Kristu Jyoti College of Management and Technology, Changanacherry 1

Intel’s 3D XPoint
A revolutionary breakthrough in Memory Technology

Jose Jestin George, MCA-S5, Kristu Jyoti College of Management and Technology

 tube filled with mercury and plugged at each end with a quartz
Abstract— The explosion of connected devises and digital crystal, delay lines could store bits of information in the form
services is generating massive amount of new data. For this data of sound waves propagating through mercury, with the quartz
to be useful, it must be stored and analyzed very quickly. 3D crystals acting as transducers to read and write bits. Delay line
XPoint technology is an extensively new class of non-volatile memory would be limited to a capacity of up to a few hundred
memory technology that can help turn immense amount of data
thousand bits to remain efficient.
into valuable information in real time. With up to 1000 times
lower latency and exponentially greater endurance than NAND.
3D XPoint technology can deliver game changing performance Two alternatives to the delay line, the William’s tube and
for big data applications and transactional workloads. Its ability Selectron tube, originated in 1946, both using electron beams
to enable high speed, high capacity data storage close to the in glass tubes as means of storage. Using cathode ray tubes,
processor creates new possibilities for system architects and Fred Williams would invent the Williams tube, which would
promises to enable new applications. be the first random-access computer memory. The William’s
3D XPoint technology innovative , transistor less cross point
tube would prove more capacious than the Selectron tube (the
architecture creates a three dimensional check board where
memory cells at the intersection of word lines and bit lines Selectron was limited to 256 bits, while the Williams tube
allowing the cells to be addressed individually. As a result, data could store thousands) and less expensive. The Williams tube
can be written and read in small sizes, leading to fast and would nevertheless prove to be frustratingly sensitive to
efficient read / write processes. environmental disturbances.

Index Terms— 3D XPoint, NAND, Flash Memory. Efforts began in the late 1940s to find non-volatile memory.
Jay Forrester, Jan A. Rajchman and An Wang developed
magnetic-core memory, which allowed for recall of memory
I. INTRODUCTION after power loss. Magnetic core memory would become the
In computing, memory refers to the computer dominant form of memory until the development of transistor-
hardware integrated circuits that store information for based memory in the late 1960s.
immediate use in a computer; it is synonymous with the term
"primary storage". Computer memory operates at a high Developments in technology and economies of scale have
speed, for example random-access memory (RAM), as a made possible so-called Very Large Memory (VLM)
distinction from storage that provides slow-to- computers.
access information but offers higher capacities. If needed,
contents of the computer memory can be transferred II. CURRENT MEMORY TECHNOLOGIES
to secondary storage, through a memory management
technique called "virtual memory". An archaic synonym for The term "memory", meaning "primary storage" or "main
memory is store. memory", is often associated with addressable semiconductor
memory, i.e. integrated circuits consisting of silicon-
based transistors, used for example as primary storage but also
In the early 1940s, memory technology often permitted a
other purposes in computers and
capacity of a few bytes. The first electronic programmable
other digital electronic devices.
digital computer, the ENIAC, using thousands of octal-base
radio vacuum tubes, could perform simple calculations Most semiconductor memory is organized into memory
involving 20 numbers of ten decimal digits which were held in cells or bistable flip-flops, each storing one bit (0 or 1). Flash
the vacuum tube accumulators. memory organization includes both one bit per memory cell
and multiple bits per cell (called MLC, Multiple Level Cell).
The next significant advance in computer memory came The memory cells are grouped into words of fixed word
with acoustic delay line memory, developed by J. Presper length, for example 1, 2, 4, 8, 16, 32, 64 or 128 bit. Each word
Eckert in the early 1940s. Through the construction of a glass can be accessed by a binary address of N bit, making it
possible to store 2 raised by N words in the memory. This
implies that processor registers normally are not considered as
 Kristu Jyoti College of Management and Technology, Changanacherry 2

memory, since they only store one word and do not include an Disk memory is what holds all of our files and programs when
addressing mechanism. Typical secondary storage devices not in use. It is the memory we are all most familiar with.
are hard disk drives and solid-state drives. We want to design computers that can store lots of data AND
operate very fast. However, when we build memory storage
There are two main kinds of semiconductor
devices, there is a trade-off between speed and memory. Either
memory, volatile and non-volatile.
you can build very large memory, like your spinning disk hard
drive, or very fast memory, like the registers in your CPU.
A. Volatile memory Memory Hierarchy lets us have the best of both worlds - speed
and size. You have a small amount of ultra-fast memory, a
Volatile memory is computer storage that only maintains its
larger amount of slower memory, and a huge amount of very
data while the device is powered. Most RAM (random access
memory) used for primary storage in personal computers is slow memory. By cleverly choosing what data to store in
volatile memory. RAM is much faster to read from and write which type of memory, we can appear to have a huge amount
of very fast memory.
to than the other kinds of storage in a computer, such as
the hard disk or removable media. However, the data in RAM Chip designers have to know a ton about how to pass data
stays there only while the computer is running; when the from registers to caches, from cache to main memory, and
computer is shut off, RAM loses its data. from main memory to the hard disk. Low-level programmers
Volatile memory contrasts with non-volatile memory, which need to be aware of how memory management works if they
does not lose content when power is lost. Non-volatile want to write programs that manipulate lots of data quickly.
memory has a continuous source of power and does not need Memory hierarchy describes each level of computer storage
to have its memory content periodically refreshed. Examples by response time. For example your RAM is faster to access
of non-volatile memory are flash memory (used as secondary than your hard drive so it is above it in memory hierarchy. It is
(memory) and ROM, PROM, EPROM and EEPROM memory important to note that capacity and complexity are intricately
(used for storing firmware such as BIOS). related.

B. Non-volatile memory
A. Internal register:
Non-volatile memory (NVM) is a type of computer memory
that has the capability to hold saved data even if the power is Internal register in a CPU is used for holding variables and
turned off. Unlike volatile memory, NVM does not require its temporary results. Internal registers have a very small
memory data to be periodically refreshed. It is commonly used storage; however they can be accessed instantly. Accessing
for secondary storage or long-term consistent storage. data from the internal register is the fastest way to access
memory.
Non-volatile memory is highly popular among digital media;
it is widely used in memory chips for USB memory sticks and
digital cameras. Non-volatile memory eradicates the need for
relatively slow types of secondary storage systems, including
hard disks.Non-volatile memory is also known as non-volatile
storage.
Examples of volatile memory are primary storage, which is
typically dynamic random-access memory (DRAM), and
fast CPU cache memory, which is typically static random-
access memory (SRAM) that is fast but energy-consuming,
offering lower memory areal density than DRAM.

III. MEMORY PYRAMID

Memory hierarchy a concept that is necessary for the CPU

to be able to manipulate data. This is because it is only able to
get instructions from cache memory. Cache memory is located
on the processor chip, and is the fastest kind of memory. As a
result of this, it is also the smallest, meaning that we cant hold
all of our processes in it at once. So we use RAM. RAM is
much larger than cache memory, but not located directly on
the CPU, so it is slower. Instructions are loaded into RAM
until the CPU needs them. They are only held as long as
power is supplied, and this is why we need Disk memory. Figure: Memory Hierarchy in Computer.
 Kristu Jyoti College of Management and Technology, Changanacherry 3

B. Cache memory: D. Hard disk:

A hard disk is a hardware component in a computer. Data is
The basic purpose of cache memory is to kept permanently in this memory. Memory from hard disk is
store program instructions that are frequently re-referenced not directly accessed by the CPU, hence it is slower. As
by software during operation. Fast access to these instructions compared with RAM, hard disk is cheaper per bit.
increases the overall speed of the software program. As the
microprocessor processes data, it looks first in the cache E. Magnetic tape:
memory; if it finds the instructions there (from a previous
reading of data), it does not have to do a more time-consuming Magnetic tape memory is usually used for backing up large
reading of data from larger memory or other data. When the system needs to access a tape, it is first
data storage devices. Most programs use very few resources mounted to access the data. When the data is accessed, it is
once they have been opened and operated for a time, mainly then unmounted. The memory access time is slower in
because frequently re-referenced instructions tend to magnetic tape and it usually takes few minutes to access a
be cached. This explains why measurements tape.
of system performance in computers with
slower processors but larger caches tend to be faster than IV. INTRODUCTION TO 3D XPOINT
measurements of system performance in computers with faster
processors but more limited cache space. Multi-tier or 3D XPoint is memory storage technology jointly developed
multilevel caching has become popular by Intel and Micron Technology Inc. The two vendors have
in server and desktop architectures, with different levels described this new technology as filling a gap in the storage
providing greater efficiency through managed tiering. Simply market between dynamic RAM (DRAM) and NAND flash.
put, the less frequently access is made to certain data or The current mainstream memory technologieshave been
instructions, the lower down the cache level the data or around for decades. While the cell designs have evolved over
instructions are written. the years to allow scaling to 20nm and below, the fundamental
physics behind DRAM and NAND operation haven't changed
C. Main Memory a bit and both technologies have their unique technological
limitations. DRAM offers nanosecond-level latency and
unlimited endurance, but this comes at the cost of large cell
The memory unit that communicates directly within the
size, cell volatility, and power consumption. Since DRAM
CPU, Auxiliary memory and Cache memory, is called main
cells need to be constantly refreshed, the cells don't retain data
memory. It is the central storage unit of the computer system.
in an off state, requiring quite a bit of power and making
It is a large and fast memory used to store data during
DRAM unsuitable for permanent storage. NAND, on the other
computer operations. Main memory is made up
hand, has much higher latency (especially write operations)
of RAM and ROM, with RAM integrated circuit chips holing
and has a limited number of write cycles, but the cells are non-
the major share.
volatile and the structure is much more efficient, enabling low
cost and suitability for storage.
RAM: Random Access Memory
Combining DRAM and NAND at the system-level
DRAM: Dynamic RAM, is made of capacitors and architecture provides the best of both worlds, which is why
transistors, and must be refreshed every 10~100 ms. It is modern computers use DRAM as a memory/cache and NAND
slower and cheaper than SRAM. for storage. However, there's still a latency and capacity gap
between DRAM and NAND, so the question arises: what if
SRAM: Static RAM, has a six transistor circuit in each cell you were to combine the best of DRAM and NAND at the
and retains data, until powered off. silicon level? The mission of next generation memory
technology across the industry has been to develop a new type
NVRAM: Non-Volatile RAM, retains its data, even when of memory that provides low latency and high endurance
turned off. Example: Flash memory. while offering a small and scalable cell size.

We have seen numerous startups, such as Crossbar and

ROM: Read Only Memory, is non-volatile and is more like a Nantero, discuss and demonstrate their next generation
permanent storage for information. It also stores the bootstrap memory technologies, but we have yet to see the established
loader program, to load and start the operating system when DRAM and NAND vendors come out with their solutions.
computer is turned on. PROM (Programmable Intel and Micron are here to change that with the
ROM), EPROM (Erasable PROM) and EEPROM (Electrically announcement of their new 3D XPoint (Cross Point) non-
Erasable PROM) are some commonly used ROMs. volatile memory technology this week.
 Kristu Jyoti College of Management and Technology, Changanacherry 4

V. 3D XPOINT ARCHETECTURE VI. HOW 3D XPOINT MEMORY WORKS

The 3D XPoint technology innovative, transistor-less cross 3D XPoint has a different architecture from other flash
point architecture creates a three-dimensional checkerboard products. It's reputed to be based on phase-change
where memory cells sit at the intersection of word lines and bit memory technology, with a transistor-less, cross-point
lines, allowing the cells to be addressed individually. As a architecture that positions selectors and memory cells at the
result, data can be written and read in small sizes, leading to intersection of perpendicular wires. Those cells, made of an
fast and efficient read/write processes. unspecified material, can be accessed individually by a current
sent through the top and bottom wires touching each cell. To
improve storage density, the 3D XPoint cells can be stacked in
three dimensions.

Detailed descriptions from Intel of how Optane works are still

notable by their absence—the company seems to have said
more about what Optane isn't than what it is—but a basic
picture is slowly being built from what Intel and Micron have
said about the technology. The memory has a kind of three-
dimensional (hence "3D") lattice structure (hence "XPoint").
Stackable layers have wires arranged in either rows or
columns, and at the intersection of each row and column is the
actual storage element: an unspecified material that can
change its resistance to different values. The details of how it
Cross Point Array Structure: does this are unclear; Intel has said it's not a phase-change
Perpendicular conductors connect 128 billion densely material, and it's different from HP's memristor tech, but it
packed memory cells. Each memory cell stores a single bit of hasn't said precisely what it is.
data. This compact structure results in high performance and
high density. The value stored in each data cell can also be written and
rewritten relatively easily. NAND flash requires a very high
Stackable: voltage to erase each cell, which allows a cell to be written
The initial technology stores 128Gb per die across two only once (flipping its value from a 1 to a 0) before it needs to
stacked memory layers. Future generations of this technology be erased again. 3D XPoint cells, by contrast, can have their
can increase the number of memory layers and/or use resistance (and hence stored value) updated between 1 and 0
traditional lithographic pitch scaling to increase die capacity. and back again without needing any erasure step.

Selector: Each cell stores a single piece of data, making a cell represent
Memory cells are written or read by varying the amount of either a 1 or a 0 through a bulk property change in the cell
voltage sent to each selector. This eliminates the need for material, which modifies the cell's resistance level. The cell
transistors, increasing capacity and reducing cost. can occupy either a high- or low-resistance state, or changing
the resistance level of the cell changes whether the cell is read
Fast Switching Cell: as a 1 or a 0. Because the cells are persistent, they hold their
With a small cell size, fast switching selector, low-latency values indefinitely, even when there is a power loss.
cross point array, and fast write algorithm, the cell is able to
switch states faster than any existing nonvolatile memory
Read and write operations occur by varying the amount of
technologies today.
voltage sent to each selector. For write operations, a specific
voltage is sent through the wires around a cell and selector.
In their 2015 announcement of the technology, Intel and
This activates the selector and enables voltage through to the
Micron claimed 3D XPoint would be up to 1,000 times faster
cell to initiate the bulk property change. For read operations, a
and have up to 1,000 times more endurance than NAND flash,
different voltage is sent through to determine whether the cell
and have 10 times the storage density of conventional
is in a high- or low-resistance state.
memory. Early products are faster and more durable than
NAND and denser than conventional memory, but they
haven't lived up to the full extent of the vendors' claims. 3D XPoint has the ability to write data at a bit level, an
advantage over NAND. All the bits in a NAND flash block
must be erased before data can be written. In theory, this
capability enables 3D XPoint to have higher performance and
lower power consumption than NAND flash.

Critically, the resistance change is persistent. Once a cell has

had its value set, it'll continue to hold its value indefinitely,
 Kristu Jyoti College of Management and Technology, Changanacherry 5

even if system power is removed. While we don't know how C. Use cases
the resistance change works, one thing that we do know is that
unlike DRAM, each data cell does not need any transistors, 3D XPoint is used as an additional layer of storage between
which gives rise to Optane's next important property: it's a lot flash and DRAM. It's a relatively common practice to tier
denser than DRAM, with Intel and Micron variously claiming storage between hard disk drives (HDDs) and flash. High-
a density improvement of four to ten times. intensity data and applications that benefit more from high
speeds are stored on the flash layer, while data and
VII. FEATURES OF 3D XPOINT applications that are accessed less frequently are put on disk.
3D XPoint is another layer of storage above flash for data and
A. Speed and performance
applications that need even greater speeds.
Intel expects the 3D XPoint Optane SSD will be used for
With the 3D XPoint architecture, data no longer has to be high-performance storage and caching, as well as to extend
stored in 4 KB blocks using a slow, file I/O stack. The new and replace memory. According to the company's projections,
technology enables small amounts of data to be written and users will be able to increase server memory by as much as
read, making the read/write process faster and more efficient eight times and displace DRAM by as much as a 10:1 ratio for
than NAND. Initial products using the 3D XPoint technology select workloads.
bear this out, though not at the speed and performance levels
Intel and Micron promised when they rolled out the
technology. D. Extensions:

While not as fast as DRAM, 3D XPoint has the advantage of Intel has provided three ways to extend memory with 3D
being nonvolatile memory. From a performance and price XPoint Optane SSDs:
standpoint, 3D XPoint technology falls between fast, but
costly DRAM and slower, cheaper NAND flash. • via an operating system paging mechanism that moves data
out to the PCIe-attached SSD when DRAM fills for a
According to Intel, the P4800X drive performed five to eight workload;
times faster than the company's NAND flash-based DC P3700 • via optimized applications; or
in internal tests at low queue depths using a mixed workload. • via Intel's Memory Drive Technology supported on its Xeon
The P4800X can reach as much as 500,000 IOPS -- or processors.
approximately 2 GBps -- at a queue depth of 11, Intel claimed.
Observers have speculated that the PCI Express (PCIe) bus In the future, it will be possible to extend memory with the
used by the P4800X is holding it back from the promised 3D XPoint DIMMs that Intel plans to release. Observers
speed of 1,000 times faster than NAND. Other system changes speculate that 3D XPoint Optane, and particularly Optane
thought to be needed for the 3D XPoint technology to meet NVDIMMs, will be used to:
higher performance goals include segregating persistent from
nonpersistent memory when handling machine check errors • expand the apparent size of DRAM;
and using a compiler that enables persistent memory to be • enable bigger, more-effective databases;
declared, along with using link editors that can build that • help overcome big data network bottlenecks;
memory into an application. The applications themselves must • facilitate high-performance computing applications;
be rewritten to eliminate file I/O and to use single instruction • extend memory and boost instance storage performance in
and vector operations. the cloud;
• provide the storage capacity and speed that hybrid clouds
Nonvolatile 3D XPoint dual in-line memory modules need; and
(DIMMs) that fit into DRAM slots and use the double data • possibly serve as primary memory tiers in hyper-converged
rate bus also may help 3D XPoint reach its full performance systems
potential.
E. Future possibilities:
B. Cost:
As of August 2017, the 375 GB Optane P4800X add-in card 3D XPoint drives excel at servicing random, transactional
is priced at $1,520, or $4.05 per gigabyte. By comparison, data sets that are not optimized for in-memory processing. For
Intel's 400 GB flash-based NVMe PCIe P3700 SSD is $879, businesses that rely on complex, random analytics, 3D XPoint
or approximately $2.20 per gigabyte. drives would be useful for performing limited real-time
analytics on current data sets or for storing and updating
Intel Optane memory for PCs is $44 for a 16 GB module and records in real time.
$79 for a 32 GB module. Intel clearly sees a broad range of other analytical uses for
its implementation of 3D XPoint. According to the company’s
Intel Optane web pages, the advance memory could be used
by retailers to more quickly identify fraud detection patterns,
 Kristu Jyoti College of Management and Technology, Changanacherry 6

by financial institutions to accelerate trading, or by healthcare X. DISADVANTAGES

researchers to work with even larger data sets in real-time.[2] According to the latest source, consumers will have to hold
Of course, none of this means NAND flash will be going away their excitement of experiencing extremely high read and
any time soon. In fact, NAND flash will continue to be useful write speeds, because 3D XPoint based SSDs might not be
for storing data that can be sequentially processed overnight. able to achieve those speeds due to interface limitations.
But for time-critical analytics, 3D XPoint drives will take their
place alongside DRAM as part of a cost-effective continuum Since SATA interfaces are unable to deliver the speeds that
of data processing options PCI Express or DD4 interfaces are able to deliver, 3D XPoint
based storage products are going to be released in the
VIII. APPLICATIONS aforementioned interfaces. At this point, PCI Express 3.0 x4
The Optane SSD Series is designed to enable data centers to can only offer bandwidth of up to 4GB/s. In contrast,
accelerate applications for fast caching and storage, increase NVDIMMs are able to deliver potentially higher bandwidth
scale per server, and reduce transaction costs for latency- and ultra-low latency.
sensitive workloads. Companies can also process larger data
sets to gain new insights that were unaffordable or impractical Another important thing is that this technology is in research
with DRAMs. Intel is planning to release additional Optane level. Performance and hardware limitations can be seen after
SSD options for both data centers and consumers over the next implementation.
several months.
The Intel Optane memory module can be used to accelerate XI. CONCLUSION
any SATA storage device installed in a 7th-generation Intel 3D XPoint is a promising form of non-volatile memory
Core processor-based platform that is designated as “Intel jointly developed by Intel and Micron. Intel claims that the
Optane memory ready.” The Optane add-in memory module memory, which it's branding Optane for commercial products,
acts as a cache to increase performance in laptops and provides a compelling mix of properties putting it somewhere
desktops. According to Intel, the Intel Optane memory PC between DRAM and NAND flash.
accelerator module provides up to 2x faster boot, 5x faster Intel clearly sees a broad range of other analytical uses for its
web browser launch, and 67% faster game launch. implementation of 3D XPoint. According to the company’s
The adoption rate for 3D XPoint memory will likely be Intel Optane web pages, the advance memory could be used
slow initially because the technology requires servers built on by retailers to more quickly identify fraud detection patterns,
newer processor chipsets. In addition, applications need to be by financial institutions to accelerate trading, or by healthcare
re-written to fully support 3D XPoint. But as enterprises researchers to work with even larger data sets in real-time.
slowly modernize their servers and software, 3D XPoint drives Of course, none of this means NAND flash will be going
and memory will inevitably become a fixture in the data away any time soon. In fact, NAND flash will continue to be
center. useful for storing data that can be sequentially processed
overnight. But for time-critical analytics, 3D XPoint drives
IX. ADVANTAGES will take their place alongside DRAM as part of a cost-
effective continuum of data processing options.
3D XPoint technology is unique because it combines the
density and non-volatility of NAND SSDs with speeds near REFERENCES
that of DRAM. That positions 3D XPoint drives into a new [1] "Intel Launches Optane Memory M.2 Cache SSDs For Consumer
storage tier, between volatile memory and non-volatile flash Market". AnandTech. 27 March 2017. Retrieved 13 November 2017.
SSDs. [2] McGrath, Dylan (28 Oct 2009), "Intel, Numonyx claim phase-change
memory milestone", www.eetimes.com
[3] Neale, Ron (14 Aug 2015), "Imagining What's Inside 3D
Intel Optane SSDs aren’t yet at a speed or price-point that XPoint", www.eetimes.com
will convince companies to swap out all their server DRAM [4] "3D XPoint™ Technology Revolutionizes Storage
for 3D XPoint technology. But savvy IT admins will be able Memory", www.youtube.com (video, infomercial), Intel
to reduce costs by replacing some DRAM with Intel Optane [5] Smith, Ryan (18 August 2015), "Intel Announces Optane Storage Brand
For 3D XPoint Products", www.anandtech.com
SSDs, and by using the new drives to enhance the [6] Peter Bright (March 19, 2017). "Intel's first Optane SSD: 375GB that
performance of their existing flash-based SSDs. you can also use as RAM". Ars Technica. Retrieved March 31, 2017.

According to Intel, 3D XPoint technology delivers:

 Up to 10x more performance than NAND over a

PCIe NVMe interface.
 About 8x to 10x greater density than DRAM.
 Exceptionally high endurance, compared to NAND
SSDs.
 Latency measured in nanoseconds (close to DRAM)
as opposed to microseconds (SSDs) or milliseconds
(HDDs)