simple way to understand the kind of application
opportunities and limits of Upmem's solution is to look at the ratio of CPU
cores to GB of DRAM. That gives you the power envelope and tells you what
problems it's best suited for. The articles
linked on Upmem's web site are very informative as far as they go."
this announcement the spectrum of
SSD processing solutions in the market had ranged in latency and benefit
The memoryfication of the enterprise and the aspiration
more within memory systems (which will lead to storage systems being an
emulation in memory and the
of the AFA as we know it) is being driven by
storage applications for big apps (as described in a slides by Parallel
Machines in February
- adding user deployable API and RAM in the flash controller (NxGn - which exited stealth
PS - "The first time I suggested to a processor
design team that they should look at adding support for solid state storage in
their new CPUs instead of just adding more cores was about 2000. I got the
response at that time - what's an SSD? And nothing more came of the matter."
- from the blog -
CPUs for use with SSDs in the Post Modernist Era of SSD and Memory Systems
Gen-Z memory fabric demo at FMS
Editor:- August 8,
of a relatively new ORG -
the Gen-Z Consortium ran
multi-vendor demonstrations this week at
FMS achieving 112GT/s.
to Gen-Z's faqs page
- the idea is to create a high bandwidth, low latency, standard for
memory-like data transfers which are media independent and can "scale
from tens to several hundred GB/s of bandwidth with sub-100ns load-to-use memory
What's a comparable context?
If you go back
in time to 2000 and think about the past but forward looking potential of
Infiniband or back
in time to about 2013 with
- it's maybe a bit like like those were in their time - but now we're looking
from a 2017 competitive needs analysis and the memoryfication of the
datacenter- so it needs to be faster.
IC Insights reports record breaking memory ASPs
July 20, 2017 - A recent
note about the memory market by IC
Insights puts an interesting spotlight on memory shipments.
other things IC Insights says:- "DRAM, unit shipments are actually
forecast to show a decline this year. Moreover, NAND shipments are forecast to
increase only 2%."
When it comes to price expectations IC
Insights says this.
"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." ...read
Editor's comments:- One message to take away from this is
that as memories have been transitioning to the next multiple of 3D layers the
chip throughput from the industry's legacy wafer fabs has stayed the same or
gone backwards due to the
extra time taken to
reliably make those extra layers to create higher bit density memories.
more blogs re 3DX
Editor:- June 6, 2017 - Some
recent blogs about Micron's
waiting on 3DXP arrays - by Woody Hutsell suggests
that 3DX is among the aspiring inheritors of the modern
space for latency sensitive customers which historically was once
associated with RAM SSDs.
Woody says - "I think the early usage for 3DXP will flow largely
to server vendors (and their suppliers)."
But he goes on to say
why he thinks that adoption in storage arrays will be slower and at a lower
scale than flash.
new memory patent for Corsair
Editor:- May 18,
2017 - News about the issuance of new patents in the SSD and memory market
- Corsair -
- described by its inventor Bobby Kinstle Senior
Project Manager in this way. "This one is for using tiny heat pipes to
remove heat from memory devices in really tight spaces."
Rambus and Microsoft extend research on cold DRAM
April 20, 2017 - Back in the early 1990s it was not uncommon to hear about
specialist server companies which were using peltier effect heat sinks to
refrigerate the fastest workstation processors so that they could be run at
higher clock speeds. But this kind of extreme approach to server acceleration
only provided short term competitive gains in a single dimension.
of the biggest bottlenecks in the past decade has been RAM architecture and
DRAM implementation itself.
A new angle on extending the performance of
recently by Rambus
and Microsoft who are collaborating on the design of prototype super cooled
DRAM systems to explore avenues of improvement in latency and density due to
physics effects below -180 °C.
A new article -
Heat Up With Cold DRAM - by Junko Yoshida
, Chief International Correspondent - EE Times - discusses these plan in
In the article - Craig Hampel,
chief scientist at Rambus, told EE Times that "Microsoft isn't alone...
heavy data center users like Google, Facebook and Amazon are all in search of
new memory architecture. Indeed, these tech giants who have primarily grown
their business via their technological prowess in software development are now
finding the future of their business growth severely constrained by hardware
advancements." ...read the article
comments:- At room temperature the main problem in DRAM systems is that in
fast clocked systems the energy required for refresh cooks the chips which
means cells lose charge faster which creates data integrity risks which in turn
needs more frequent refresh. So even if you have a miraculous packaging
technique which can sandwich more chips into a box - DRAM loses out to other
memory technologies which don't require refresh - when the scale of the
installed capacity in the box is high.
If you freeze DRAM then the
refresh cycle can get longer (which means you can pack more capacity in a box)
but also the native transit time for data in the copper interconnects and
silicon gets faster too.
Although Rambus and Microsoft are pitching
this a progressive research exercise I disagree that it will provide a general
solution for data intensive factories.
While it's a good thing for
researchers to play around and explore the limits of what can be done with all
kinds of memory devices - I think that the answer to greater performance lies in
new architectures rather than freezing old ones.
low yield at sub 20nm is root of DDR4 shortage says DRAMeXchange
April 14, 2017 - Quality problems in
DRAMs which have been
sampling this year at the new sub 20nm generation from major suppliers is at
the heart of the issues discussed in a new -
view blog by DRAMeXchange - which concludes that the contract prrice
of 4GB DDR4 DRAM modules will rise 12.5% entering 2Q17.
research director of DRAMeXchange said - "PC-OEMs that have been
negotiating their second-quarter memory contracts initially expected the market
supply to expand because Samsung
and Micron have
begun to produce on the 18nm and the 17nm processes, respectively. However both
Samsung and Micron have encountered setbacks related to sampling and yield, so
the supply situation remains tight..." ...read the
inside SSD pricing,
storage market research
2017 will be crossover revenue year for DDR4 says IC Insights
April 13, 2017 - A new
about the DRAM market by
- DDR4 prices in 2016 fell to nearly the same ASP as DDR3 DRAMsAs a result,
IC Insights expects DDR4 to become the dominant DRAM generation in 2017 with 58%
market share versus 39% for DDR3.
- Following a year of extraordinary gains in pricing, a boost to DRAM supply
in the second half of 2017 could lead to reduced ASPs and the inevitable start
of a cyclical slowdown in the DRAM market.
BeSang says 3D Super-DRAM could fix multi-billion dollar money
pit of memory industry's fab capacity roadmap
15, 2017 - Just as we're starting to get used to a world view that memory
not be enough to make all the memory parts needed - and that a pragmatic
global optimization from the user point of view may be to plan ahead for
memory systems which use tiering, flash as RAM, freshly minted shiny
nvms and new
SSD aware software to get
more storage and processing
done with less
chips - a journey which - depending who you are - begins or ends with the
idea of reducing
the ratio of DRAM to storage - and just as we're getting our heads adjusted
to the huge investments which would be needed to make DRAM technology
better and to believe that no sane investor (not even a
VC who loves SSDs)
would want to toss their money in that direction - a seemingly different
alternate get out of jail free option is offered in a new blog by Sang-Yun Lee,
CEO - BeSang -
in EE Times -
Among other things Sang says...
you consider planar DRAM shrinking from 18nm to 16nm, then, 20% more
dice-per-wafer could be achieved. To do so, multi-billion dollar should be
invested for R&D and
is required. In case of 3D Super-DRAM, it needs less than $50 million for R&D
and no EUV; and even so, it could produce 400% more die-per-wafer."
at the risk of repeating some of that:- 4x as much DRAM from the same fabs
without huge investments... How is that possible? ...read
Editor's comments:- You can get an idea of
the complex decision matrices facing memory makers. In past decades the product
types which determined the demand mix for memories (PCs, phones, servers) were
few in number and had predictable roadmaps. Now big demands for memory are
coming from cloud, IoT and new intelligence based markets which are creating
entirely new ratios and rules of what is possible with memory systems.
the RAM memory mix - 5 years back, 5 years forward -
February 22, 2017:- In
February 2012 -
that the percentage of its enterprise SSD systems which were pure RAM SSDs had
declined to 10%. And 45% of the systems it was shipping (at that time) were
all flash arrays.
That was a useful way of assessing progress in the
succession of flash
in the enterprise over the original
RAM SSD market.
the perspective of 2017 we now see of course that what was good for storage
(capacity and IOPS) is
good too for
latency - as flash has started replacing DRAM as random access memory in
high capacity RAM systems ranging from single servers to multiple racks.
because the low energy requirement of nvms (which don't need gas guzzling
refresh) means you can fit more raw memory capacity into a single motherboard.
And even the higher raw access times of flash (compared to DRAM) look good in
comparison to box hopping fabrics.
(And other cost savings kick in
One day in the future on this page we will be reporting when DRAM
(in external chips and DIMMs) has reached the point where it is only 10% of all
native main memory too.
How do banks use big memory systems to detect and prevent fraud?
January 9, 2017 - In the early 2000s I started hearing stories from vendors
about how their fast memory systems were helping banks to not only ease the
choke points in their transactions but also provide insights into fraud
A new white paper GridGain
Systems provides a good introduction and synthesis of
various roles of in-memory computing in accelerating financial fraud detection
and prevention (pdf) which includes many named bank examples.
This paper describes how in memory computing provides the low latency data
sharing backbone which is needed to enable pattern detection for fradulent
activity to be assessed in real-time while at the same time enabling genuine
transactions to proceed quicky.
Among other things, the paper says...
move from disk to memory is a key factor in improving performance. However,
simply moving to memory is not sufficient to guarantee the extremely high memory
processing speeds needed at the enterprise level... Clients who have implemented
the GridGain In-Memory Data Fabric to detect and prevent fraud in their
transactions have found that they can process those transactions about 1,000
times faster." ...read
the article (pdf)
Virtium announces 64GB very low profile industrial DDR4 RAM
December 13, 2016 - Rated at industrial temperatures Virtium today
the imminent availability (in January) of 64GB DRAM modules - VLP RDIMM and
Mini-RDIMM - which have been developed to provide high-performance memory to
a different approach to 3D SCM?
29, 2016 - The different semiconductor technology approaches to storage class
memory of 3 large hopefuls in the market (WD,
Intel / Micron) are compared
and contrasted to a different tunneling approach which is claimed to provide
greater endurance - in a recent blog -
Mechanical Advantage: A Revolution through Evolution for Storage Class Memory
by Andrew Walker,
Founder and CEO of Schiltron.
Andrew says his company's approach to 3-D memory is "designed to wring
every ounce of advantage out of Quantum Mechanical (QM) tunneling." ...read
Rambus and Xilinx partner on FPGA in DRAM array technology
October 4, 2016 - Rambus
a license agreement with Xilinx
that covers Rambus' patented memory controller, SerDes and security
Rambus is also exploring the use of Xilinx FPGAs in its
Data Acceleration research program. The SDA - powered by an FPGA paired
with 24 DIMMS - offers high DRAM memory densities and has potential uses as a
CPU offload agent (in-situ
can memory do more?
Editor:- June 17, 2016 - When
all storage is memory based - are there new design techniques which can push
back the boundaries of what memory can do?
That was the catalyst for my new blog on StorageSearch.com -
Should we set
higher expectations for memory systems?
All the marketing noise
coming from the DIMM wars market (flash as RAM and Optane etc) obscures some
important underlying strategic and philosophical questions about the future of
SSD. Can we think of software as a heat pump to manage the entropy of
memory arrays? (Nature of the memory - not just the heat of its data.)
we be asking more from memory systems? ...read the blog
We just made the first tri-state DRAM chip in the world
June 2, 2016 - In his recent linkedin note -
chip gonna rock the DRAM industry - Wayne Zhang,
President and CEO at Encrip
enthuses about his company's new tri-state DRAM technology which he says
can work with any standard process - even 10nm.
Wayne says - "For
the same capacity DRAM chip, with using our patented technologies, we could
reduce the memory array area up to 36%, we could reduce the power consumption up
to 40%, we could also increase the chip access speed." ...read
DIMM wars at battery scale - FLC from Marvell
May 12, 2016 - When thinking about
SSD / SCM DIMM
wars - most of the buzz in the past year has been focused on the impacts of
replacing DRAM with
flash at the
server and cloud
levels. But the same concepts can be applied (albeit with different
efficiency gains) at
the implementation level of battery powered embedded devices and wearables.
a recent blog -
Marvell FLC Redefines Main Memory - by Hunglin Hsu, VP
provides authoritative examples of the replacement ratios possible in a phone
A strategic lesson to guide future designers is that even
while getting a 50% power consumption reduction (due to flash as RAM) it
is also feasible to increase application performance at the same time because
the software can work with a larger memory capacity (due to the lower
cost of flash
Among other things Hunglin says - "With FLC, better
performance can be achieved by reporting to the operating system a larger than
physically implemented main memory. The operating system is thus less likely to
kill background apps, which is why the fast app switching is possible. The FLC
hardware does all the heavy lifting in the background and frees up the tasks of
the operating system." ...read
DRAM's indeterminate latencies and the virtual memory slider
Editor:- March 2, 2016 - in the new page blog on StorageSearch.com I cast an eye
on the latency specific defects in DRAM system behavior which are among
the many technology enablers of the emerging tiered memory / flash as RAM
We've been accustomed to think of DRAM as the simple
predictable latency memory (compared to
flash). But server
motherboard memory system latency hasn't improved for over 10 years. Memory
systems got bigger and bandwidth got faster but worst case latencies can
sometimes be worse than they used to be - due to interference effects caused
by complex data queuing patterns.
If you haven't noticed these
problems - congratulations!
It means you might not notice (or care)
when the virtual memory slider moves in the cheaper direction towards
memories like flash. ...read the
retiring and retiering RAM
Editor:- December 3, 2015-
Retiring and retiering RAM is one of the ideas discussed in my new home page
blog on StorageSearch.com -
the big SSD
ideas of 2015.
Netlist allies with Samsung to codevelop flash-as-RAM DIMMs
November 19, 2015 -
how it's going to enter the storage class memory SSD DIMM wars market. This by
way of a 5 year joint development and license with Samsung which also
brings to the table $23 million of funding. The companies expect to sample
products in 2016.
2015 was a
signficant kick-start year for the server memory market.
retiering enterprise DRAM was one of the three big SSD ideas of the year.
DIMM wars in
SSD servers how significant is Memory1?
3D X-Point could shrink DRAM market by 1/3 in 5 years- says
Editor:- October 23 , 2015 - Coughlin Associates
has recently published a new
report on Emerging Non-Volatile Memory and Spin Logic (163 pages,
$4,000). The memories addressed in this report
(pdf) include PRAM, RRAM, MRAM, STT MRAM as well as the recently announced
3D X-Point Technology.
3D X-Point Technology will have a big impact
on DRAM growth (with DRAM
sales down $6.7 billion to $15.6 billion due to XPoint by 2020) with XPoint
revenues of $663 million to $1.5 billion by 2020.
MRAM and STT MRAM
revenue is estimated at $1.4 billion to $3.2 billion by 2020. Manufacturing
equipment revenue for MRAM and STT MRAM production is estimated to be between
$159 million and $294 million by 2020. See also:-
market research news,
Adaptive Dynamic Refresh in DRAM
Editor:- October 14,
2015 - I expected most of the practical iinovations in
architecture to come from the enterprise market.
But there's an
interesting exception from Green
Mountain Semiconductor which is revealed in a new paper -
DRAM for High-reliability IoT, Automotive and Control System Applications (pdf).
designs memories for industrial.
embedded and custom
systems. The innovation discussed in their paper is the use of adaptive dynamic
refresh as a collaborative technology with ECC which can react to ECC errors by
tuning the refresh rate.
GMS says the strategy is - "Increase refresh rate if too many fails and
reduce rate if too few fails, always guaranteeing refesh rate mimics cell fail
distribution. Self-calibrating system, no need for tightly calibrated
the article (pdf)
SanDisk and HP ally in SCM DIMM wars
October 9, 2015 - SanDisk
and HP yesterday
a long-term partnership to collaborate on a new technology within the Storage
Class Memory category.
The companies say it will center around HP's
Memristor and SanDisk's ReRAM memory technology and manufacturing and design
Editor's comments:- In the summer
Micron established the
precedent that it's now OK to talk about futuristic memory roadmap intentions
as long as they include a big dollop of memory types which are less well known
that flash - because most of the press and business analysts treat it with just
as much seriousness as if you were talking about something which you can ship
This is part of the pre-shooting, phoney war about how the
industry is going to phase in a new level of big memory which from the software
point of view has similar R/W characteristics to RAM - but which from the
capacity point of view - is closer to flash than it is to DRAM. And in
competitive terms will work better than existing memory types in some types of
applications and not at all well in others.
SanDisk already has a good
view of the possibilities in this market via its ZetaScale software - which
provides big data RAM virtualization using any type of flash SSD. And
conversations with customers of its memory channel storage codeveloped with
Diablo - must have reinforced SanDisk's confidence in new uses for DIMMs.
(Although SanDisk's ULLtraDIMM is a flash based SSD - which can't do byte writes
in the same way that Memory1 can.)
So... what could HP bring to
this party for SanDisk?
You need a friendly bios
and platform and routes to market when you're trying to launch a new
Memristor? - This press release had to include
some kind of technology input from HP to make them feel better. If you had said
"toner cartridges" instead it would have been just as deliverable
today - except that everyone knows the printers are now going in a different
direction. Maybe the draft press release did have toner cartridges as the
placeholder and they just slipped that memory jibber jabber in at the last
minute before pressing send.
- "Low power is at the center of HP's ReRAM technology. HP's
presentation pointed out that a lot of the time and energy of computation is
used by the OS moving data between the various levels of the memory hierarchy
of existing computer architectures." - ReRAM
Forum (July 2014)
- "We're the world's largest purchaser of DRAM and the second largest
buyer of flash and (with Memristors) we're trying to disrupt and re-arrange our
supply chain" - said HP - reported in the article -
to replace flash and SSD in 2013 (October 2011) on Electronics Weekly
nvRAMs - state of the semiconductor market readiness
September 17, 2015 -
Gaps in the
memory hierarchy have created openings for new types of memory is a new blog
LaPedus, Executive Editor - Semiconductor
Engineering - which is flavored with some strong opinions from
Mark says - "after numerous delays, a new wave of
next-generation, nonvolatile memories are finally here. One technology, 3D NAND,
is shipping and gaining steam. And 3 others - Magnetoresistive RAM, ReRAM and
even carbon nanotube RAMs - are suddenly in the mix." ...read the article
Diablo aims to shrink enterprise DRAM market with flash as memory
August 12, 2015 - Diablo
has launched an assault on the enterprise server RAM market with the launch
of a new DRAM compatible emulation memory module called "Memory1"
which replaces DDR-4 DRAM
- with byte addressable flash.
In a new blog on StorageSearch.com
- DIMM wars in
SSD servers - how significant is Diablo's Memory1? I discuss the
potential impact of this technology and make some guesses about how Diablo
has managed to replace DRAM with flash. ...read the
data compression techniques in memory systems
May 26, 2015 - Inside the
SSD controller brain
the compressibility of data is one of the tools which can go into the mix of
optimizing performance, endurance and competitive cost.
A recent paper
Survey Of Architectural Approaches for Data Compression in Cache and Main Memory
Systems by Sparsh Mittal
and Jeffrey S. Vetter in
IEEE Transactions on Parallel and
Distributed Systems - reviews the published techniques available and places
their relevance in the context of real and future memory types and applications.
The survey covers applications from embedded systems upto supercomputers. In
addition to being useful resource directory of related papers the article
gives you a brief description of many compression techniques, where you might
use them and what benefits you might expect.
list of articles and
books by Sparsh Mittal which among other things covers caching
techniques, reliability impacts and energy saving possibilities in a wide
range of server architectures.
flash backed DIMMs - a new directory from StorageSearch.com
October 21, 2014 - Although StorageSearch.com
has been writing about flash
backed DRAM DIMMs since the first products appeared in the market - I didn't
think that subject was important enough before to rate a specific article or
market timeline page. (Unlike
SSDs - which became 1 of the top 10
viewed by readers after having had its own directory page since
However, sometimes a
market is defined as
much by what it isn't as by what it is.
And so - to help clarify the
differences between these 2 types of similar looking storage devices (one of
which I think is much more significant than the other - but both of which are
important for their respective customers) I have today created a directory
page for hybrid DIMMs
etc - which will act as the future pivoting point for further related
Samsung in volume production of 3D DDR4 RDIMMs
August 27, 2014 - Although the main interest in DDR4 RDIMMs - from an SSD
market perspective - will be in how that interface opportunity gets leveraged
channel flash SSDs - let's not forget that the motherboard slots - which
will enable that market - have been designed for
DRAM. So the DRAMs will
come first and are an important part of the countdown to the new DDR4 flash DIMM
In that context I'd like to mention that Samsung is today
celebrating "a new milestone in the history of memory technology" with
the announcement that the company is in volume production of the
industry's first 64GB DDR4
RDIMMs (DRAM) that use 3D "through silicon via" (TSV) stacked die
package technology and 20nm class die geometries.
Samsung says that
the new 64GB TSV module performs 2x as fast as a 64GB module that uses
conventional wire bonding packaging, while consuming approximately 1/2 the
Editor's comments:- Samsung describes this announcement as "historic"
and I was content to include that positioning statement in the news above -
because much of what Samsung has done in the past has indeed had historic
For more examples - see "Samsung
historic" which gives you search results from the
AgigA Tech samples 1st DDR4 NVDIMM
6, 2014 - AgigA Tech
today announced that it is now sampling the industry's first DDR4 Nonvolatile
DIMM (NVDIMM) to key OEMs and development partners.
is there a market for I'M Intelligent Memory inside SSDs?
June 4, 2014 - Are there applications in the SSD market for DRAM chips which
integrate ECC correction inside the RAM chip - and which plug into standard
That was the question put to me this afternoon by Thorsten
Wronski - whose company MEMPHIS
Electronic AG distributes I'M Intelligent Memory in
Thorsten told me he's had a good reaction from the SSD
companies he's spoken to - which is why he phoned.
But in a long
conversation about the economics and architectures of end to end
in SSDs and the different
ratios of RAM cache
to flash in SSDs - I told him that my initial reaction was he should look
at embedded applications - which depend on the
reliability of a
single SSD - rather than enterprise systems in which the economics analysis for
arrays point to a system wide solution rather than a point product fix.
interesting thing is he said he's done tests on the new I'M memory as drop in
replacements for unprotected memory designs- in which he accelerated the likely
incidence of error events by increasing the interval between refreshes and
raising the temperature.
Here's what he said.
assembled a standard 1GB unbuffered DIMM with 8 chips of 1Gbit ECC DRAM. Then we
put this into a test board and ran RSTPro (a very strong memory test software).
No error found.
Next we put the whole board into a temperature chamber
at 95°C, which normally requires the refresh rate to be doubled (32mS
instead of 64mS). No error found.
Finally we wrote a software to change the refresh-register of the CPU
on the board, so we were able to set higher values. The highest possible was
750mS, so the DRAM did almost not get any more refreshes. Still it continued
working in RSTPro without a single error for 24 hours.
We tried the same with Samsung and Hynix modules, but none of them
came even close to those results. Most failed at refresh-rates of 150 to 200 mS,
which is not bad indeed. Many more tests will follow."
Editor's comments:- the reason I mention this - is because
adapting the refresh rate was one of the things mentioned in my recent blog -
Are you ready to
However - most of the leading SSDs in
don't have RAM caches for other reasons (to reduce the physical space, power
consumption, hold-up time, or because don't need the performance). So I told
Thorsten I don't see an industry wide demand inside SSDs. But some of you
might already have thought of applications.
I'M ECC DRAM product
big memory makers still need to recoup investments in flash and
DRAM before switching to newer technologies
Editor:- May 22, 2014 -
today published a new article -
Big Memory Shift
Ahead - which looks at the cost, scaling and other technical pressures on
legacy nand flash and DRAM which are the object of attack by other memory
"If these new memories really are as good as the
claims, why are we not seeing them in production applications today?" -
says Brian Bailey,
Technology Editor - Semiconductor
Engineering who wrote the article. "The answer appears to be inertia."
CEO of Intermolecular
expands on this by suggesting - "NAND flash is a $30B industry that has
tens of billions of dollars in capital infrastructure that would need to be
retooled. The big 4 players represent 95% of the market and they have a lot of
existing investment. The entire cost equation is CapEx, so they need to milk the
tail of the revenues as long as they can." ...read the article
Are you ready to rethink RAM?
Editor:- April 2, 2014
- We've all got used to the idea that a
revolutions has been playing out in the enterprise server market centered
around flash SSDs and - in that context - the developments in DRAM
technology have sounded reassuringly boring and predictable.
are you ready to
rethink enterprise DRAM architecture too?
The state of blue sky
thinking about enterprise DRAM - what is it really for? - and the changes
that could lead to - are discussed in a new blog on StorageSearch.com ...read the article
Samsung in volume production of 20nm DDR3 RAM
March 11, 2014 - 40 years ago in the
days of MOS LSI - whenever semiconductor companies like
Intel wanted to
characterize a new semiconductor production process and establish the "safe"
design rules for manufacturability at ever smaller chip geometries (aka "shrink")
the circuit and product of choice for the fab architects was memory - even if
the eventual product for the wafer fab was going to be a microprocessor.
recently - in the past few years - if you've been looking at all those "nm"
(nanometer) numbers in the news stories about IT related chips you can
hardly fail to have noticed that it's been the flash memory devices which have
been at the leading edge of the numbers. And when looking at production
devices - flash has been about 2 years in advance of DRAM and server CPUs.
often heard on these pages that it's only by
safe design rules used in preceding generations that interesting new SSDs
come to market.
And a big part of the to-do list for any
SSD controllers is to
cope with a predictable scale and style of expected memory defects and
virtualize them away - creating a usable base level storage device.
in line with this trend - Samsung today
it's using 20nm technology in the production of new 4Gb DDR3
way this pattern has been going in recent years is that the first volume uses
of new silicon geometries go into consumer markets - where if there's a data
upset - you can see something wrong happening (blue screen or freeze) turn the
power off and try things again. After several quarters of doing this - the chip
bakeries have finely tuned their recipes and are ready to guarantee a less
crumbly dough mixture for use in the enterprise.
- Samsung was doing volume production of 10nm flash (used in consumer
eMMC SSDs for mobile
phones) in November
But if these concepts
are new to you - it's not worthwhile memorizing them. Because
3D nand flash changes the
priority of future enhancements towards a preference for building upwards
in more layers instead of merely thinning sideways.
Hybrid Memory Cube gets x2 speedup
25, 2014 - Although the market for Hybrid Memory Cube compatible
RAM has barely begun - a
new Gen2 specification was
today which doubles the fastest short-reach data performance (previously
15Gb/s) upto 30Gb/s. See also:-
Micron samples 2GB HMC
Editor:- September 25, 2013 -
it's sampling the company's first implementation of the
Hybrid Memory Cube (high
density chip stacking architecture
standard) which was
launched in October
(short reach) HMC provides 2GB DRAM in a BGA - with upto 160 GB/s
DRAM technology won't advance soon - says Micron
August 20, 2013 - In recent years the SSD market has become nearly 100%
flash (and nv memory)
focused - with little or no mention of
DRAM based SSDs. The
reason is that nearly every company whose product line used to be mainstream
RAM SSD - has pulled out of that market or discontinued enhancements to those
products. Flash SSDs are more economic and easier to sell.
mean to say that the role of DRAM in SSD systems has entirely disappeared. It
still appears as a cache
or tier in many flash SSD arrays and the existence of some small percentage
of DRAM is assumed in SSD
and also in (flash based)
sent out a useful signal of where its own DRAM roadmap is going in
article yesterday in EETimes - which reports an interview with
Micron's president Mark
Adams who said - "There will be no new greenfield DRAM fabs for the
foreseeable future. We are hitting something of a lithography wall in DRAM where
shrinks are getting tougher and gains are not as attractive, so people are not
as financially motivated to invest in new fabs. Also we see planar DRAM advances
will end in the next 3 to 5 years, so you probably cannot get ROI in a new
planar fab." ...read
SMART samples ULLtraDIMM SSDs
8, 2013 - SMART
Storage Systems today
it has begun sampling the first memory channel SSDs compatible with the
interface and reference architecture created by Diablo Technologies.
first generation enterprise
(ULL = ultra-low latency) can be deployed via any existing DIMM slot and
provides 200GB or 400GB of enterprise class flash SSD memory with upto 1GB/s and
760MB/s of sustained read/write performance, with 5 microseconds write latency.
Throughput, IOPS and memory capacity all scale with the number of ULLtraDIMM
deployed in each server.
comments:- With the current design -only one DIMM slot in each server has to
be reserved for conventional DRAM. Apart from that constraint any DIMM slot can
be used for either flash or DRAM as deemed necessary for the application.
more about the potential of this technology, the thinking behind it and the
competitive landscape relative to
PCIe SSDs etc see my
earlier articles on the
Hybrid Memory Cube spec ready for chip designers
April 3, 2013 - back in
October 2011 - I
reported on this page the formation of a new industry
ORG - the Hybrid Memory Cube Consortium
- which could have an impact on future SSD packaging densities.
takes a while to get these things going - but according to
press release this week by one of the founding companies - Micron - the 100 plus
companies which are collaborating in this enterprise have agreed on an
A key feature of the new multiplane memory
architecture is that distributed memory controllers in an HMC module will
handle the data I/O packet requests for the bunch of stacked memory chips in its
own vault. This is similar to the distributed intelligent data mover concept
which is already used in all proprietary
SSD controller designs - because it's the only way you can get good
aggregated global system performance while also dealing with low level
local memory management issues at low latency.
As with earlier
generations of remote distributed memory interfaces - such as
InfiniBand - HMC is
designed to optimize the request of small packets - which in the case of HMC is
16 to 128 bytes of data.
With today's semiconductor speeds -
accessing the data in those distributed memory chips within the same HMC module
presents similar technical problems to distributed memory cards in traditional
computer designs - because traversing inches of physical space at high speed is
as difficult as moving data across tens of feet at slower speeds.
has been born as a DRAM
technology - but don't ignore it - just for that reason. (Or because the data
packet sizes are small compared to the block sizes in
nand flash.) If and when
these HMC packaging ideas result in viable products - the ideas and
methodologies will spill into SSDs too -regardless of what the underlying
memories used in SSDs may be at that time.
It's all about speed and
scalability. According to the HMC
faqs page - A single (1st generation) HMC unit can provide more than 15x
the bandwidth of a DDR3 module. See also:-
SSD interface glue chips.
Micron sources power holdup technology for NVDIMMs
November 14, 2012 - Micron
has signed an
agreement with AgigA Tech to
collaborate to develop and offer nonvolatile DIMM (NVDIMM) products using
AgigA's PowerGEM (sudden power loss
controller and holdup modules).
Viking ships nv 8GB DDR3 DIMM
Editor:- October 18,
Viking Modular Solutions
said it is shipping
extension of their nv module range.
ArxCis-NV plugs into standard
RAM sockets and provides
2GB to 8GB RAM which is backed up to SLC flash in the
event of a
power failure - while the memory power is held up by an optional external
25F supercap pack. Viking says these new memory modules can eliminate the need
for battery backup units in servers and the maintenance logistics associated
with maintaining them. They are specified as being maintenance free for "5
years @ 60°C".
Editor's comments:- will these new
modules replace batteries in
RAM SSDs? - I doubt it
- because of scalability issues - like managing a spiderweb of 100+ dangly
bits of wire when you have a terabyte of RAM. Having said that - there are many
applications which only use a small number of memory chips which could benefit
from such a product.
Hybrid Memory Cube will enable Petabyte SSDs
October 7, 2011 - Samsung
and Micron this
week launched an new industry initiative - the Hybrid Memory Cube Consortium
- which will standardize a new module architecture for memory chips -
enabling greater density, faster bandwidth and lower power.
is unlike anything currently on the radar," said Robert Feurle,
Micron's VP for DRAM Marketing. "HMC brings a new level of capability to
memory that provides exponential performance and efficiency gains that will
redefine the future of memory."
Editor's comments:- HMC
may enable SSD designers to pack 10x more
RAM capacity into the same
space with upto 15x the bandwidth, while using 1/3 the power due
to its integrated power management plane.
The same technology will
enable denser flash SSDs too - if flash is still around in 3 years' time and
hasn't been sucked into the obsolete market slime pit by the
lurking nv demons
which have been shadowing flash for the past 10 years and been waiting for each
"next generation" to stumble and be the last.
management architecture integrated in HMC and the density scaling it allows
for packing memory chips (without heat build-up) are key technology enablers
which were listed as some of the problems the SSD industry needed to solve
in my 2010 article -
this way to the
| "Across the whole
enterprise - a single petabyte of SSD with new software could replace 10 to
50 petabytes of raw legacy HDD storage and still enable all the apps to run
much faster while being hosted on a shrunken population of SSD enhanced
|meet Ken and the
enterprise SSD software event horizon|
|Partial list of past and present RAM
manufacturers - mentioned in
storage news /
AGIGA Tech ,
Apacer Memory America,
ATP Electronics, Austin Semiconductor,
Avant North America,
Micro Memory Bank,
Ramaxel Technology ,
Rocky Mountain Ram,
Memory, Silicon Power
Micro Systems, Spansion,
Viking Modular Solutions,
White Electronic Designs,
Winbond Electronics ,
Z Tech International.
|when you need the fastest access|
|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. |
of the 2017 memory shortages|
|No longer a commodity: new
applications and new memories have created a more customized, system approach to
memory solutions which have de-commoditized the DRAM and NAND markets. Memory
manufacturers must offer more options and carry a more diversified product
Lane... Far from a Leisurely Stroll (February 2016)|
|"We propose the design
and an implementation of a bulk parallel external memory priority queue to take
advantage of both shared memory parallelism and high external memory transfer
speeds to parallel disks. Our experimental results show that in the selected
benchmarks the priority queue reaches 64% of the full parallel I/O bandwidth of
SSDs and 49% of rotational disks, or the speed of sorting in external memory
when bounded by computation."|
|A Bulk-Parallel Priority Queue in
External Memory with STXXL (pdf) (2015)|
|"Among 129 DRAM
modules we analyzed (comprising 972 DRAM chips), we discovered disturbance
errors in 110 modules (836 chips). In particular, all modules manufactured in
the past two years (2012 and 2013) were vulnerable, which implies that the
appearance of disturbance errors in the field is a relatively recent phenomenon
affecting more advanced generations of process technology. We show that it takes
as few as 139K reads to a DRAM address (more generally, to a DRAM row) to induce
a disturbance error."|
Errors in DRAM - an Experimental Study (pdf) (2014)|
| "Our past work
showed that application-unaware design of memory controllers, and in particular
memory scheduling algorithms, leads to uncontrolled interference of applications
in the memory system." - said Onur Mutlu, Carnegie Mellon University.|
|Are you ready to
|RAM in a historic market
by Zsolt Kerekes,
|RAM - Random Access Memory
- is the fastest type of storage.|
It's implemented by silicon chips
which can contain upto thousands of millions of storage bits (gigabits)
connected in a randomly accessible array.
The "random access"
part of the RAM name was to differentiate RAM from other early types of
memory which had different interface characteristics. (Mostly block based - and
sometimes with the blocks having device dependent R/W timing and location
RAM was easier to write software for - even if it was
more expensive. So it became the preferred standard for software.
has equal (symmetric)
read and write access times (unlike
flash memory). Other
significant differences to flash are:-
- the data stored in a RAM is only maintained while the device is powered
up (is volatile)
- RAM doesn't suffer from write wear-out (endurance)
RAM products have different
designs and are optimized for various markets (such as servers, notebooks and
graphics cache) based on their speed, cost, interface and capacity.
- RAM is typically more expensive than flash for the same capacity, and
typically uses more electrical power. The exception is smaller capacity memories
inside a chip where the complexity of managing flash memory incurs more overhead
than the much simpler overheads in RAM.
has been the dominant type of random access memory from 1972 until today
(2015). That was because it offered lower cost and higher storage capacity than
competing memory types which had similar R/W characteristics.
period 2003 to 2015
many new types of randomly writable memory emerged (which used different on
chip storage characteristics). At various times these alternative memory types
claimed they would displace DRAM or flash. But they had no impact on the
revenue of the 2 dominant semiuconductor memory types.
If there was a
threat to DRAM upto 2015 - it didn't come from physical random access memory
Instead important lessons learned from virtualization
techniques (mostly arising from the
PCIe SSD market
experience from around
indicated that in large datacenters - a proportion of DRAM could be
functionally replaced by flash.
The intrinsic defficiencies of flash
(block write rather than byte write, and slower writes than reads) were capable
of being statistically offset most of the time by adding a new latency tier in
As flash cells had already surpassed DRAM cells
in terms of how many data bytes could be packed in a chip - this was an
approach which promised lower systems costs.
But it wouldn't provide
satisfactory performance for all applications and even when it could do so -
it would require significant investments in rewriting software.
the close of
entering a new age in the enterprise DRAM market in which 2 things will change
in the next 5 years.
- flash will continue to displace a proportion of DRAM in the enterprise
due to virtualization efforts and new
SSD software and also
by plug compative devices.
- some of the alternative physical random access memories are again being
hyped (by their advocates) as being nearly in a market readiness state to
compete with DRAM in server sockets on a cost basis.
news archive 2000 to 2015
circuit RAM timeline 1961 to 1998 (pdf)
|"...The RAM market faces
disruptive challenges from SSDs
- just as hard disks
have done. At some time during the next 5 years - most of the world's new
memory will be deployed inside an SSD or an SSD controlled loop. Owning an
SSD brand will be as important in the new market for memory makers as getting
designed into tier 1 server slots was in the past. Commercial RAM makers will
have to re-engineer themselves into SSD companies - or risk lower profit
margins from selling to SSD brands at spot market prices from outside the SSD
|...Editor:- talking to a market strategist in one
of the world's biggest seminconductor companies in June 2011.|
|"I'm often asked if I could foresee
how important (DRAM) would become," said Robert Dennard
(the inventor of DRAM) on receiving a lifetime achievement award in 2005.|
"I knew it was going to be a big thing, but I didn't know it would grow to
have the wide impact that it has today." ...read the