The NVM Insider, Issue 1 - Page 3 - Outside Thoughts

Outside Thoughts:  The Changing Face of Embedded Memory

Rich Wawrzyniak, Senior Analyst ASICs and SoC, Semico

 One of the underlying facts of the current System-on-a-Chip (SoC) market is that the use of embedded memory, in all its different variations, is growing. There is not a single SoC in the market today that does not employ some form of embedded memory at some density level whether it is for main memory, data cache, security or other uses. In fact, the ITRS roadmap from 2005 shows the die area given over to embedded memory content for the typical SoC reaching approximately 94% by 2014. While this may be overly optimistic (especially considering the current trend toward employing potentially 100’s of CPU cores for multicore applications), it is reasonable to expect the die area dedicated to embedded memory on many parts to exceed at least 70% in this time frame.

One of the major reasons for this growth is the increase in device functionality and the richer feature sets that are possible in devices that approach and exceed 100 million gates. As this device complexity rises, so too, must the memory density increase to support the new levels of functionality found in these silicon solutions. Previously, the answer to support these higher levels of performance would have been to simply increase the density of the external memory in the system. Now however, with shrinking power budgets and form factors and the constantly increasing pace of device integration, discrete memory is becoming less of a viable option for many end applications.

The other interesting trend is regarding the rise in the usage of non-volatile embedded memory.

Up to around 12 years ago, the embedded memory found on most SoCs consisted of SRAM and mask ROM (MROM). A few parts used DRAM with MROM. There was very little usage of other non-volatile embedded memory types until fairly recently – about 6 to 7 years ago. The densities for non-volatile embedded memory were always small compared to the volatile embedded memory alternatives and mostly used for security purposes. This has started to change as NOR-Flash is now available in 65nm process geometries, but densities still remain relatively low compared to discrete devices. This is mostly due to issues with die area inefficiency and the fact that the charge pump used for Flash does not scale well as geometries migrate downward.

Given the desirability of having embedded non-volatile memory as part of the silicon solution for many of the new, emerging portable applications, yet considering the issues with scalability and die area inefficiency, the search for the perfect embedded memory has yielded some interesting results. A great deal of research and development has been done in the areas of Magnetic RAM (MRAM), Phase Change Memory, Spintronics, Carbon Nanotubes, etc. All these memory variants possess the capability of providing relatively good density combined with non volatility. In addition, MRAM, Spintronics and Carbon Nanotubes have acceptable access times. However, they all possess at least one undesirable feature: circuit complexity, and in the case of Phase Change Memory, a wear-out mechanism. These are not very desirable features for the ‘universal memory’ the industry has been searching for over the last 30 years.

A New Voice is Heard

Why then, is a relatively obscure announcement from HP Labs in Palo Alto, CA cause for interest? The announcement in question concerns something innocuously dubbed the ‘Memristor’, a circuit that ‘remembers’ changes in the current that passes through it by changing its resistance – hence the term memristor.

Professor Leon Chua, while at the University of California, Berkeley back in 1971, mathematically predicted there should be a fourth type of electronic component to complement the resistor, capacitor and inductor. This was named the memristor for memory-resistor. While theoretically possible, the predicted behavior of the ‘memory resistor’ was never experimentally observed to any great degree or understood as such. That all changed with the announcement by Senior Fellow R. Stanley Williams at HP Labs at the end of April, 2008.

Essentially, Mr. Williams and his group at HP had been looking for materials that exhibited the ‘memristic’ effect for use in HP’s nanowire crossbar switches by examining different materials at nanoscale geometries. Now they have found it.

The memristor created by HP consists of a sandwich of nanowires made out of Titanium Oxide that, in its pure stare, is highly resistive. Dopants of other materials added to the TiO2 can make it highly conductive. In a high electric field, the dopants are not stationary and drift in the direction of the current. Putting the dopants on one side of the TiO2 then applying the current causes them to move to the other side which is pure TiO2, and lowers the resistance. Reversal of the current causes the dopants to reverse direction and increases resistance, setting the stage for 1’s and 0’s to be created.

Advantages of Memristors

• Non Volatile
• Large density possible
• Relatively fast access times at <50ns
• Very simple circuit topology
• Scales with process geometry
• No extra control circuitry required (unlike Flash)
• No exotic materials required
• No wear-out mechanism detected so far
• Relatively low power
• Can be used for Logic circuits and, potentially, for Analog circuits as well

Why is This Important?

More and more SoCs are finding their way into portable and mobile devices where the addition of non-volatile memory is desirable to accommodate ‘instant-on’ functions, or to add other desirable capabilities – programmability being one of these. These devices are hampered today by the inability to instantiate large amounts of non-volatile memory without incurring yield issues or being very die-area inefficient which can cause cost issues. A memory type that exhibits the features listed above becomes very attractive when considering all the additional functionality it could bring to ASIC and SoC designers.

The key to all of this will be how manufacturable the memristor is and how fast will it ultimately be able to function? We are at the very beginning in the life of the memristor and the search for other materials to use besides Titanium Oxide has just started. It is very possible that different materials will yield faster switching times and become even more useful to designers.

Two other important points to mention are that a) this material dose not seem to have a wear-out mechanism unlike Phase Change memory and b) it can be used in the same manufacturing process to create logic functions and even perhaps analog functions. This can be a rather powerful combination when one considers that even today, there are trade offs necessary between fabbing parts with embedded memory in either a Logic process or in a Memory process. Life just became much simpler for both ASIC designers and process engineers.

To be fair, the industry has heard claims regarding the discovery of the ‘universal’ memory before and there always seems to be some flaw that prevents the newest claimant from living up to its advance billing. Certainly, it’s the early days for the memristor and much additional research into, and characterization of, this material remains to be undertaken before anyone can say “mission accomplished.” However, the promise of this material can be truly great and can have a tremendous impact on the embedded memory market if it turns this promise into actual capabilities.

This is definitely one area to keep an eye on in the future.