It's about eventual density. It's presumable, that these memristors will someday be at least as small as individual transistors on silicon. Try and emulate that functionality in the silicon space of one transistor. Remember, these are VERY new. They will come down in cost, and they will shrink, in time. The first LED cost $200 EACH, back in the 1960s. It's the same story here. It costs as much as an Intel CPU because the parts are large, cumbersome, have extra redundancy to account for yield losses, and each part is individually probed, conditioned, and tested, and then bonded based on the yield results... It's an expensive process. It's also very, very new. People look at these and often consider the digital switching aspect, but they have an analog range that can be tapped with a very low current compliance. You can use these to create very intriguing parts like neurons, that have their "weight" shift with subsequent firings, and not have to emulate that characteristic digitally. There are more abstract applications where the analog features are a benefit. Pattern recognition is just one area where these parts could excel. They bring to the table the degree of fuzzy processing that neurons excel at, and they do it without emulation, using efficient discrete hardware, with switching compliance currents measuring in the nano and micro amp range, at under a volt.
There is also the fact that they are
non volatile, storing their analog state indefinitely, without power. Even flash loses it's data integrity after a certain period of time. These won't lose their data. Emulate that digitally, without requiring a power source.
Further communication with the company reveals nothing surprising. They roll custom silicon, and their biggest early application has been ASICs for the US DoD. They are also working on developing a memristor based FPGA like device. DoD work is nothing new... The government is typically first to fund new tech, and to push it's limits. The same was true with the early transistors and early integrated circuits as well.
Memory is the "easiest" solution. You get fast, one device memory cells with great speed, retention, and non-volatility. A single element, that will someday presumably be no larger than an on die transistor, able to hold a state, without power, with the density of flash and the speed of DRAM. The ability to perform logic, at the active memory, that also doubles as storage... That's a huge deal.
One proposed concept, would have a massive FPGA like memristor matrix that can reconfigure it's logic, so you essentially write a free area of your unified working/storage memory to perform a pipelined logic function, that is non volatile (by nature of memristors), that simply feeds your initial memory down a pipeline, to be calculated and stored at the end of the pipeline. Shift data into the pipeline and shift results out, and just watch a massively parallel task complete as memory is read on one side and filled up on the other. You could feed the results into another pipeline, configured for another processing task. Make a loop in the memory. Feed some data back forward, make a hardware video encoder that places the logic right in the memory area containing the video. Maybe you wanna play a game. Your memory could be reconfigured into a massively parallel shader pipeline. Why send data over PCIe lanes to a GPU. Just configure the GPU logic right in the unified system memory. That's the kind of long term potential that we may see with these memristors.
No drive fetches.
No sending or reading data out over busses to external storage or devices.
No fetching from RAM into the CPU registers.
No sending the results back to RAM.
You just create the logic block you want, right IN the memory space, and start shifting data through it. It's no longer necessary to even rely on general purpose functions... You can create ASIC like hardware logic functions, as needed, thanks to the FPGA like potential of such a future system configuration. Look up crossbar latches for more info on this concept.
Hewlett Packard gets it... even if they don't get branding...
*cough* Keysight *cough* ... They first developed the memristor in 2007ish, over 10 years of prior work led to its development, and now, after another 8 years of directly developing the tech, they've decided to gamble most of their chips on the tech. As in, reportedly, 75% of their computer division's R&D is focused on this new project. I don't know if they plan to use the in situ logic elements yet (that's a tech that might still need time to develop), but they are already proposing an all new concept for a next generation server that uses memristors for unified storage and system memory. They simply refer to it as "The Machine". They claim that it will eliminate the waste of seeking data over busses, from drives, and they also claim up to 90% increases in efficiency, due to the fact that memristors only require current flow to change, not to hold state. It's one of the big benefits of non volatile active processing elements. You only need power to change states or to read them.
Basically, HP realized that 5% margins on rebadged intel hardware was not a worthwhile future for their server business. They decided that they would devote their resources into creating a technology that had the potential to completely displace the competition. Stop and consider, if you were a data center, and HP offered you a machine that used up to 90% less energy, and was orders of magnitude faster at accessing and delivering stored data to customers... Would you be interested? That's HP's gamble. If they succeed, they escape intel's hold on the market, and return to being the market leader, with an innovator new tech. If they fail... They dumped 75% of their R&D for however many years they develop this.
I personally wish them luck. Those claims can't be emulated. It takes new hardware to pull that off.