You need to run with a better crowd. ;) I've never seen that in my life--UBC, Microtel, Stanford, and IBM Research were all pretty collegial. (Some of the IBM product divisions were a bit less so--there were folks there who seemed to think their job was to throw rocks at everybody else's ideas, but they weren't scientists.)
Thirty years ago, I did take a guy apart at a microcontamination conference because he was lying through his teeth and trying to sell snake oil for $$$. But only that once, and not just because he was wrong.
I'd be gentler about it today, but it really needed doing.
Cheers
Phil Hobbs
--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC / Hobbs ElectroOptics
Optics, Electro-optics, Photonics, Analog Electronics
Briarcliff Manor NY 10510
http://electrooptical.net
http://hobbs-eo.com
I've seen some shocking rudeness among physicists. One poor undergrad was making an oral presentation, to get into grad school, when the dean said "I've heard enough" and walked out. The guy was crushed. And I was at a meeting, about a project at CERN, when my physicist friend made a small mistake and was publicly savaged by a rival. Both events impressed me. I've seen a few milder cases too.
You're an honorary circuit designer, so you'd never do anything really nasty.
--
John Larkin Highland Technology, Inc
Science teaches us to doubt.
Claude Bernard
When I was taking Ron Bracewell's graduate Fourier Transforms class (back around 1984ish), the subject of the politics of tenure came up. He said,
"If you're in a place where tenure committees knife people, you're much better off getting knifed."
Cheers
Phil Hobbs
--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC / Hobbs ElectroOptics
Optics, Electro-optics, Photonics, Analog Electronics
Briarcliff Manor NY 10510
http://electrooptical.net
http://hobbs-eo.com
Right. My elderbrat moved from Austin to California to get a tenured position in biology, and lost to some ugly politics. So she started her own DNA consultancy, which works pretty well. It's nice to have her and the grandkids and whats-his-name nearby.
I agree. Patent applications are stunningly daunting.
... as it happens... recently been notified that my first ever 2 patent applications have been granted...
The fortunate bit is that I did practically nothing in the whole application process... the company has a dedicated guy to do that...
...yes... unfortunately... until recently I have not been clever another to find an existing idea to copy and rework it for a new purpose until now...
Its actually one of real value. Typically TCXOs can get temperature compensated to around 100 ppb to 300 ppb due to multiple perturbations in the crystal. This technique allows one to get a 10 times better post compensation. The first prototype asic actually worked!
The best ideas are usually not patentable...for example... my latest asics have an oscillator topology resulting in at least 10 dB better close in phase noise, yet the topology is too conventional to be patent worthy, even though it don't seem to have been used for this purpose before.
Just keep it a secret. If it's patented, it's hard to discover infringements and sue them. And in the US, you have to keep paying the patent office or it expires.
I've known guys with patent addictions, with something like 40 silly patents to their names.
I discovered accidentally that I'm the co-inventor on one patent. I must have signed something that allowed the outfit to do that.
So.... Barrie Gilbert wrote papers in the 70s of how to linearize a differential gain controlled amp...
The standard problem to be solved is having a very low noise front end gain controlled amp that is linear but with a wide input range. A full gilbert cell (input diodes) itself adds a fair bit of noise . A typical application is to do variable tuned, narrowband RF channel filtering digitally, e.g. mobile phones, which is pretty much impossible to do in analog. The input range being uVs to 100s mv. If the input distorts, the digital filtering wont work.
A diff pair can only take about 50mV max whatever the gain control tail current does.
Gilberts's idea is to use several diff stages in parallel but with different offset voltages. Thus forming an offset sum of tanh() functions. As the gain of one drops off, another takes over. I have a SS example... somewhere.
Now... xtals have up and down frequency shifts (100 ppbs) with temperature, all different from unit to unit, sitting on top of a main 3rd order chebychev 20 ppm shape. We use up to 4th order chebys for the main compensation, but to compensate these wiggles is not practicable with more chebys. The inflection points are all over the place.
Thus I realised that I could pinch the idea to use the offset diff pairs to form summed tanh() correction curves to take out the xtal wobbles.
The 1st application actually got rejected, not surprisingly, the idea of using S curves to compensate processes had already been patented...However very surprisingly, our man reworked it to make it more specific and bobs your uncle....
For me, what is significant is that.....
Usually, top brass pass down the chain decide what they want to produce, and the surfs churn the handle for them.
In this case, I was pondering the problem of poor yields due to the variability of the xtals and the knowledge of Gilberts technique popped up in my noggin as a solution. I explained the idea to my local work mates, we did some simulations with real xtal data and tanh() in Excel and convinced ourselves that it would work. We presented the idea to top brass and they said go ahead, make a test ASIC. Although a few teething issues with the ASIC, it worked well enough to prove the concept worked.
So, this is a method that should give at least a 5 times improvment in temperature stability, yet originated from a surf with no remit....
Its like writing a song...one takes bits from other peoples songs and make new one of your own.... ;-)
Yes.
Most patents, truly, achieve nothing.
The is a patent where the patent referenced my website on a start-up circuit!
That's a cute technique, thanks. My underwater projects are always the best too.
Cheers
Phil Hobbs
--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC / Hobbs ElectroOptics
Optics, Electro-optics, Photonics, Analog Electronics
Briarcliff Manor NY 10510
http://electrooptical.net
http://hobbs-eo.com
Given that you can approximate a function out of N piecewise segments of arbitrary length and slope, is there a process to locate the segment endpoints for minimum (peak or RMS maybe) error? I usually do it by eyeball but there must be a better way.
Overlapped TANHs is cute. One could slide them around (offset) and adjust their slopes (tail currents).
For one-offs, I generally use the Nelder-Mead downhill simplex optimizer, specifically routine AMOEBA from Numerical Recipes in C, 2nd Ed. The main bit of cleverness required is in making a vaguely-reasonable penalty function for it to work on.
Functions made of sums of A*tanh(b*x + c) for random A, b, and c, are going to have a lot of flat spots, so the penalty function may have them too. For that sort of thing, NR2 gives a simulated-annealing version of Nelder-Mead that they call AMEBSA (amoeba plus simulated annealing--they're from the FORTRAN-77 era when the first six characters of all function names had to be unique.)
Unconstrained optimizers such as Nelder-Mead can run off chasing unphysical solutions, so I generally use folding to impose an allowed solution space. In this problem you want the tail current of the diff pairs to be within some reasonable range, say Imin to Imax. On each iteration of the optimizer, you reflect the candidate point off the boundary, i.e.
Itail[i] = Imin + fabs( Itail[i] - Imin );
Itail[i] = Imax - fabs( Imax - Itail[i] );
This messes up fancy optimizers really badly, but Nelder-Mead doesn't even notice.
This is how the situation tends to end up, but there's often a stage where several theories look equally convincing, and the one that gets used depend s on the situation being looked at. In quantum mechanics, Dirac famously pr oved that the three different theories around at one point were all the sam e theory expressed with different mathematics.
happen in person several times.
That's more turf war behavior than anything specific to science.
Not the ones I've known. Getting an advanced degree does put you in contact with enough people with advanced degrees to let you realise that many of t hem aren't all that smart. Being smart helps, but being persistent and cons cientious is a lot more important.
Being smart in the way John Larkin is doesn't get you an advanced degree.
It's usually fairly easy to work out when you have got working examples tha t embody the different solutions.
That's not the problem. Scientists spend a lot of time reading the literatu re on the problem they are trying to solve, and that exposes them to a lot of old-fashioned circuit design. They don't spend a lot of time reading abo ut the latest integrated circuit, or cute ways of using it to solve a probl em that isn't obviously related to one they are trying to solve. They publi sh what they can get to work - and graduate students can spend a lot of tim e tweaking, and aren't in a position spend a lot on the latest and most exp ensive integrated circuits (particularly of their supervisor has never hear d of it) .
You will find a couple of my published comments to that effect if you do it thoroughly.
Clearly the patent lawyers involved weren't ingenious enough.
But there's the fact - well-known to US patent lawyers - that "everything i s obvious to the Supreme Court".
By the time something has been written up clearly enough to make it fit to patent, anybody remotely skilled in the art can understand that is a good i dea, and correspondingly obvious.
The GaAs single crystal pulling machine I worked on in 1986 was covered by a patent on the idea that if you waited long enough the extra weight accumu lated by a growing single crystal would apply enough extra force on the sus pension system to exceed any short term variations in the surface tension p ulling on the circumference of the solid crystal being pulled out of the ba th of liquid GaAs.
This struck me as being blindingly obvious at the time, though the 1975 pat ent was worded in a way that obscured what was actually going on.
There is an ostensibly better way of doing it, and some Scandinavian worked it out around 1986, and published it the journal being edited by the guy w ho had got the 1975 patent. The paper wasn't written with the idea of contr olling crystal growth in mind, so the application wasn't all that obvious. When I saw it, I figured that the approach was patentable, and sent a memo to my boss to that effect. Three months later I got called into his office to talk with the editor/inventor who was trying to sell us his patent based on the same paper, which he'd put in rather earlier. It was a fun conversa tion. As far as I know, the idea didn't work in practice, but it was exceed ingly cute.
Yes. There is. The "piecewise" bit was a key issue being solved though. One needs a smooth transition to get a low dV/dT slope. The slope error is typically more of a problem for the frequency stability that the actual stability itself
Exactly. The blocks have programmable slope, offsets and max current, with polarity.
One uses the solver in Excel!
One constructs a table with the data of the curve that is to be compensated. Then construct a sum of tanh() functions with co-efficients
A_n.tanh(B_n.T+C_n)
One tells Excel to find the best co-efficients that minimises the summed squared error of the original function and the approximating function.
Its really simple. One just selects the co-efficients to be varied and Excel just churns out the answer. No math knowledge is required at all.
It practice, one actually uses measured data tables from the blocks. It doesn't actually matter what the exact shape is so long as it is S like.
Yeah... it has been considered... several times....stunningly difficult to to this digitally within the noise and current constraints.
The target is of the order of -175 dBc flatband phase noise, -85 dBc close in . Any on board processing generates switching noise all over the place just for starters. 10 nV/rthz on a supply is a problem. Step changes in frequency are disastrous, so the DAC output can't wiggle about much. 1uv at
1Hz on the vcxo line is a problem thus with a 1V full scale the A/D and DAC needs 20 bits at least, realistically 24 bits. Analog has "infinite" resolution...
MEMS oscillators do do it all digitally, so its not entirely hopeless. They use two oscillators next to each other. One with a deliberate large temperature coefficient. The difference in frequency gets you temperature. No A/D needed, just a counter.
Memory lifetime is a also a problem... telecoms stuff specs 10 to 20 year lifetimes. We use fuse blown rom. Typically only need a few hundred bits, not bytes.
ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here.
All logos and trade names are the property of their respective owners.