This is very different materially from the TiO2 based ones used in HP-derived designs, though I'm not sure what differences that will lead to in practice.
TLDR: There is a major mismatch between the device physics of the academic group and what is being marketed/sold.
If the actual device they were selling represented what they did in this technical work, I would be very, very excited. The I/V curve looks great.
BUT, compare that figure to the I/V curve represented in the longer datasheet (linked to in parent comment). A one million fold increase in resistance in 5nS- That isn't analog at all. It could be fairly emulated as almost a flip-flop with just a few transistors!! Although a binary memristor at scale might be more interesting as a memory (eg ReRAM), at this scale meant for neuromorphic analog applications, it becomes utterly uninteresting.
So to be honest I am baffled how they went from such a promising inspiration to a completely bland packaged device. Two possibiliies come to mind: 'they' are entirely different groups, or a major mis-marketing is occurring in the interest of gaining some funds for future R&D.
There is no Machine. There never was any Machine. The Machine is an urban myth, a grand illusion, something to give your life meaning, but which is in fact not there.
Someone probably made a nice sum from the temporary stock bump...
I have a friend in HP who actually had a job on The Machine until the "hiatus." She got switched to another project when it got put on hold... but HP was actually assigning human resources to it at one point, not just making press releases.
My read on The Machine was that HP was developing the system in parallel with the memory component.
The idea was "imagine if you had a server with an insanely huge amount of non-volatile SRAM (that could hold OS, applications, and files), and no hard disk"
It was the "huge amount of non-volatile SRAM" part that the memristor was supposed to deliver and didn't.
I thought they just pushed that part of the project back. Didn't they announce that "the Machine" is still coming, just that first iterations will have a more traditional architecture?
Hello Everyone. My name is Terry Gafron. I am the CEO of the company in Boise selling memristors. You all make some very good points. We are a few honest guys, trying to bring this technology to life. We do sell research grade memristors. Leon Chua has bought some from us and we are working with him to refine their characterization. While many of us are from Micron, there was no IP package from them. We developed this with a local university and are working very hard on this. We are a bare bones start-up, bootstrapping our way through this. You are correct in that there are still some discrepancies we need to resolve, but we are working on this. We have been doing contract work for DoD for the past few years, developing memristor based processors for them. We have some significant IC design experience with this sort of thing, but there is still so much that needs to be done. We sell bare bones devices we scraped up the cash to build. We do not make colossal claims. Our aim is simply to drive a stake in the sand. If you have questions, you can always send me a note at terry.gafron@bioinspired.net. Thanks all!
Can somebody explain what does this mean for the field of AI (short and longer term) besides (I guess) the obvious advantage of being able to implement software algorithms in hardware?
Its not just for the AI field, if you read their About Us page you will see that a very interesting application has to do with UAV and algorithms for motion detection. The chip set design space for this is very low energy but efficient/accurate at a certain trained task, which is what makes bio-inspired applications so compelling. But going back to AI, I would say this:
Short term: a wide variety of demonstrators of various 'easy' problems, eg image and simple logic function pattern matching, in chip as opposed to in theory or software like you say. This can be done with dozens or a few hundred memristive devices.
Medium term: more interesting (my maybe not trying NP hard problems yet) application in image, sound processing and other non-linear signal processing that begin to rival best approaches coming from dozens of GPU cores. Such applications might require tens of thousands or even hundreds of thousands of devices. Pattern matching and classifications at this stage might rival that of a very simple brain or a small component of the neo-cortex.
Longer term: Unknown, but some possibilities when you use tens of millions of such devices may include full (neo)-cortex implementation in hardware, remember that an entirely new generation of non-linear algorithms that implement due to topological and temporal effects at the nano-scale may allow for better results than what we get given an equivalent number of transistors (yes, transistors are also nano-devices, but despite a small community working on sub-threshold switching, they are being used in an almost uniformly binary approach).
It reminds me a lot of a box for a GPU or fancy motherboard. The logo even looks exactly like he kind of one a second-tier video card manufacturer would have.
A friend emailed them for small quantity pricing a while ago.
"Thank you for contacting us. Our 16 pin package is the newest offering in our Neuro-Bit line of memristor products. Single packages are $240.00 USD each. We are offering a 5% academic discount on all of your purchases. The 16 pin dip will be available for shipping late-May and we are taking pre-orders now."
Eh, depends on where you're coming from. These clearly aren't "production grade." Maybe not even "academically curious grade." Until you can at least get budgetary pricing without sending someone an email, it probably isn't even worth investigating unless you're just intensely curious.
Never mind that $240 for 8 fragile, difficult-to-work-with memristors is definitely not cheap.
While I agree that they're not production grade, $240 is well within the budget of, say, a EE research lab that wants to experiment with memristor-based circuit topologies and has the equipment+expertise to handle them properly.
Points well taken, but if they're the only people selling anything like these, they can set whatever price they like. Higher prices also restrict their support efforts to serious people who know how things work.
Since it's Boise, I'll put money on these guys being either former HP or Micron employees that probably got their memristor IP as part of their golden parachutes now that "The Machine" is gone.
Data sheet: http://nebula.wsimg.com/35f756676bf3290bebff9c9ac0f9e3ea?Acc...
User manual: http://nebula.wsimg.com/6dba75009009af7a59036365876b3f66?Acc...
Demo video: https://www.youtube.com/watch?v=0wQviPeaNpw