Architectural Solutions for NanoMagnet Logic

The successful era of CMOS technology is coming to an end. The limit on minimum fabrication dimensions of transistors and the increasing leakage power hinder the technological scaling that has characterized the last decades. In several different ways, this problem has been addressed changing the architectures implemented in CMOS, adopting parallel processors and thus increasing the throughput at the same operating frequency. However, architectural alternatives cannot be the definitive answer to a continuous increase in performance dictated by Moore's law. This problem must be addressed from a technological point of view. Several alternative technologies that could substitute CMOS in next years are currently under study. Among them, magnetic technologies such as NanoMagnet Logic (NML) are interesting because they do not dissipate any leakage power. More- over, magnets have memory capability, so it is possible to merge logic and memory in the same device. However, magnetic circuits, and NML in this specific research, have also some important drawbacks that need to be addressed: first, the circuit clock frequency is limited to 100 MHz, to avoid errors in data propagation; second, there is a connection between circuit layout and timing, and in particular, longer wires will have longer latency. These drawbacks are intrinsic to the technology and for this reason they cannot be avoided. The only chance is to limit their impact from an architectural point of view. The first step followed in the research path of this thesis is indeed the choice and optimization of architectures able to deal with the problems of NML. Systolic Ar- rays are identified as an ideal solution for this technology, because they are regular structures with local interconnections that limit the long latency of wires; more- over they are composed of several Processing Elements that work in parallel, thus exploit parallelization to increase throughput (limiting the impact of the low clock frequency). Through the analysis of Systolic Arrays for NML, several possible im- provements have been identified and addressed: 1) it has been defined a rigorous way to increase throughput with interleaving, providing equations that allow to esti- mate the number of operations to be interleaved and the rules to provide inputs; 2) a latency insensitive circuit has been designed, that exploits a data communication protocol between processing elements to avoid data synchronization problems. This feature has been exploited to design a latency insensitive Systolic Array that is able to execute the Floyd-Steinberg dithering algorithm. All the improvements presented in this framework apply to Systolic Arrays implemented in any technology. So, they can also be exploited to increase performance of today's CMOS parallel circuits. This research path is presented in Chapter 3. While Systolic Arrays are an interesting solution for NML, their usage could be quite limited because they are normally application-specific. The second re- search path addresses this problem. A Reconfigurable Systolic Array is presented, that can be programmed to execute several algorithms. This architecture has been tested implementing many algorithms, including FIR and IIR filters, Discrete Cosine Transform and Matrix Multiplication. This research path is presented in Chapter 4. In common Von Neumann architectures, the logic part of the circuit and the memory one are separated. Today bus communication between logic and memory represents the bottleneck of the system. This problem is addressed presenting Logic- In-Memory (LIM), an architecture where memory elements are merged in logic ones. This research path aims at defining a real LIM architectures. This has been done in two steps. The first step is represented by an architecture composed of three layers: memory, routing and logic. In the second step instead the routing plane is no more present, and its fe