/ more pages
Princeton's in-memory accelerator uses switched caps to take
neural matrix MACs out of the digital box
Editor:- November 22,
2018 - In memory computing acceleration is one of those architectural ideas
which - continuing in the footsteps of the
era of SSDs - now looks well placed to be the focus of creating
which leverage the new loci of latency unbottlenecking opportunities
revealed by the incremental memoryfication of data processing.
recent years we've seen advanced design groups leveraging algorithm aware
approaches to neural network processors which enable them to usefully deploy low
resolution MACs (such as Google's TPU).
And stretching this idea
outside the binary box some hybrid implementations have mixed analog and
digital computing elements.
For example the ReRAM ML engines
mentioned in 2 news
stories in May 2018.
Similar to that vein of thinking - but using
different underlying technology - researchers at
Princeton University recently
published a paper (a Microprocessor
implemented in 65nm CMOS with Configurable and Bit-scalable Accelerator for
Programmable In-memory Computing (abstract) /
which describes a new type of in-memory-computing accelerator which is
integrated in a programmable processor architecture which uses charge domain
compute elements built around 8 bit resolution ADCs and switched capacitors to
perform multiplication and accumulation.
The authors say their 65nm
CMOS reference design of a 590kb in-memory-computing accelerator shows an
energy efficiency of 152/297 1b-TOPS/W and throughput of 4.7/1.9 1b-TOPS
(scaling linearly with the matrix/input-vector element precisions) at VDD of
the paper (pdf)
In a related non technical interview article
by Princeton communications -
memory and computation, programmable chip speeds AI, slashes power use -
one of the researchers - Naveen Verma
- Associate Professor of Electrical Engineering - says (among other things)...
computing has been showing a lot of promise in recent years, in really
addressing the energy and speed of computing systems. But the big question has
been whether that promise would scale and be usable by system designers towards
all of the AI applications we really care about. That makes programmability
CPUs for use with SSDs
are we ready for
infinitely faster RAM?
Modular showcases 96GB Gen-Z Memory Module
Editor:- November 12,
2018 - SMART Modular
it is showcasing its Gen-Z Memory Module this month at
SC18 in Dallas.
ZMM demo unit
comes in a PCIe HHHL form factor. It operates at data rates up to 40GB/s and
with less than 400ns load-to-use memory capacity of 96GB of DRAM. The ZMM uses
(PCIe bridge) fabric memory controller which supports up to 384GB of DRAM.
Spin Memory announces new agreements related to its MRAM
November 11, 2018 - Spin
Memory (formerly called Spin Transfer Technologies) today
a $52 million Series B funding round and an agreement with
Applied Materials to create a
comprehensive embedded MRAM solution and also a commercial agreement with
Arm for the licensing of
Memory's Endurance Engine design IP to address SRAM applications in SoCs.
processors in in the Post Modernist Memoryfication Era
university researchers disclose delusion of security blanket
delivered by popular self encrypting SSDs
Editor:- November 8,
2018 - Researchers at Radboud University, in
the Netherlands have recently
a paper - self-encrypting
deception: weaknesses in the encryption of SSDs (pdf).
conclude - "we found that many hardware implementations of full disk
encryption in SSDs have critical security weaknesses, for many models allowing
for complete recovery of the data without knowledge of any secret... This
challenges the view that hardware encryption is preferable over software
encryption. We conclude that one should not rely solely on hardware encryption
offered by SSDs."
After discovering the ineffectiveness of the
protection offered by the encryption hardware in the models affected - the
authors alerted the authorities and agreed to wait 6 month before publishing
their paper so as to enable the manufacturers - whose products had been tested
(Micron and Samsung confirmed the vulnerability issues) to work on firmware
The paper also examines commonly held assumptions about the
security of encryption engines in various states when analyzed from particular
attack directions. ...read the
Editor's comments:- as with all
aspects of SSD design
- the correlation between what the customer expects and what does actually
happen in the real world depends on the quality of the designers and the
verification process. That's why it's not unusual in mission critical projects
SSDs for the customer verification process to take longer in elapsed time
than it took to design the original SSD.
fast purge SSDs,
Why consumers should
expect to see more flaky SSDs (2009)
Everspin announces new fab deal to make more MRAM
November 5, 2018 - Everspin
it would augment the production capacity for its Toggle MRAM by means of a
multi-year partnership to use the fabs of SilTerra
(Malaysia based semiconductor manufacturing services company). Initial
production is expected to start in calendar year 2020. Everspin said it will
continue in parallel to operate its manufacturing line for Toggle MRAM in
Understanding Macro Memory Market Trends
Moore's Law is
history - Newton's 3rd may be better
Editor:- November 1, 2018
- Although memory cell behavior is best understood by referring to quantum
physics there seems to be a role for Newtonian Mechanics in the
memory markets - especially when it comes to patent disuputes and tariff
I refer of course to that well known phrase I learned at
ever action there is an equal and opposite reaction."
earlier versions of this SSD news page - now in
- the simmerings of the memory patent and tariff disputes between the US and
China went something like this...
The next development on this tariff ping pong is
reported in a story (October 30, 2018) in EETimes -
US Bans Exports
to Chinese DRAM Maker - about Fujian Jinhua - which has been declared a
threat to US national security.
- July 2018 -
Micron's memory sales in China were the object of a preliminary injunction
obtained by UMC.
EETimes says - "The move means
that U.S. firms need a license for all exports, re-exports, and transfers of
commodities and software to Jinhua. Such license applications will be reviewed
with a presumption of denial..."...read the article
re Micron's "soft-announcement" of a DRAM replacement
Editor:- October 30, 2018 - A new article -
Divorce - Micron and Intel See Different Futures - by William
Tidwell - who is a regular memory market commentator on
SeekingAlpha.com - discusses among
other things:- the need for in-memory processing and signs that
Micron is working on a
new SCM memory architecture.
Re the interdependency of memory and
processor architecture Tidwell says - "...memory in today's system
architecture is locked in a fatal embrace with the CPU. It is stranded."
as part of a multi-page analysis and detailed speculative look ahead as to
what Micron will do with the fabs it gets from buying out Intel's stake in the
joint IMFT venture - Tidwell says - "Under promise, over deliver - is the
right strategy for Micron in the wake of Intel's 3DXPoint misadventures. I
believe this is what they are doing and that gives investors reason to hope that
the New Memory will be commercialized in FY '21." ...read
what's RAM really?
what's Metadata-as-a-Service got to with SSDs?
October 24, 2018 - There's a small number of new companies which attract my
attention because of the impact which one or more their founders has had on the
SSD market in the
So although it's not strictly speaking an SSD company - and I'm counting down
the weeks to my retirement in 2019 - and trying not to start too many new
strands of content which I can't complete tidily - I was very interested to see
these stories about a new company called Hammerspace
which provides a sequel to the question of what happened next to one of the
esteemed co-founders of Fusion-io and Primary Data -
that person being David Flynn
who has probably done more work on figuring out and managing the intricacies of
data and the latency and repurposing cost-benefits of legacy and new dynasty
architectures scrambled in the archeological mix of big data infrastructuer than
anyone else on the planet. Here are the links.
also:- SSDs in the cloud,
re SSD everywhere software,
Violin acquires X-IO Storage
Editor:- October 16,
2018 - Violin Systems
it has signed a letter of intent with X-IO Technologies to acquire
X-IO Storage, the
division of X-IO Technologies that developed the Intelligent Storage Element
(ISE) flash and hybrid storage array product lines. Financial terms were not
acquisitions in the
storage arrays and flash caches - timeline 2010 to 2017
$100 million AI investment fund launched by Micron
October 10, 2018 - Micron
that it will invest up to $100 million in startups with a strong focus on AI and
machine learning through its strategic investments entity, Micron Ventures.
And in a related announcement, the Micron Foundation launched a $1 million grant
for universities and non-profit organizations to conduct research on AI.
also:- VCs in SSDs
value of infinitely faster RAM?
joins the memory latency to do list
after AFAs -
what's the next box? - cloud adapted memory systems
Megabyte went through
| no more Red Hat jokes -
Novembers of yore in SSD market history|
||Solid Data Systems
discussed the advantages of SAN SSD accelerators in the article -
Solid State File-Caching for
Performance and Scalability.|
||Texas Memory Systems
announced the industry's first NAS SSD for enterprise acceleration.|
produced a 64Mb ferroelectric RAM chip using Ramtron's patented FRAM
Intel announced an
agreement to up a new jointly owned company IM Flash Technologies.|
Texas Memory Systems
demonstrated its first native
rackmount SSD accelerator.
M-Systems which had
been a pioneer in the use of MLC in SSDs. |
||Google opened its doors to
developers who wanted to work with a new operating system
Chrome OS which was
designed at the outset around solid state storage and which specifically
excluded the paradigm of hard
published a paper which examined the possible future advantages of using arrays
of storage class memory nodes as a cost effective archiving solution compared to
hard drives and tape.|
||First mention of
Diablo Technologies on
StorageSearch.com with news of a $28 million funding round.|
||NetApp underwhelmed SSD analysts with the
announcement that it had shipped 59PB of SSDs in the past 3 years.|
||Steve Wozniak rejoined the old "band"
of the enterprise flash accelerator songbook writers from Fusion-io in the
newly emerging Primary Data.|
||Netlist revealed how it was going to enter the
storage class memory SSD DIMM wars market by way of a 5 year joint development
||Elastifile announced it had been granted a US
patent for a method of managing flash-aware adaptive data management in
cloud scale meshes. |
||IntelliProp demonstrated a memory controller for
the emerging Gen-Z memory fabric.|
I like the idea that "fastest" claims are still commenting on Violin Systems - October 3, 2018
hard currency in