LEVERAGING CHIP VARIABILITY
Embodiments are described that leverage variability of a chip. Different areas of a chip vary in terms of reliability under a same operating condition. The variability may be captured by measuring errors over different areas of the chip. A physical factor that affects or controls the likelihood of an error on the chip can be varied. For example, the voltage supplied to a chip may be provided at different levels. At each level of the physical factor, the chip is tested for errors within the regions. Some indication of the error statistics for the regions is stored and then used to adjust power used by the chip, to adjust reliability behavior of the chip, to allow applications to control how the chip is used, to compute a signature uniquely identifying the chip, etc.
- 1. (canceled)
- 2. An apparatus comprising:
a memory controller configured to control the operation of a dynamic random access memory (DRAM) device comprised of memory regions configured to be controlled by the memory controller, the memory controller comprising logic configured to; provide an error-prediction profile comprising error-prediction metrics respectively corresponding to the memory regions, each error-prediction metric representing a degree of likelihood that a corresponding memory region will produce a bit error at a given time, the bit error comprising an erroneous toggle of a bit value that is being retained by the DRAM device in the corresponding memory region, wherein the error-prediction metrics are predictive of potential bit errors in the memory regions represented by the error-prediction metrics, respectively, and wherein the error-prediction metrics exist concurrently with respect to each other; identify, according to the error-prediction profile, from among the memory regions, a target memory region, the target memory region identified based on a corresponding error-prediction metric; and based on the identifying, cause a change to an error sensitive factor (ESF) associated with the DRAM device, the ESF comprising a memory refresh rate, wherein the change to the ESF comprises an increase of the memory refresh rate.
- View Dependent Claims (3, 4, 5, 6, 7, 8, 9, 10, 11, 12)
- 13. A method to control operation of a dynamic random access memory (DRAM) device comprised of memory regions, the method comprising:
accessing error-prediction values associated with the DRAM device, the error-prediction values configured to determine that a memory region is likely to fail based on error characteristics of the respective memory regions, each error-prediction value corresponding to a respective memory-region-specific likelihood of corruption of data being retained by the DRAM device; and monitoring the error-prediction values and identifying the memory region according to the error-prediction values; and according to the identified memory region, changing an operating parameter of the DRAM device, the operating parameter comprising a refresh rate of the DRAM device.
- View Dependent Claims (14, 15, 16, 17, 18, 19)
- 20. An apparatus comprising:
a dynamic random access memory (DRAM) storage device comprising memory regions configured to store data written to the DRAM storage device and to provide data read from the DRAM storage device; a memory controller configured to control operation of the memory regions; the apparatus configured to, when operating; provide a set of error-prediction values for the respective memory regions, each error-prediction value changing in correspondence with probability that a corresponding memory region will be associated with an erroneous bit flip within the memory regions; select a target memory region according to the error-prediction values; and based on the selected target memory region, cause a change to an operating parameter of the DRAM storage device.
- View Dependent Claims (23, 24, 25)
- 21. An apparatus according to claim 21, wherein the error-prediction values are derived by counting accesses to the memory regions.
This application is a continuation of allowed prior application Ser. No. 15/462730, filed Mar. 17, 2017 (attorney docket 329638-US-CNT2), which is a continuation of patented prior application Ser. No. 14/604,608, filed Jan. 23, 2015 (attorney docket 329638.03), which is a continuation of patented prior application Ser. No. 13/791,479, filed Mar. 8, 2013 (attorney docket 329638.02), which is a divisional of patented prior application Ser. No. 12/819,100, filed Jun. 18, 2010 (attorney docket 329638.01). The aforementioned applications are incorporated herein by reference in their originally filed forms.
Chips, and in particular memory chips, are manufactured and provided to purchasers with a guarantee that a chip will most likely operate without errors if operated within parameters specified by the manufacturer. That is to say, a chip may be designed to provide error-free data storage if the chip is within a certain age, temperature, voltage, clock frequency, refresh rate, or other similar error-sensitive operation factor. DRAM chips in particular are overdesigned so that every cell among the billions on a chip loses data only with a very low probability under normal operating conditions. However, almost all of the memory cells in a typical DRAM chip can hold their value even if the operating conditions are changed. If a DRAM chip is refreshed at a below-recommended rate, almost all of the cells in the chip will continue to accurately retain their assigned values. As a result, typical DRAM chips are consuming power in the form of refresh power that is unnecessary for their effective operation.
Not only are chips often over-provisioned with respect to current or actual conditions or ESF values, they may be over-provisioned for certain portions of a chip. Chip manufacturing is an intricate process with considerable intra-process variation. There are variations from chip to chip. Chip manufacturers set a threshold for rejecting chips that do not conform to a specified quality threshold (measured in terms of performance, errors-per-chip, etc.). However, such thresholds are usually set low to ensure adequate yields, especially for commodity chips like DRAM chips. As a result, there is considerable variation even among chips that pass quality control. Some chips may have areas or regions that are more error-tolerant than other regions and therefore may require less power.
Embodiments described herein relate to leveraging variations in chips for both power optimization and unique identification.
The following summary is included only to introduce some concepts discussed in the Detailed Description below. This summary is not comprehensive and is not intended to delineate the scope of the claimed subject matter, which is set forth by the claims presented at the end.
Embodiments are described that leverage variability of a chip. Different areas of a chip vary in terms of reliability under a same operating condition. The variability may be captured by measuring errors over different areas of the chip. A physical factor that affects or controls the likelihood of an error on the chip can be varied. For example, the voltage supplied to a chip may be provided at different levels. At each level of the physical factor, the chip is tested for errors within the regions. Some indications of the error statistics for the regions are stored and then used to adjust power used by the chip, to adjust reliability behavior of the chip, to allow applications to control how the chip is used, to compute a signature uniquely identifying the chip, etc.
Many of the attendant features will be explained below with reference to the following detailed description considered in connection with the accompanying drawings.
The present description will be better understood from the following detailed description read in light of the accompanying drawings, wherein like reference numerals are used to designate like parts in the accompanying description.
As noted, regions 122 of a chip may have varying reliability. Some regions 122 may have higher or lower probabilities of errors than other regions 122 under a same set of ESFs. Although an ESF may be measured in a way that suggests an ESF is the same over all regions 122, in reality, at any given time, an ESF'"'"'s measureable value at different regions 122 may actually vary. A chip may have hot spots, areas where voltage is slightly higher or lower, and so on. Nonetheless, for purposes herein, an ESF may be treated as a uniform property. For instance, if a chip is measured as having a given input voltage, the voltage will vary within that chip in a consistent way. When a chip is sensed as having a single apparent chip-wide temperature at different times, if other ESFs are substantially unchanged, any region will have nearly the same temperature at those times, even if the temperature of the region differs from the measured or sensed temperature. Thus, it may be seen that if a chip'"'"'s error behavior is to be modeled in terms of ESFs, a single same measure of an ESF will be sufficient to model the error behaviors of the regions of the chips.
Referring again to the example of
Given the ESF value, the process then involves measuring 146 error behaviors for various regions of chip 138 at the ESF value. The error measuring 146 can be performed in numerous ways. A data pattern such as all-1s, all-0s, random bits, interleaved strings of 0s and 1s, etc., is stored to and read from 148 the chip 138. When the data is read, the read values are compared against the original data. Each region being tested has an error counter. When an error is encountered (e.g., a word, bit, etc. does not match its original value), the error counter for the corresponding region is incremented 150. The storing, reading, and counting are repeated and the number of loops (reads and writes) is stored 152 along with the error information or error count of each tested region. Note that for each store and read 148 iteration and/or for each varied value of the ESF, the data pattern may be varied, for example, the data pattern may be randomly selected from a dictionary of patterns, or alternated between all-zeros and all-ones, etc. In addition to the write-read-compare technique, in cases where errors are more likely to be related to logic delays and clock rate, bit scan chains and test patterns can be used to generate and seek errors.
In one embodiment, the ESF is tested for different values of the current ESF; if there are other 154 values to test, the ESF value is varied 144 (e.g., voltage is lowered from 1.7 V to 1.6 V), and the measuring 146 and storing 152 are repeated for that value. In one embodiment, it may be sufficient to test for only one value of the ESF. For example, the behavior of a chip family or material or the general affect of the ESF may allow error behavior of a region to be extrapolated to other values of the ESF.
If 154 the ESF is to be tested for multiple values, the varying 144, measuring 146, and storing 152 are repeated as needed. When errors have been measured and stored 152 for the value(s) of an ESF, if 156 there is another ESF to test, then another ESF is selected 142, and the varying 144 and measuring 146 are repeated. Otherwise, the process is finished 158.
From the numerous store, read, and compare cycles, the error behavior of a region can be derived from the number of cycles and the number of cumulative detected errors. For example, at a given ESF value, one region might have 1 error in 1,000,000 cycles, and another region might have 5 errors in those same 1,000,000 cycles. Errors might also be characterized in terms of clock cycles, overall test time, errors per amount of data read/written, and so on.
It should be noted that the regions of a chip can take any form. A region may be determined by physical construction of the chip, by addressing features, etc. A region or area need not be a contiguous unit (either physically or in address space), but as used herein will be considered to include any collection of areas or parts of the chip, e.g., interleaved segments or words, etc. Furthermore, not all parts of a chip need to be measured or included in the chip'"'"'s error function. For example, regions that have similar error traits may be grouped as a single region. Or, regions that have a similar error rate function may be represented by one of such regions. Note that in some instantiations regions may have different sizes.
An error profile for a chip may have multiple dimensions. For example, temperature may be varied by 5 degree increments, and each increment may be measured for errors over a range of voltages. Thus, an error function may map a region, temperature, and voltage to an error rate. For ease of discussion, examples will have a single ESF. As used herein, an “error rate” includes not only a rate of errors, but also a probability of an error, a measure of reliability, and so forth.
When an error profile for a chip has been obtained, the error profile may be used for many purposes.
The adjusting 194 may also involve determining if there is a level of an adjustable ESF that can be set to a new value in view of a desired reliability or power setting. For instance, if an application or operating system running on a computer using the chip specifies a desired level of error probability or error tolerance (e.g., high/medium/low, or some specific number such as a probability of error), then the error profile can be consulted to identify regions of the chip (or corresponding pages of memory) that would operate with acceptable reliability at a different ESF value.
It will be appreciated that the ability to obtain error information for regions of a chip can be used in almost any way to reduce power consumption by the chip. In yet another embodiment, a chip may have a special portion integrated with error-correction coding (ECC) functionality in the chip. The ECC portion of the chip may be supplemented with regional error information, thus allowing the ECC to operate with greater reliability or efficiency. If the ECC functionality of the chip is detecting more or less errors, the chip may use this information to adjust an ESF to be able to run using more or less power.
In one embodiment, a current ESF value may be used. For instance, a temperature of the chip may indicate, according to the error profile of the chip, that all regions are currently able to operate at an acceptable level of reliability, perhaps as a result of a low temperature value. In response, an operating parameter of the chip, such as memory refresh rate, may be adjusted to lower its power consumption. This example shows that it may be helpful to vary one ESF (e.g., refresh rate) in response to an external setting of another (e.g., temperature) that cannot be controlled.
The error profile of a chip may also be used to allow an application to select memory based on reliability needed by the application. For example, an application may have some data for which reliability is important. The operating system may have a memory manager that obtains information about what portions of physical memory (in the form of one or more memory chips with error profiles) are more or less reliable at a given current level of an ESF (or even without an ESF value). The operating system may then use this region-specific reliability or error rate/probability information to assign memory. When the application requests low reliability memory, the memory may be drawn from regions of one or more chips that are determined to be less reliable (perhaps at a given current ESF measurement or value). When the application requests high reliability memory, regions are selected and assigned accordingly.
In another embodiment, a global approach may be taken. Power consumption may be reduced across an entire computer or subsystems thereof. For example, a clock rate might be lowered. The clock rate is passed to the error rate functions of the computer'"'"'s memory chips to determine which regions of one or more chips are below a desired level of reliability. Those regions are then flagged by the operating system or memory controller as unavailable. Thus total available memory size is traded for reduction of power.
In another embodiment, a computer or software thereon may make dynamic adjustments to a setting that affects power consumption by analyzing the amount of memory that is being used. When there are substantial portions of memory not in use, the system may look for regions that have an error rate below some threshold at a given value of the setting. Or, the system may pick an incrementally lower or higher value of the setting, use the error profile of a chip to determine error rates for regions at that level, and select the regions that would operate with sufficient reliability. In this way, a system can lower its power consumption while assuring that only memory that has a desired error rate or lower is used. If memory consumption increases, more memory may be made available by increasing the setting (thereby increasing power use) and finding and returning to service regions that, according to their error profile and new setting, will be sufficiently reliable.
An error profile of a chip may be used for other purposes besides providing reliability granularity and power adjustments. Due to variations in manufacture processes and materials, the error profile for each chip at each setting of any one or more ESF values differs. Thus, the error profile can be used to uniquely identify a chip therefore may be used as a unique hardware or computer identifier which can be used to detect counterfeit chips, detect software piracy, verify identity for online transactions, and others. As described above, an error profile can be computed by measuring error rates at one or more ESFs such as input voltage, refresh rate, and/or temperature (which will be referred to as x). In one embodiment a mean error rate for the entire chip is measured directly or computed from the error rates of a known profile. The mean error rate for the chip at x will be referred to as E(x). The error rate of a region Ri will be: e(x, i), which is the error or reliability function for the chip. A signature bit for a region Ri will be computed as:
f(x, i)=sign(e(x, i)−E(x)).
In other words, a bit for a region is 1 if it is above the average and 0 if it is below the average. The overall chip fingerprint is a concatenation or set of the bits of the regions, e.g., f(x, 1), f(x, 2), . . . f(x, K) across K regions of the chip. Of course because each region'"'"'s error rate may be thought of as a random variable, any variety of signature functions may be derived for the error profile. Moreover, it is possible that the signature can be computed for an assumed or baseline x (representing the value(s) of the ESF(s) of the error profile)), rather than a measured x. Because the error rates for a region and a chip tend to be monotonic, for some chips the error rate for a region will usually be above the average (or below the average) for the range of x. In one embodiment, a region may be selected for inclusion in the signature only if its error rate differs from the mean by more than some threshold. Finally, a comparison between a pre-computed signature and a current computed signature can be performed using a Hamming distance or the like, thus allowing a match to be determined even if some small number of bits differ.
Note that error profiles of a chip may change with time. If the change is significant, an error profile can be recomputed. The error profile may also be designed to model age such that as the chip ages its error profile adapts to the age of the chip.
It should be noted that embodiments described above may be performed by special-purpose on-chip hardware, by software or controller hardware that is not part of the chip, or a combination of both. In a hardware embodiment, a chip may be provided with a special region that is able to compute a profile as discussed above. In another hardware embodiment, a dedicated part of a chip may be designed to run at different parameters and dynamically measure the error rate thereof dynamically. For example, a small region may be able to change its own input voltage or refresh rate and then check for errors in the region. This error rate may then serve as an error profile for the chip.
In another embodiment, an error profile may be based on hierarchies of regions.