The invention concerns the field of CAM (Content-Addressable Memory) memory cell, and advantageously a TFET 7T-TCAM (Ternary CAM).
Conventional 16T TCAM cells are expensive in terms of area and power consumption. The emergence of the IoT market pushes the development of more compact and energy efficient memory solutions.
The increasing demand for longer battery life time and, implicitly, for energy efficient devices, drives the research in the field of low power design in particular for the Internet of Things (IoT) era. Content addressable memory is a special type of memory which is accessed by data in place of physical location. This is an attractive feature which makes CAMs popular in high speed hardware based search implementations, such as look-up tables, data compression, image processing, network routers, etc. Speed of search in CAM comes at the cost of high power consumption. Efforts were made to optimize power consumption at both system and circuit level using various circuit techniques, like dynamic voltage frequency scaling, power gating, etc. and/or sacrificing bit cell area to provide less leakage while maintaining sufficient performance.
Binary CAM is the simplest type of CAM which uses data search words consisting entirely of 1s and 0s. Ternary CAM (TCAM) allows a third matching state of “X” or “don't care” for one or more bits in the stored data word, thus adding flexibility to the search.
Tunnel Field-Effect Transistors (TFETs) are better in terms of leakage because of different than CMOS in working principle. TFETs operate by band-to-band tunneling and, therefore, the theoretical subthreshold slope (S) is not limited to 60 mV/dec as in the CMOS case. It should be noted that fabricated TFETs with S as low as 30 mV/dec have already been measured. While a large variety of TFETs is present in the literature, a world performance record was established when integrated with FDSOI technology. Progress on TFET devices has encouraged research on TFET circuits and the few published reports on TFET circuits mostly focus on the design of TFET SRAM cells. Most of the reports on TFET SRAMs revealed difficulties in obtaining sufficient stability in read and write operations. As the stability in both operation modes is inherently low due to the electrical performance of TFETs and the low supply voltage, it is difficult for circuit designers to find the best balance between read and write. Due to lower than MOSFET ON current of TFETs and lower supply voltages, the performance of the TFET SRAMs is limited and cannot match the CMOS SRAM performance. In the document N. Gupta et al., «Ultra-Low Leakage Sub-32 nm TFET/CMOS Hybrid 32kb Pseudo Dual-Port Scratchpad with GHz Speed for Embedded Applications», in ISCAS, IEEE, 2015, 8T TFET Dual-Port SRAM is presented which overcomes the half-selection and write-disturb problems and provide lower than CMOS static power consumption (<5 fA/bit) even at 1V supply at the cost of increased bitcell area; performance is improved by using high supply voltages and dual ports.
In the document US 2011/0299326 A1, a 4T-TFET SRAM bit cell is proposed using negative differential resistance (NDR) property of TFETs in reverse bias. However, the architecture proposed in this document suffers from stability and performance issues. In order to maintain data during read operation, read current should be less than the hump current (in pA range) provided by NDR. Such constrain leads to an extremely slow read with the risk of data corruption while executing the operation. Moreover, the TFET transmission gate for data access limits the maximum operating voltage.
An aim of the invention is to propose an ultra-low standby power and compact CAM cell suitable for low voltage applications, and which requires less number, e.g. more than 50% reduction, of transistors in comparison to conventional CAM bit cells.
The invention proposes a CAM memory cell comprising at least:
This cell takes advantage of the TFETs negative differential resistance (NDR) property to form the latch. This latch comprises N first TFETs and is able to store a data the value of which corresponding to one of N possible stable values which can be stored in the (N−1) storage nodes.
When the cell corresponds to a TCAM cell, the cell stores data in the ternary latch and avoids the use of a second latch to store mask bit.
An advantageous embodiment of the invention corresponds to TFET/CMOS hybrid TCAM cell.
The memory may be designed using Si-TFETs and MOSFETs which can be fabricated together in FDSOI CMOS process.
The CAM memory cell may be such that:
In an advantageous embodiment, the write block may comprise (N−1) first data lines such that each one of the (N−1) first data line is connected to one of the (N−1) MOSFETs.
In this case, the gate of each of the (N−1) MOSFETs may be connected to a write word line in which a write control signal is able to be applied. This write control signal may control the triggering of a writing of data into the latch.
In a particular embodiment, the CAM memory cell may be such that:
This particular embodiment may correspond to that of a TCAM.
The output block may comprise one or several second TFETs each having its gate connected to one of the (N−1) storage nodes, one of its source or drain connected to a match line and the other one of its source or drain connected to a second data line in which the search data is able to be applied, and wherein an electric potential of the match line is able to discharge when the search data does not correspond to the data stored in the (N−1) storage nodes. In this case, the output block is able to fulfill a search/lookup table function, which is able to compare a search data with the data stored in the latch and to output a value corresponding to the matching or the mismatching between the searched data and the stored data.
According to a particular embodiment, the CAM memory cell may be such that:
This particular embodiment may correspond to that of a TCAM.
In an advantageous embodiment, the first and second data lines may be the same data lines. However, in a variant embodiment, it is possible that the first data lines don't correspond to the second data lines.
In another configuration, the output block may comprise one or several second TFETs each having its gate connected to one of the (N−1) storage nodes, one of its source or drain connected to a read bit line and the other of its source or drain connected to a read word line in which a read triggering signal is intended to be applied, and wherein an electric potential of the read bit line is able to discharge or not according to the value stored in said one of the (N−1) storage nodes. In this configuration, the output block is able to read the value of the data stored in the latch and to output this stored value.
In a particular embodiment, N=3 and the CAM memory cell forms a TCAM.
The invention also concerns a CAM (or TCAM) memory array comprising several CAM memory cells as disclosed above, wherein the CAM memory cells are arranged according to an array of several rows and several columns.
This invention will be understood easier in view of the examples of embodiments provided purely for indicative and non-limiting purposes, in reference to the appended drawings wherein:
Identical, similar or equivalent parts of the different figures described below have the same numeric references for the sake of clarity between figures.
The different parts shown in the figures are not necessarily drawn to scale, so as to make the figures more comprehensible.
The different possibilities (alternatives and embodiments) must not be understood to mutually exclude each other and can, thus, be combined with each other.
TFETs are reverse-biased p-i-n gated junctions that operate by tunneling effect, in which the electrostatic potential of the intrinsic region is controlled by a gate terminal. The TFETs used in the CAM memory bit cells described below may be calibrated and designed on data similar to that disclosed in the document C. ANGHEL et al., “30-nm Tunnel FET with improved performance and reduce ambipolar current”, IEEE Transactions on Electron Devices, 2011.
For example:
For the region III, the reverse biased output characteristic is named “unidirectional” due to the fact that the gate loses the control over the TFET for high negative drain voltages.
The band-to-band tunneling current dominates the hump. The current severely reduces and attains its minimum in the flat current region, whereas the tunneling current is suppressed due to the non-overlapping bands. The turn-on of the p-i-n diode is dominated by the thermionic emission over the barrier.
The cell 100 comprises three first TFETs 102.1, 102.2 and 102.3 serially connected one to the other and forming a latch 103. TFETs 102.1 and 102.2 are of the p type, and the TFET 102.3 is of the n type. The source of the p-TFET 102.1 is connected to the drain of the p-TFET 102.2, forming a first storage node 104.1. The electric potential in the first storage node 104.1 is named Q0. The sources of the p-TFET 102.2 is connected to the source of the n-TFET 102.3, forming a second storage node 104.2. The electric potential in the second storage node 104.2 is named Q1.
The first TFETs 102.1-102.3 are biased by a first supply electric potential VDD applied one the drain of the p-TFET 102.1 and a second supply electric potential forming a reference potential (ground) GND applied on the drain of the n-TFET 102.3. Thus, the first TFETs 102.1-102.3 are biased such that they are always in reverse bias VDS range, i.e. VDS≤0 for n-TFETs and VDS≥0 for p-TFETs. In addition, GND is applied on the gates of the p-TFETs 102.1 and 102.2, and VDD is applied on the gate of the n-TFET 102.3, such that first TFETs 102.1-102.3 are in forward bias VGS range, i.e. VGS≤0 for p-TFETs and VGS≥0 for n-TFETs.
Due to NDR property (see
The latch 103 formed by TFETs 102.1-102.3 can store ternary data using above mentioned three combinations of data. For example, ‘0’ is represented with {00} on {Q0Q1}, ‘1’ is represented by {11}, and ‘X’ or ‘don't care’ by {10}. However, it is possible to consider that the three valid storage combinations are associated differently to the values ‘0’, ‘1’ and ‘X’.
As shown in
The cell 100 also comprises an output block 106. In this first embodiment, the output block 106 is able to fulfill a search data function, that is outputting a signal which is representative of the matching or the mismatching between a search data and the data {Q0Q1} stored in the latch 103.
In this first embodiment, the output block 106 comprises a n-TFET 108.1 and a p-TFET 108.2, named second TFETs 108.1-108.2. The gate of the n-TFET 108.1 is connected to the second storage node 104.2, and the gate of the p-TFET 108.2 is connected to the first storage node 104.1. The drain of the n-TFET 108.1 and the source of the p-TFET 108.2 are connected to a match line ML. The source of the n-TFET 108.1 is connected to a data line DL providing the search data to the output block 106. The drain of the p-TFET 108.2 is connected to a complementary data line DLB providing the complemented value of the search data to the output block 106.
With this output block 106, the match line ML discharges only if either {Q0Q1} is {11} and search data is ‘0’ (i.e. DL=0 and DLB=VDD) or {Q0Q1}={00} and search data is ‘1’ (i.e. DL=VDD, DLB=0), i.e. mismatch condition. In case of match of data (matching between the searched data and the stored data), invalid search data ({01}) or don't care combinations {10} on {Q0Q1}, ML will remain precharged.
Read waveform is shown in
The cell 100 also comprises a write block 110 which is able to provide a data to be stored in the latch 103.
In this first embodiment, the write block 110 comprises a first n-MOSFET 112.1 and a second n-MOSFET 112.2. One of the source or drain of the first n-MOSFET 112.1 (the drain in
Since the drain current ID of the first and second n-MOSFETs 112.1 and 112.2 is much higher than the current through the drain current through the TFETs 102.1-102.3 (i.e. hump current), it results in fast writing of data into storage nodes 104.1 and 104.2.
Write operation for different combination of values is shown in
Wordline boosting is used to overcome the limitation of using n-MOS for writing ‘1’ on storage nodes 104.1 and 104.2 ({Q0Q1}={11}).
In the first embodiment shown in
In the first embodiment previously disclosed, the latch 103 of the cell 100 comprises three TFETs 102.1-102.3 forming two storage nodes 104.1, 104.2 able to store three possible combinations of values ({00}, {10} and {11}). More generally, the latch 103 of the cell comprises N TFETs 102.1-102.N forming N−1 storage nodes 104.1-104.(N−1) able to store N possible combinations of data, with N≥2. When N=2, the cell 100 corresponds to a binary CAM. When N=3, the cell 100 corresponds to a ternary CAM.
All first TFETs 102.1-102.N of the latch 103 are in reverse bias VDS and forward bias VGS. Thus VGS<0 and VDS>0 for a p-TFET 102, and Thus VGS>0 and VDS<0 for a n-TFET 102.
Valid stable values on each storage node correspond to either VDD or GND.
Concerning the possible combinations of data, a stable combination of data is one fulfilling the following condition: V(104.1)≤V(104.2)≤ . . . ≤V(104.N−1).
Advantageously, the latch 103 can comprise eight TFETs 102.1-102.8 such that eight stable combinations of data can be stored in the seven storage nodes 104.1-104.7 formed by these TFETs 102.1-102.8. These eight stable combinations may correspond to the eight possible values of the 3-bits word. The table below represents these eight stable combinations of data:
In the first embodiment previously described, the output block 106 comprises elements able to fulfill a search data function, which may considered as corresponding to a lookup table function. As a first variant embodiment, it is possible that the output block 106 comprises elements forming a read port, which is able to output the value(s) stored in the latch 103, whatever this value. For example, the output block 106 can comprises one or several TFETs 114, each having its gate connected to one of the storage node(s) 104, its drain connected to a read bit line 116 (RBL) and its source connected to a read word line 118 (RWL).
According to a second variant embodiment, the search data function of the output block 106 may be fulfilled by TFETs such that, for one storage node 104, a n-TFET 120.1 and a p-TFET 120.2 have their gates connected to this storage node 104.
The behavior of this output block 106 is close to that of the output block 106 previously disclosed on
As for the output block 106, the write block 110 may comprise elements which are different than those previously disclosed in relation with the first embodiment of the
An example CAM memory array 1000 is shown in
Generally, proposed cell 100 is ultra-low standby power and works for wide voltage range from 0.4V to 1V. Evaluated standby power is <2 fW/bit, which shows 3 decades improvement in comparison to state of the art low standby TCAMs. Less than 1 ns read and write time is achieved at 0.6V supply with match line load for 256 bit words.
More generally, the proposed cell 100 may be made with different TFET technology and fabrications. The values of the features of the TFETs may also be different.
Number | Date | Country | Kind |
---|---|---|---|
1652054 | Mar 2016 | FR | national |
1659649 | Oct 2016 | FR | national |
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/EP2017/055812 | 3/13/2017 | WO | 00 |