letters to nature ..............................................................
Programmable computing with a single magnetoresistive element A. Ney, C. Pampuch, R. Koch & K. H. Ploog Paul-Drude-Institut fu¨r Festko¨rperelektronik, Hausvogteiplatz 5-7, D-10117 Berlin, Germany .............................................................................................................................................................................
The development of transistor-based integrated circuits for modern computing is a story of great success. However, the proved concept for enhancing computational power by continuous miniaturization is approaching its fundamental limits. Alternative approaches consider logic elements that are reconfigurable at run-time to overcome the rigid architecture of the present hardware systems1. Implementation of parallel algorithms on such ‘chameleon’ processors has the potential to yield a dramatic increase of computational speed, competitive with that of supercomputers2. Owing to their functional flexibility, ‘chameleon’ processors can be readily optimized with respect to any computer application. In conventional microprocessors, information must be transferred to a memory to prevent it from getting lost, because electrically processed information is volatile. Therefore the computational performance can be improved if the logic gate is additionally capable of storing the output. Here we describe a simple hardware concept for a programmable logic element that is based on a single magnetic random access memory (MRAM3,4) cell. It combines the inherent advantage of a non-volatile output with flexible functionality which can be selected at run-time to operate as an AND, OR, NAND or NOR gate. MRAMs consist of giant-magnetoresistive5,6 (GMR) or tunnelling magnetoresistive7 (TMR) elements which are composed of two magnetic layers (1 and 2 in Fig. 1). These are separated by a nonmagnetic spacer or a tunnelling barrier that suppresses magnetic coupling, and guarantees that the magnetizations M 1 and M 2 of the two layers can be rotated independently. The resistance of a magnetoresistive (MR) element depends on the relative orientation of M 1 and M 2 and is considerably lower for parallel alignment. The two magnetic states—parallel and antiparallel, characterized by low and high resistance—can be identified with a logical 1 and 0, respectively. For writing a bit, a current with a magnetic field that is sufficient to rotate the magnetization of one of the two layers is required. Since the direction of the magnetization is maintained when the current is turned off, the information is non-volatile and can be read out repeatedly by measuring the resistance of the MR element without the periodic refreshing needed in CMOS (complementary metal oxide semiconductor) technology. Furthermore, MR elements offer enhanced logic abilities compared to the rigid architecture of transistor-based logic elements where the functionality is fixed by the wiring. In two recent studies it was proposed to form a programmable spin-logic element by altering the switching threshold of an MR element8, or by combination of several MR elements8,9. However, we will show that a single MR element with two independent input lines (A and B in Fig. 1) is already sufficient to provide three different functionalities: the conventional storage bit as well as AND and OR gates selectable simply by the addressing (‘set’) procedure. By adding a third independent input line C the logic gates NAND and NOR can also be realized. The MR element required for the proposed logic gate (Fig. 1) consists of two magnetic layers with different coercive fields H c1 and H c2 where H c1 , H c2. The input lines A and B are operated with positive or negative currents ^I A and ^I B of equal magnitude. Each of the two currents alone is not able to generate a magnetic field sufficient to reverse M 1, because the coercive fields H c1 and H c2 are NATURE | VOL 425 | 2 OCTOBER 2003 | www.nature.com/nature
too large. However, I A and I B together are able to rotate M 1 but not M 2. For rotation of M 2 an additional current ^I C on input line C is necessary. We now discuss the realization of the AND and OR functionalities by particularly using the non-volatility of MR elements (see also Fig. 2). AND gate. See blue arrows in Fig. 2a. We assume that the magnetization M 2 of the lower layer points to the right; according to the chosen convention it is positive (þM 2). Antiparallel (parallel) alignment of the upper layer, that is 2M 1 (þM 1), corresponds to a logical 0 (1) as output of the MR element. Magnetization with negative (positive) sign is accomplished by negative (positive) currents at the input lines (^I A, ^I B) identifiable with logical 0 (1) of the inputs. Before the logic operation the system is set to the antiparallel configuration, that is 2M 1/þM 2, by applying 0 at both inputs A and B. This configuration corresponds to an output 0. In the next step, the logic operation takes place: inputs A and B are addressed independently with a 0 or a 1. The direction of the magnetization remains unchanged if 0 is applied at both inputs A and B. The same holds if 1 is applied either at input A or at input B. The magnetization of the upper layer can only be switched to þM 1 by applying a logical 1 at both inputs A and B. Only then does the output change to 1 (parallel configuration). The corresponding logic table given in Fig. 2a (blue entries) can easily be recognized as the binary logic AND function. OR gate. See red arrows in Fig. 2a. To realize the logic functionality of an OR gate the MR element has to be set to a different, but again well-defined, initial state before the logic operation: the magnetization of the upper layer is set to þM 1 by applying a 1 simultaneously to inputs A and B; the magnetization of the lower layer remains unchanged at þM 2. The parallel configuration þM 1/þM 2 with the output 1 is maintained even when the input currents are turned off—a direct consequence of the non-volatility of MR elements. The subsequent logic operation proceeds as follows: applying a 1 to both inputs A and B repeats the set sequence and the output remains 1. With only one input being 1 and the other 0, the resulting field is not sufficient to rotate M 1, thus keeping the output at 1. Only after applying negative currents (that is 0) to both inputs A and B, þM 1 changes to 2M 1. Then the two magnetizations are aligned antiparallel, which corresponds to an output 0. The resulting logic table is also given in Fig. 2a (red entries) and represents the logic OR function. The preceding discussion illustrates the inherent properties of MR elements. If they are employed as spin-logic gates, the operation has to proceed in two steps, namely the ‘set operation’ followed by the ‘logic operation’. The first step (where both inputs are addressed together) allows us to choose between the AND and OR functionality by setting the magnetization of the upper layer to 2M 1 or þM 1, respectively. This means that for the different functionalities AND
Figure 1 Schematics of a programmable spin-logic device based upon a single MR element with three independent input lines A, B and C and an output line.
© 2003 Nature Publishing Group
485
letters to nature and OR the output has to be preset to 0 and 1, respectively, before the logic operation. Then in a second step both inputs are addressed separately and the preselected logic operation is performed. Although at first sight a two-step procedure appears to be a drawback in speed compared to CMOS technology, our concept offers several advantages: (1) since the output is non-volatile, it can be repeatedly read out via the GMR or TMR effect (analogous to an MRAM). Consequently, a time-consuming transfer of the output to a cache memory is not required. (2) With a non-volatile logic gate the different stages of a computation need not necessarily be synchronized to a single clock. Therefore the gates can be operated asynchronously, which is expected to yield low power circuits10 and support parallel computation. (3) The set procedure makes runtime programmability feasible, which is expected to increase the computation efficiency by up to three orders of magnitude owing to massive parallel and pipelined execution2. The proposed logic element therefore represents the simplest solution of a programmable spin-logic gate based on a single conventional MRAM cell. By using only the ‘set’ and ‘read’ steps, the device acts as an ordinary MRAM bit.
Figure 2 Principle of operation of the MR-element employed as logic gate. a, Set operation by switching the magnetization of the upper layer (M 1) to preselect between the logic functions AND (blue arrow) and OR (red arrow) by simultaneously activating inputs A and B with currents þI or 2I. The subsequent logic operation for AND (blue) and OR (red) is executed by the input currents ^I of lines A and B leading to different magnetization alignments with resulting high or low resistance (R) identified with the output. The logic input and output values are summarized in the logic table. b, Negation of the output (orange) is achieved by reversing the magnetization of the lower layer (M 2) by simultaneously addressing all three input lines. The subsequent logic operation proceeds analogously to a. 486
NAND and NOR gates. See Fig. 2b. So far we have assumed that the lower layer 2 has the magnetization þM 2. The logic functionality (AND, OR) of the MR element is solely defined by the relative orientation between the magnetization of the upper layer M 1 and the input currents I A and I B. Reversal of M 2 only affects the value of the output changing the magnetoresistance from low to high, or vice versa. Reversal of M 2 therefore corresponds to a negation and turns AND to NAND and OR to NOR (see Fig. 2b). For rotation of M 2 currents of equal sign have to be applied to all three input lines to overcome the coercive field H c2. We illustrate the complete sequence of a logic AND operation followed by a NOR operation in Fig. 3: in a first ‘set’ step 2M 2 or þM 2 is selected by activating all three input lines—(i) in Fig. 3— leading to a negation of the output or not. In a second set step (ii), layer 1 is programmed by simultaneously addressing the input lines A and B, yielding either the (N)AND or (N)OR functionality; input line C is not addressed. The device is now pre-programmed. After the set-procedure the ‘operation’ step takes place by independently using lines A and B as logic inputs. The last step is to read the output by measuring the magnetoresistance between a single input line, for example, C, and the output line (see Fig. 1). The resulting signal can be amplified if required and distributed to other MR elements, for example, by repeated read-out. The new concept for a universal, non-volatile spin-logic gate, presented here, is based on a single MR element with three input lines and one output line. Whereas an MR element possesses four distinct magnetization states, the output distinguishes only between two states, namely antiparallel and parallel, irrespective of the individual magnetization directions. Each of the four states, however, represents a different logic functionality, making the MR element suitable to serve either as AND, OR, NAND, or NOR gate or as a conventional MRAM cell. The logic functionality has to be predefined before the logic operation is executed. The MR element has several advantages compared to common logic devices: (1) a single MR element is sufficient to realize and store the four basic logic functionalities. Since at least two transistors are needed for each CMOS logic gate, the integration density is effectively increased. (2) The output is non-volatile and repeatedly readable without refreshing, which reduces the heat evolution. (3) After the logic operation the device is again in a defined state for a logic operation which can be directly (that is, without a further set-
Figure 3 Sequence of set, logic, and read operations for AND and NOR. Two initial set steps are performed: (i) to enable/disable negation of the output by addressing all three inputs, and (ii) to choose between (N)AND and (N)OR using only two inputs. The third step is the logical operation analogous to Fig. 2. The last step is the read-out of the output via the GMR/TMR effect. The two different coercive fields of the layers H c1 and H c2 (dotted lines) and the resulting magnetic field H of the given input currents þI and 2I are indicated in the upper panel.
© 2003 Nature Publishing Group
NATURE | VOL 425 | 2 OCTOBER 2003 | www.nature.com/nature
letters to nature step) used by new types of algorithms. (4) Aside from the input and output lines no further means are necessary for the logic operation, for example, additional magnetic fields, switching thresholds, or voltages. (5) The switching frequency of magnetic films can be pushed to several GHz, in principle allowing fast operation11. (6) Since only short current pulses are required to rotate the magnetization, the energy consumption is further reduced. (7) The size of an MRAM cell—commercially available in 2004 (ref. 12)—can easily be scaled down to less than 100 nm (refs 13, 14) to further increase the integration density. (8) Our approach a priori does not necessarily require semiconductors to fabricate a logic gate. (9) The run-time programmability will increase the computational efficiency. Our new approach has the potential to induce a paradigm shift from transistor-based logic to magneto-logic, where the programmable functionality and non-volatility are as important as miniaturization and clock speed. A Received 12 February; accepted 1 September 2003; doi:10.1038/nature02014. 1. Prinz, G. A. Magnetoelectronics. Science 282, 1660–1663 (1998). 2. Sidhu, R. P. S., Mei, A. & Prasanna, V. K. in Field-programmable Logic and Applications (eds Lysaght, P., Irvine, J. & Hartenstein, R. W.) 301–312 (Lecture Notes in Computer Science 1673, Springer, Berlin, 1999. 3. Parkin, S. S. P. et al. Exchange-biased magnetic tunnel junctions and application to nonvolatile magnetic random access memory. J. Appl. Phys. 85, 5828–5833 (1999). 4. Gru¨nberg, P. Layered magnetic structures: history, highlights, applications. Phys. Today 54, 31–37 (2001). 5. Baibich, B. N. et al. Giant magnetoresistance of (001)Fe/(001)Cr magnetic superlattices. Phys. Rev. Lett. 61, 2472–2475 (1988). 6. Binasch, G., Gru¨nberg, P., Saurenbach, F. & Zinn, W. Enhanced magnetoresistance in layered magnetic structures with antiferromagnetic interlayer exchange. Phys. Rev. B 39, 4828–4830 (1989). 7. Moodera, J. S., Kinder, L. R., Wong, T. M. & Meservey, R. Large magnetoresistance at room temperature in ferromagnetic thin film tunnel junctions. Phys. Rev. Lett. 74, 3273–3276 (1995). 8. Black, W. C. Jr & Das, B. Programmable logic using giant-magnetoresistance and spin-dependent tunneling devices. J. Appl. Phys. 87, 6674–6679 (2000). 9. Richter, R. et al. Field programmable spin-logic based on magnetic tunneling elements. J. Magn. Magn. Mater. 240, 127–129 (2002). 10. Martin, A. J. Proceedings of SSGRR 2000, International Conference on Advances in Infrastructure for Electronic Business, Science, and Education on the Internet; at khttp://www.ssgrr.it/en/ssgrr2000/papers/ 185.pdfl (2000). 11. Gerrits, Th., van den Berg, H. A. M., Hohlfeld, J., Ba¨r, L. & Rasing, Th. Ultrafast precessional magnetization reversal by picosecond magnetic field pulse shaping. Nature 418, 509–511 (2002). 12. Motorola Inc. Press release. IEEE International Solid State Circuits Conference (San Francisco, June 2001); at khttp://www.motorola.com/mot/document/content/0,1028,372,00.docl (2000). 13. Goronkin, H., von Allmen, P., Tsui, R. K. & Zhu, T. X. Nanostructure Science and Technology (eds Siegel, R. W., Hu, E. & Roco, M. C.) 67–92 (National Science and Technology Council (NSTC) Committee on Technology and The Interagency Working Group on NanoScience, Engineering and Technology (IWGN), 1999); at khttp://www.wtec.org/loyola/nano/05_01.html (1999). Copyright is held by: WTEC, Loyola College (Maryland). 14. Compan˜o´, R. Technological Roadmap for European Nanoelectronics at kftp://ftp.cordis.lu/pub/ist/docs/ fetnidrm.zipl (2000).
Competing interests statement The authors declare that they have no competing financial interests. Correspondence and requests for materials should be addressed to R.K. (
[email protected]).
..............................................................
Controlling anisotropic nanoparticle growth through plasmon excitation Rongchao Jin, Y. Charles Cao, Encai Hao, Gabriella S. Me´traux, George C. Schatz & Chad A. Mirkin Department of Chemistry and Institute for Nanotechnology, Northwestern University, Evanston, Illinois 60208, USA .............................................................................................................................................................................
Inorganic nanoparticles exhibit size-dependent properties that are of interest for applications ranging from biosensing1–5 and catalysis6 to optics7 and data storage8. They are readily available in a wide variety of discrete compositions and sizes9–14. Shapeselective synthesis strategies now also yield shapes other than NATURE | VOL 425 | 2 OCTOBER 2003 | www.nature.com/nature
nanospheres, such as anisotropic metal nanostructures with interesting optical properties15–23. Here we demonstrate that the previously described photoinduced method23 for converting silver nanospheres into triangular silver nanocrystals—so-called nanoprisms—can be extended to synthesize relatively monodisperse nanoprisms with desired edge lengths in the 30– 120 nm range. The particle growth process is controlled using dual-beam illumination of the nanoparticles, and appears to be driven by surface plasmon excitations. We find that, depending on the illumination wavelengths chosen, the plasmon excitations lead either to fusion of nanoprisms in an edge-selective manner or to the growth of the nanoprisms until they reach their lightcontrolled final size. The photoinduced synthesis of silver (Ag) nanoprisms involves the preparation of a colloidal suspension of Ag spheres (diameter ,10 nm), followed by conversion of the spheres to larger prism structures with visible light. In a typical experiment, colloidal Ag nanoparticles passivated with sodium citrate and bis(p-sulphonatophenyl)phenylphosphine dihydrate dipotassium (BSPP) (diameter 4.8 ^ 1.1 nm, standing solution) were irradiated with a narrowband light source (using a 150 W xenon lamp with a light output ,12 W) with an optical bandpass filter (centre wavelength 550 nm, width 40 nm) for ,50 h (see Supplementary Information). Transmission electron microscopy (TEM) shows that the colloid formed consists of two different size distributions of nanoprisms (Fig. 1a and inset), with the smaller particles (designated as type 1) and the larger particles (type 2) having average edge lengths of 70 ^ 12 nm and 150 ^ 16 nm, respectively. These structures tend to form stacks, so that edge-on views allow the precise determination of nanoprism thickness23. Although the average edge lengths for the type 1 and type 2 nanoprisms are significantly different, their thicknesses are almost identical (9.8 ^ 1.0 nm) (Fig. 1b, c). The bimodal particle growth process also has been monitored by ultraviolet–visible–near-infrared (UV–vis.–NIR) spectroscopy (Fig. 2a). During the reaction, one sees the disappearance of the plasmon band at ,395 nm (characteristic of the spherical silver particles) and the formation of two new, strong plasmon bands at 680 nm and 1,065 nm that are associated with the type 1 and type 2 nanoprisms, respectively (see below). The band for the type 1 prisms is initially centred at l max ¼ 680 nm and gradually blueshifts to l max ¼ 640 nm. This blue-shifting correlates with the tip sharpness of the nanoprism features; rounding is known to lead to blue-shifting24. The second strong band at l max ¼ 1,065 nm is assigned to type 2 particles (see below). In addition to the two strong surface plasmon bands, one can observe two other weak resonances at 340 and 470 nm, respectively (Fig. 2a, spectrum 6). To gain further insight into the optical spectra of the solution with the bimodal particle distribution, we carried out theoretical modelling using a finite-element-based method known as the discrete dipole approximation (DDA)24–26. The calculated spectrum shows plasmon bands that reproduce the experimentally observed spectrum (compare Fig. 2b and Fig. 2a, spectrum 6), confirming our peak assignments. The first three peaks in the spectrum of the colloid containing both type 1 and type 2 particles, centred at 340 nm (outof-plane quadrupole resonance), 470 nm (in-plane quadrupole resonance) and 640 nm (in-plane dipole resonance)24, result from the type 1 nanoprisms; in the case of the type 2 nanoprisms, only the strong dipole resonance at 1,065 nm is clearly observed. Quadrupole resonances, which occur at 340 nm and 600 nm (weak) in the spectrum of the solution of the type 2 nanoprisms, are overlapped with plasmon bands from the type 1 nanoprisms (Fig. 2b). The time-dependent optical spectra thus suggest that the process is bimodal, rather than unimodal as would be expected in the case of conventional Ostwald ripening13,14. The bimodal growth of Ag nanoprisms is not caused by the wavelength dispersity of the excitation beam (550 ^ 20 nm). Indeed, when a monochromatic laser beam (l ¼ 532.8 nm, the
© 2003 Nature Publishing Group
487
