#E50#E50
Q: What is “Run & Sample”?
IDDQ Measuring IDD and IDDQ
|
In the DCPMU window, we find a feature called “Run & Sample” when either “Measure DUT Supply” or “Measure Iddq” is selected. It appears in the lower portion of the window, like this: 
This Q’nApp will illustrate this function by way of example.
IDD Measurements of NAND FLASH In our quest to measure IDD during the programming phase of a NAND type FLASH (Toshiba part TC58128AFT) we discovered a degree of volatility we seldom see with any CMOS or bipolar device. The measurement mode was Sample While Running. Employing the original test files, we observed readings where the relative IDD measurements could vary by more than a magnitude. When changing the vector file so that the program sequencing of vectors provided tight looping on the programming function of the FLASH, the variations were significantly reduced. Nevertheless, we still observed measurements indicative of something less than good methodology. Attempts to use the PC controller for the
purpose of discerning the unexpected results were unsuccessful.
In retrospect, we concluded this was so
because of accumulated timing uncertainties ranging from the PC’s
caching modes to response time aberrations inherent in
interfaces between computer and external devices. To combat this
deficiency, the HRISC – a 64-bit Harvard type RISC
processor residing in the periphery of the test head – was
utilized. Among the salient features of the HRISC is
the quality that every basic instruction takes exactly the same time,
with
accuracy only dependent on the frequency accuracy of a crystal
oscillator. Furthermore, the HRISC is tightly coupled
with the DC PMU measurement resources*. Results of measurements
along with their
interpretations are the principal subject of the remaining text.
*While featuring hardware multiply and a
“barrel shifter”, the HRISC was designed to act as the controlling
device of
high speed digital I/O devices.
Test Setup The tester employed was a HILEVEL Dragon with 128 pins and with USB connecting to the PC. HiLevel Technology Polska designed a special DUT board for the sake of establishing a MultiSite backdrop that in the lab would closely emulate a production environment. This investigation applied MultiSite mode with two sites, though the combined results are not always shown.
(Note: Subsequently,
tests were also conducted
using single site mode; the results are provided in an addendum.) The aforementioned DUT board has
sixteen
sockets, each of which can house a 128Mb or 1Gb NAND FLASH.
Connections between tester and DUT board are
through FLEX circuits that combine the utility of high quality
transmission
cabling and a LoadBoard. The FLEX
circuits connect directly to the pin-electronic cards, thus bypassing
the
standard LoadBoard. While a FLEX
circuit looks like a cable, it is indeed a printed circuit board that
facilitates a 50Ω environment; in a sense, the FLEX circuits act as a
LoadBoard. All in all, the purpose of
this setup is to create a MultiSite environment that meets or exceeds
the
electrical characteristics of “Direct Docking”; we call it “HILEVEL
Firm
Docking”.
Power and ground connections also
extend directly
from the pin electronic cards. The
ground is a solid layer, while a 20 mil trace is used for connecting
power. Power sense lines are merged
with power lines at the far end of the cable.
Each DUT has two decoupling capacitors: One 0.1µF and one 0.01µF. There is no “hand-wiring” on the DUT board, implying a 50Ω setting throughout, excepting at the socket and two connectors; both connectors, the one connecting to the pin electronic card and the other connecting to the DUT board, provide one ground pin for every signal. Indeed, the type of connectors utilized causes negligible discontinuity of the transmission line environment. |
Methodology
& Measured Data Importing the acquired data into
Microsoft Excel,
Figure 1 depicts a striking periodicity of power consumption of the
FLASH
device when subjected to repeated Erase operations. While varying
between approximately 4mA to
slightly above 10mA, the surges in current into 330µs spikes
occur roughly in 1
millisecond intervals. Referring to Figure
1, a “time tick” corresponds to a program loop of the HRISC, and, for
the sake
of completeness, the corresponding HRISC microprogram is shown in
Figure
2. Each “tick”, the total time between
samples, is 712 HRISC cycles. With each
cycle being 60 ns, a “tick-time” is consequently 42.720
µs. The majority of that time (41µs) is just
waiting for the ADC to become ready*.
* A mode of the ADC is available whereby the HRISC simply waits until the ADC has completed the conversion; in the final implementation this mode will be used. While it reduces test time, it may also have the somewhat disconcerting effect of test times varying for same type devices. The HRISC operation is invoked when the PC
issues an HRISC command. The HRISC
responds by acquiring 1000 samples. The
results are first stored in the HRISC’s local memory and later
transferred, as
a burst via the USB connection, to the PC controller. The PC in
turn computes min/max values as well as an average of
all 1000 samples. The “raw” data
accumulated for a total of ten operations (each of which with 1000
samples) are
furnished in Appendix A (because of its length we suggest it be left
unprinted). Again let us emphasize that our investigation
employed MultiSite mode, using two sites.
In this mode, the sampling is done simultaneously for all sites (one
ADC
per site) and the data is actually gathered simultaneously for the two
ADCs in
our investigation. The result is
depicted in Figure 3. The striking
part of this picture is the skew between spikes for the two ADC
readings. Seemingly, the erase operation for one FLASH
is done more frequently than for the other FLASH! At least, one
device lags the other, and indeed in a cumulative
manner. Measurements of IDD when programming
the FLASH were also made. The picture
that emerged as seen in Figure 4 is very different from that of Figure
1. Data appears almost random, though
“modulated in some fashion”. However,
the maximum reading of the current is very similar in the two operating
modes. The differences between the two
pictures are almost certainly the reflection of the different ways of
writing
data; yet and again, the max values remain basically the same for the
two
modes. On the other hand, the min value
was measured to be virtually zero for the specified range during erase
while clearly non-zero during programming. Are these max values real? Would they be higher if unfiltered by decoupling capacitors? Or are they spikes that show an unrealistic representation due to the presence of wire inductance? Above all,
which quantities should be used as a
representation of max IDD value? And how many times do we need to
sample? In answering these so imperative questions,
we turn to the data in Appendix A. The
peak values are repeated periodically and thus presumed to be a
reasonable
representation of worst case programming current for a specific
device.
Besides, this value also agrees with typical
values reported in the device specification (by Toshiba). What
constitutes a sufficient number of
samples is a topic for statisticians to ponder; nevertheless, a
thousand samples will take less than 50 milliseconds and seemingly
gives
adequate results. In 10 independent
cases of sampling 1000 times, the difference in maximum reading for the
10
cases were no more than 4 ADC counts (50µA).
In light of this we tentatively conclude that 1000 samples will be both
economical and sufficiently accurate, though it may become subjected to
digital
filtering in the final implementation.
See also: Qe50.zip is a zipped Word file of this Q'nApp. Click your browser's Back button to return to the Q'nApps index. |
|
Notes:
ADC = 2900: IDD = 10.427 mA ADC = 2400: IDD = 4.272 mA 1 Time Tick = 42.720 µs. Figure 1: ADC reading during repeated ERASE operations |
|
DO_SAMPLE_POWER_WHILE_RUNNING: Figure
2 – HRISC command to execute 1000 samples from
ADC
|
 Figure 3 – Simultaneous ADCs readings during ERASE operations |
 Figure 4 –
ADC reading during repeated Program
operations
|
Conclusions and Further Questions To reach the final conclusions we solicit the opinions of those who by their experience can shed further light on the topic. It would for instance be nice to know the type of measurement methodology used by Toshiba. It would also be nice to know what kind of resources other testers have to make the above measurements; in particular, is the statistical approach suggested here necessary. And is it sufficient? On the specifics, we would like to know if
there is
any a’ priori “digital way” of determining when the IDD will be
the
highest – in which case we would only need to make the measurement at
that
moment. Suspecting that the device
“independently and capriciously chooses its own agenda”, we reluctantly
and
hesitantly conclude that sampling is the “only way to go”. Upon feedback to all the aforementioned assertions (perhaps of dubious quality) we will embark on the “new and alternate” implementation of the “Sampling while Running” mode of IDD test. |
Addendum
|
Surprisingly, using the
original single site test we found the results to be almost identical
to those
of MultiSite. However, when comparing
results obtained using the same SET file as for MultiSite, we
found the
measured values to be approximately 30% higher.
Then by adding decoupling capacitors to the DUT board, the
difference was narrowed to less than 10%.
Details are provided in the subsequent text.
While we did investigate erase
mode as well as program
mode, we focus on the former since both modes exhibit virtually the
same
maximum readings and the former (erase
mode) provides a simpler basis for analysis.
The tester was an ETS770
and a general purpose DUT board was used with signals hand-wired to the
socket. Power and ground were provided
using a flat copper tape in series with round wires (< 3cm). A total of four decoupling capacitors were
initially employed (2 of 0.1µF, 2 of 0.01µF), each connected
between power and ground
of the strips of copper tape*. Also, a fixed
range (20mA) was applied for
taking DCPMU measurements. At first we ran the test
that was initially designed for power measurements (it didn’t have a
tight
programming loop). When running the
test a thousand times (using Repeat Test in AutoTest), we got the
following
results: 10.200mA
(988 cases) 10.010mA
(7 cases) 9.810mA
(5 cases) While stable and
remarkably similar result to that of MultiSite, the measurements
nevertheless
exhibited an “interesting” anomaly discussed below.
In light of later discoveries, the “remarkable
consistency”
between single and MultiSite just seems like “lucky coincidence”**. |
Another crucial experiment
makes us suspect that we can only hope for adequate results, and not
necessarily accurate results. By adding
two 0.68µF
capacitors, the above 10.200mA reading was reduced to 8.330mA. As one could expect, the peak was dampened
by the extra capacitor. What would
the
optimum capacitor be? The HRISC microcode for Sample
while Running in single site mode is now tentatively completed. In the 20mA range, the time to measure takes
about 75ms, which includes starting the test, taking a thousand
samples, and
computing the proper result. Almost half
the time (33µs) goes to starting and completing the test;
4ms is what the HRISC uses to compute the max value while the rest (38µs)
is dedicated to the actual Sample
while Running. * A detail of potentially great importance in explaining the qualitative differences between measurements in MultiSite and single site. ** The above stability would indeed have been remarkable if the measurements were off by 1 and 2 counts. But as it turned out, the aberrations differed by 16 counts and by 32 counts (one count representing about 12µA in the 20mA range); curiously, they were off, relative to the first case, by 16 times 1 and 2 respectively. Perusing through the collected data, the appearance of a fundamental flaw was so strong that we were compelled to run a basic test with resistors to check out the hardware. For DC measurements pertaining to the resistor, the results were as predicted; one is almost left to ask: Is the behavior of the ADC reflecting “something like” it cannot handle rapid changes? We received the answer from our expert. The 574 ADC (a well-known “work-horse”) is not very fast and does not have the sample-and-hold feature that we sorely need in this dynamic application. Fortunately, another ADC, the 1674, does have sample and hold circuitry and features a 10µs max conversion time (versus 35µs for the 574). To what extent HILEVEL may make corresponding changes in its hardware is partially subjected to further testing using the 1674 ADC. |
 Figure 5 – ADC reading of erase in single site using original decoupling capsThe results show instability akin to transients of voltage readings of a bad high-speed pin driver. By adding two 0.68µF decoupling capacitors, the results changed significantly as shown in Figure 6. And with the addition of these capacitors, the measurements got much closer to those obtained with MultiSite – the difference was reduced to less than 10%. We suspect more decoupling caps will further “clean up” the spikes and might indeed render results that match those of MultiSite. Figure 6 – ADC reading of erase in single
site when adding decoupling
caps
Final conclusion is yet to come! |
