Přepis řeči - MULTI-RATE POLYPHASE DSP AND LMS CALIBRATION SCHEMES FOR OVERSAMPLED DATA CONVERSION SYSTEMS

0:00:13	and
0:00:13	um
0:00:14	so a graph from everyone and things but anyway style
0:00:17	to are representing my work on a turn rate only phase
0:00:21	just S and elements rician scheme
0:00:23	for sampled data the conversion systems
0:00:26	and this work has been done at a university of washington seattle um
0:00:30	with michael you ten can which a ring and permission might by of that bit out
0:00:36	as a motivation of my talk call
0:00:38	present a do we need this yes is good and a lot so that its what's a role of
0:00:43	um
0:00:43	dish the uh D S B about terms for than about circuits and there
0:00:47	and we'll go with the brief or real
0:00:49	but it easy to a you that we have selected which in this case would be a segment of thirty
0:00:53	C architecture
0:00:55	then will propose a low and the lms calibrated to the analog and
0:01:00	and
0:01:01	then after that will look into that is to back and use be regarded them
0:01:05	which uses is under rate and the missed estimation for
0:01:08	and finally will complete
0:01:11	so
0:01:12	and
0:01:13	the requirement for as we going to more and more seven and a meter process so you can process
0:01:19	we find that
0:01:20	a life of an analog design a is becoming harder and harder because
0:01:24	for even simpler blocks like operational amplifier
0:01:28	you find that the voltages is that shrinking
0:01:30	you are strings are reducing and so
0:01:33	the challenges is becoming higher and hard to reach a higher again
0:01:36	at the same time for the should design a is becoming a relatively easy you have a much higher sampling
0:01:41	speed
0:01:42	and it can be that
0:01:43	um you can really a not achieve i to but
0:01:48	so that
0:01:48	the basic motivation for this work is this to use a core design approach
0:01:53	um a to be use that dish to the S got tends to kind of correct then a lot of
0:01:57	functions that we can
0:01:59	so before or we step into they at C architecture it is a very general overview of
0:02:04	different type of D Cs
0:02:06	and this architecture is just a sorry
0:02:09	the still clutching is actually a a just compare is actually it
0:02:13	for a low to medium band with application
0:02:16	and for high resolution
0:02:17	applications on a so
0:02:19	it's a very generate go we just to give an idea of we stand
0:02:22	and if you look we have
0:02:23	different types of a arctic for a T Cs like flash
0:02:27	i line
0:02:28	successive approximation just tear and segment the at C
0:02:31	and they have been compared for different parameters like throughput resolution solution they can see power here
0:02:37	and
0:02:37	for
0:02:38	like for low to medium and at that location and high resolution
0:02:42	you find a sigma delta a C a one of the most of a lot to that
0:02:46	we can have a
0:02:47	and so you know use we are actually targeting like for wireless lan uh approximation of like nice i like
0:02:53	that
0:02:54	and eleven it affected them
0:02:58	before we step into the at C are look into that
0:03:01	if see architecture the some critical implementation choices is that we should we need to decide
0:03:06	and to of the definitions that we have a
0:03:08	as for or something ratio
0:03:10	which is defined
0:03:11	as as a ratio of sampling frequency or what twice a signal time
0:03:14	no what's simple of this number is is that
0:03:17	i do a so you find that you can get a high
0:03:20	uh
0:03:22	at the same time
0:03:23	for white that of that iterations
0:03:25	it means that we need a a has in frequency
0:03:29	because if a is increasing
0:03:30	it's got at first increases as well
0:03:33	so here in this design
0:03:35	the focus
0:03:35	more on maximizing minimizing the power
0:03:38	rather than maximizing this P
0:03:40	and so that's of reason real select a low sir of one eight for was design
0:03:46	and the question that we want to answer is
0:03:48	how little can we should be in of a pen
0:03:51	like a is it why able to use like
0:03:53	around ten a can be enough ten or twenty
0:03:57	and so will will like is twenty six db gain or of easy but generally like a sigma don't applications
0:04:02	the are in the range of like fifty to seventy db
0:04:06	which is a to achieve a as we use a scaled down the supply
0:04:10	so we'll see that if such a thing just yeah use a to be all them uh oh and that
0:04:14	such a again
0:04:15	if you try to use the in and bold is in the system can you can you can be them
0:04:19	which today
0:04:20	and and the whole
0:04:21	uh process we power
0:04:23	yeah fish
0:04:25	before we four still uh so this like of a simple um evil you
0:04:30	all a second order
0:04:31	see my that the U C to large
0:04:34	so that i made a C it can be just
0:04:36	plus five in terms so
0:04:38	it's a two input
0:04:40	one our system
0:04:41	so these are like to uh a time samples sampled input
0:04:45	and then you get a a a a two a a and it can be it in terms of runs
0:04:49	functions
0:04:50	well the signal process
0:04:51	S T F
0:04:52	and noise times
0:04:54	you
0:04:55	so will will to of this out
0:04:57	and a be five
0:04:58	so if you look into the right side here so
0:05:01	so a large
0:05:02	score the V four
0:05:03	oh
0:05:04	no that's what we is thing but is just it's of C
0:05:08	you take that out of the sum
0:05:09	and you give it to the to integrate is
0:05:11	do of one
0:05:12	two
0:05:13	let's signal that someone O G N S one year
0:05:16	time being
0:05:17	so this is there since that to integrate as
0:05:19	well
0:05:20	i not as the
0:05:21	we take the output of the someone and we some all the all
0:05:24	a and you give it to a for quantizer
0:05:27	for like to D C
0:05:28	and the we do a C is then
0:05:31	back
0:05:31	to at least a to compare compared
0:05:33	oh sorry
0:05:35	and by to
0:05:36	which is then some time
0:05:39	they have to the right is of the still a for all us
0:05:43	as you can see from the
0:05:45	from the signals signs here
0:05:46	on the Y axis is the a is a range of and you hear
0:05:50	and the X is just a times
0:05:53	so you see that the was we one and B two
0:05:56	have a very small signals planes
0:05:58	which means that we do i mean it to use as the requirements for the
0:06:02	for the integrate is that we have a
0:06:04	all
0:06:05	and it also
0:06:06	the using this string mean
0:06:08	that's so and using
0:06:10	a and ones on the entire assist
0:06:12	or
0:06:13	that that's one of the pretty good advantages of having a a for long
0:06:17	the second row
0:06:18	that if you can the process
0:06:20	from input of the quantizer
0:06:22	to a the the second integrate two
0:06:25	we find that
0:06:26	the last
0:06:27	is just a quantization noise
0:06:30	so we'll see what's a man to this is to
0:06:32	this feature in the next i
0:06:34	but is the two things that's well
0:06:37	and
0:06:38	so for this system
0:06:39	we can and a like the signal as a function which is the cost
0:06:42	from the input that
0:06:44	to but
0:06:45	skis
0:06:45	and the also
0:06:46	and to yeah i post
0:06:48	a given by a second one scene
0:06:51	if you like to do not at here for we find
0:06:54	while see that all put yeah training
0:06:56	but
0:06:57	position of two done
0:06:58	X of the month like of the signal process
0:07:00	okay
0:07:01	so what implies lies is as the input
0:07:04	goes to all without
0:07:05	a frame
0:07:07	and the one position on
0:07:08	is not filtered by a signal
0:07:10	i
0:07:11	um
0:07:12	so you can think of this as like it basically a portion that in but quantisation noise
0:07:16	with
0:07:17	to go that
0:07:18	and in that was is it is a resolution of eighties
0:07:22	the next question that the we have to so is is this
0:07:25	system
0:07:27	sufficient for for a requirement of realising a one bit
0:07:30	snr or S india
0:07:33	so think is like this system but then always are of eight can realise and nine but that's the not
0:07:37	so we need an additional two bits
0:07:39	so what should we do
0:07:41	so what we do here is we
0:07:43	we into it to a large you just on a stochastic signal at at at C
0:07:48	so it's very simple like take the system from the previous slide
0:07:51	we are another system a
0:07:55	and then to do that
0:07:56	what we do is at the input of this second stage
0:07:59	we just stick you would you all
0:08:01	Z you minus two from the previous it
0:08:03	and added to the input of the second stage
0:08:06	now
0:08:07	feed you it in a more or any uh a a little bit more mathematics a
0:08:12	and if you yeah to digital filters
0:08:14	S of two D which actually matches
0:08:17	uh a that a lot signal as a function of the second stage and then T of one D which
0:08:22	match is the noise
0:08:23	a function of the first stage
0:08:24	and we some though two
0:08:26	why of C
0:08:27	then after to doing some at which we go now
0:08:30	we find that the we can get a higher order of my thing which implies a high resolution
0:08:35	ones
0:08:36	uh let's just finish the system here
0:08:38	and so once we get this out
0:08:40	this is just a decimation estimation for to and it that's mixed down by eight
0:08:44	as you can see a three state to to the first as the cascade to get the call to just
0:08:48	but you for that about how when only face but to
0:08:51	the the estimates it's by two for by a single uh a five compensation for
0:08:55	but isn't it split into three state is is just to my as the power
0:08:58	for
0:09:00	let's look into the summation here and you see here is that the input signal
0:09:05	X of Z
0:09:06	C just a a segment as a function
0:09:08	and
0:09:08	as we saw the previous slide
0:09:10	the signal process
0:09:11	and is you need so the signal that signal X of C
0:09:14	goes out with that
0:09:15	you need
0:09:17	a second term here
0:09:19	which is a forced one quantization noise as you in this is a ideal system
0:09:23	would find that
0:09:24	if you assume that the and T F one of me is equal to T of one of the
0:09:28	and S you of two of these equal to have two of this at this and sits out
0:09:32	and so there's is not one position on the first to from position noise doesn't come
0:09:38	the total component reduced you to of C
0:09:41	sees a four door that a function here
0:09:43	and it effectively means that now we are able to
0:09:46	noise shape by a fourth order
0:09:49	so
0:09:49	that's it's it's a good thing in the sense that we have
0:09:53	we can just take to stable second out of trance that systems
0:09:56	put them together and get a higher order
0:09:59	a a higher or the noise shaping which implies a a high snr
0:10:02	what what
0:10:03	the disadvantage of the system is this one will be in a cyst uh it is the design
0:10:08	it will introduce mismatch between the and a lot
0:10:11	a fancy
0:10:12	chen and that is to
0:10:14	function
0:10:15	so how does that come to the house
0:10:17	so as we see you know like for a very high gain or and we see that
0:10:21	the noise or or for uh
0:10:24	a noise of the system is done to but i'm long
0:10:28	so here here it's six
0:10:29	this is the signal back from zero
0:10:32	to ten made
0:10:34	and on the bias
0:10:34	the or like to rent C V
0:10:36	D
0:10:37	so right now
0:10:38	we can see that for a very large gain it's done it but um and noise
0:10:42	but at that it use uses a pretty six db
0:10:44	the one position noise
0:10:46	starts don
0:10:47	on
0:10:49	on the right to easy you why that's a reason
0:10:51	so we see that
0:10:52	that approximation emission that we came a are cross across year one minus the was where
0:10:58	all of it true because that are outside
0:11:00	as an iir are filter
0:11:02	because of that that's that's thing
0:11:04	and the gain of whatever
0:11:05	"'cause" the system that's
0:11:07	to right
0:11:08	so what happens is that one position as of the first stage weeks without it
0:11:12	and it produces
0:11:13	by uh
0:11:14	has an impact on a phone
0:11:19	so
0:11:19	what we do is we take the help of and of the audit
0:11:23	to really
0:11:24	or a the performance impact that we have
0:11:27	so i'll go to the
0:11:28	the detail in the blocks of the plots in the next slide
0:11:31	what
0:11:32	in this like we just want to get a general overview of what we want to do
0:11:35	so what we do is we deactivate the input and we inject
0:11:38	at to to do great
0:11:40	no no no C
0:11:41	which is generated on to
0:11:43	from a linear feedback shift register
0:11:46	at the same time we take the sequence
0:11:48	to uh
0:11:49	to a set a file for this
0:11:51	uh i T have to the in in of one B
0:11:53	the output of this a five for is compared with that
0:11:56	a with that and a lot of by one of the Y to see
0:11:59	and then we use that a missing write them to get you can be the distance
0:12:03	a
0:12:04	so right
0:12:05	so to summarise rate yeah actually using
0:12:07	sign to the ms one
0:12:10	for
0:12:10	a listener was it's one of the most easiest to implement that are most of if
0:12:16	and it's it's from you got to maybe compute the next year or efficient
0:12:20	from the previous state
0:12:21	use that the patient size
0:12:22	you know as there
0:12:23	if getting in the design and
0:12:26	uh and that it right C
0:12:28	we take the sign of
0:12:29	yeah yeah
0:12:30	that
0:12:32	the way is about the works is it each be because of the operates
0:12:36	to it finds that error is minimised that a miss things
0:12:39	and the coefficients is actually get a response to the
0:12:43	by so as a function of the signal doesn't
0:12:45	that we expect on the
0:12:47	so after the long i do calibration
0:12:50	you see that the to uh signal thus
0:12:53	this same as that of one
0:12:54	and for what they stiff and then you
0:12:57	oh
0:12:59	there is much what do you know how we implement the
0:13:02	uh
0:13:03	the filters
0:13:04	so we inject the thirty two bit to a L that it and m-sequence a thing
0:13:09	yeah
0:13:11	screen and rate to five to
0:13:13	the output of this a five
0:13:15	is that can be a the output
0:13:18	that you see why
0:13:19	the Z Y two Z
0:13:20	and there is then you meant to have a isn't going
0:13:23	and the coefficients of an adaptive guys
0:13:26	so to to use the truncation errors
0:13:28	in there five for it is use an eight to five
0:13:32	one disadvantage that the
0:13:34	system them can have a
0:13:35	is that you in for a set of the asians
0:13:37	we are actually sound the system
0:13:39	i a sampling frequency of the N T one fifty two hundred meg
0:13:44	a what as is that
0:13:45	you do you have a very low or and
0:13:47	because now we are using very
0:13:50	simple simple fires and optimising and
0:13:52	the
0:13:53	oh
0:13:53	a channel addition
0:13:54	but
0:13:55	three can
0:13:56	approximate three times if you do a
0:13:58	a force that
0:13:59	as compared to that
0:14:01	so what should we do
0:14:02	how should be uh
0:14:04	uh or should be in signal processing techniques here
0:14:07	to make sure that you get a good efficient says
0:14:10	what we do a is in that to mention the police that you just
0:14:14	about
0:14:14	oh
0:14:15	we are using a noise cancellation the also
0:14:18	which is same as
0:14:19	basically C the N T F and S T E
0:14:22	uh yeah on again as and C S
0:14:25	and the are basically using that a
0:14:27	as we talked in the previous slide
0:14:29	no but ones get a is given decimation fig
0:14:32	which on
0:14:34	so the proposed approach what we do it
0:14:38	we use one these techniques
0:14:39	do really to do a convolution of
0:14:41	the noise
0:14:42	the noise cancellation fit
0:14:44	and india for this stiff
0:14:45	yeah
0:14:46	and that is for that
0:14:48	and get a what if easy
0:14:51	then we do the same per
0:14:53	we reject the a random sequence
0:14:55	we get the desired signal
0:14:57	and B J the all output of the four and a stage
0:15:01	and we make sure that sees the C
0:15:04	and then compared what the stages
0:15:06	do you do in N
0:15:07	and
0:15:09	now this approach is very similar to what a good subband adaptive filter
0:15:13	but
0:15:13	one pretty good point two
0:15:15	that we are using or sampling in this case
0:15:18	so using or so
0:15:20	i
0:15:20	it tells us to use the you using
0:15:22	right
0:15:22	and it helps service
0:15:23	to ensure
0:15:25	that of
0:15:25	that
0:15:26	the white one of C
0:15:28	oh which we used to basically
0:15:30	uh this you estimate
0:15:32	is not a very aggressive anti aliasing filter
0:15:34	and it just has to be very many
0:15:37	so
0:15:37	that this approach
0:15:39	for a or so as well
0:15:43	so that is the proposed a one
0:15:45	some
0:15:47	and
0:15:47	the the system that we show that us like reject
0:15:51	it and see
0:15:52	here
0:15:53	we are a a lot of the estimate it
0:15:55	and we lost
0:15:55	a if it's like filter
0:15:57	take out right
0:15:58	the same time we take the out of that of
0:16:01	well as the
0:16:02	this yeah
0:16:03	so for or or in this this meeting but
0:16:06	a of this
0:16:09	right to that it are and take their
0:16:11	compute coefficient
0:16:12	so we do that for both the parts of are here
0:16:15	and finally
0:16:16	we can speak
0:16:19	good thing about this is we found that
0:16:21	but the meant in a similar
0:16:23	it's you using this to two at which means that you can
0:16:26	re read to not to my
0:16:28	based on the whole system in innovation
0:16:33	um
0:16:34	do does have a small one right
0:16:36	because of the increase in the number of taps of you and C T E M Z but was that
0:16:40	have can is based on
0:16:42	and it can he's the optimized when we do not you
0:16:45	issue
0:16:46	we friend the power decreases when you decoding
0:16:49	case
0:16:50	oh
0:16:51	that's television we shouldn't a points the number of points in or
0:16:55	are pretty much the same for all all the different techniques
0:16:58	so
0:17:00	is proposing but the system that we have
0:17:02	yeah we have that a lot
0:17:03	yeah
0:17:04	we we see that
0:17:05	using the low you know that
0:17:07	introduces some some in what errors which is a mismatch in the
0:17:10	and and and to process function
0:17:12	and here as the but different flavours of the back and that we have a
0:17:15	and to again an F P G
0:17:17	and we got and mention that a misty with the one face but to a and what if is better
0:17:22	or
0:17:26	there is a
0:17:27	they should signal to noise ratio in of you of this nation
0:17:30	so without any yeah
0:17:31	a missing uh without any kind vision find a signal to noise ratio
0:17:35	yeah
0:17:36	as actually don't around fifty six points
0:17:38	db
0:17:39	after calibration
0:17:40	uh
0:17:41	we see that
0:17:42	all the three seems give hear a D or form
0:17:44	eleven bits
0:17:46	oh six is a T V C
0:17:47	i
0:17:48	i
0:17:48	C six
0:17:49	i
0:17:53	um
0:17:54	so here is a matrix fixed calibration
0:17:57	sampling frequency V using as
0:17:59	one
0:17:59	have
0:18:00	oh of the option that i
0:18:03	i
0:18:04	what
0:18:05	what
0:18:06	or supply voltage
0:18:07	one point two five or to
0:18:10	a spare
0:18:12	i meant one had in a seamless
0:18:16	as a comparison
0:18:17	the it's uh
0:18:18	a a here is the forced to be so we use like a in a six db
0:18:23	but it's just a lot of them
0:18:25	to yes
0:18:26	and before that up to my
0:18:28	the rest of the cost of i'd a reference for years
0:18:32	and
0:18:32	actually a uh and the quantizer
0:18:35	and i think this and the number can be really optimized to last
0:18:39	what's for a forty five
0:18:40	design
0:18:41	yeah
0:18:44	we uh as compared to the other the words we see that we have a
0:18:47	can't
0:18:48	uh a or savings for that
0:18:51	for the by
0:18:52	we do a
0:18:53	approximate analysis of finding the power for years
0:18:58	so the use that are
0:18:59	i see that
0:19:01	uh you
0:19:02	a power as it can be at fourteen time
0:19:04	uh a small errors
0:19:06	G
0:19:07	i become got this the actions of you for many words
0:19:12	what is
0:19:12	you get on that
0:19:15	a before for the demise
0:19:16	as this
0:19:18	so we do not that it at your had
0:19:20	because of some
0:19:21	for does that you have to put to match
0:19:23	it
0:19:26	so is a final completion format
0:19:28	um are don't was to the use of a six E uh you know
0:19:33	sees
0:19:34	you said that
0:19:35	addition white them to go go from the gain errors
0:19:37	and uh we proposed something
0:19:39	techniques
0:19:40	which uses like month at it uh are it important is
0:19:44	only polyphase
0:19:45	i
0:19:47	and we find that
0:19:48	using this you
0:19:49	that as in using at or
0:19:51	the machine
0:19:52	especially for high fit
0:19:54	by
0:19:57	thank you
0:19:58	if you have any question
0:20:02	sure
0:20:08	oh
0:20:13	yeah
0:20:21	also
0:20:22	yeah we compared to the convention
0:20:26	the
0:20:27	then use
0:20:28	i don't
0:20:30	i a moving bad to are of people
0:20:33	yeah the whole system can see
0:20:37	uh
0:20:38	okay
0:20:38	okay
0:20:42	i
0:20:46	just space
0:20:53	and
0:20:57	min
0:20:58	yeah i would say like a uh i think though the presence and this project has been on like a
0:21:02	design between that a lot in by
0:21:05	so if you compare the uh
0:21:07	i figure of matter it's with the current architecture use one just
0:21:10	the an up front and that we have
0:21:12	it's actually not uh it's not
0:21:14	as good as
0:21:15	scum it's and the ball but but it's not as good as the best ones and there
0:21:19	but the more mean as is here is how can you really meant like a back and
0:21:23	when you have a like a very cheap and a
0:21:32	i
0:21:33	oh
0:21:34	i
0:21:35	oh
0:21:35	thank you

MULTI-RATE POLYPHASE DSP AND LMS CALIBRATION SCHEMES FOR OVERSAMPLED DATA CONVERSION SYSTEMS

DSP Algorithm and Architecture Optimization for Hardware Implementation