Logic-in-Memory architecture made real

The current trend for intensive computational architectures is to adopt massive parallelism, with several concurrent tasks performed simultaneously, as done for example in GPUs. This approach has many advantages, such as the reduced design time given by circuit replication and an increasing in computational speed without the need of higher frequency. It has however evidenced an important bottleneck in data exchange between memory and processor. We envisage a revolutionary path for the future relation between memory and logic in parallel processors, where a new type of architecture exploits the principle of caching to the limit. Our Logic-in-Memory (LIM) architecture mixes logic and memory in the same device, removing the bottleneck of other existing parallel solutions. The architecture we propose, here in its preliminary version, has an array organization and each element in the array is based on three blocks: a logic unit for processing, a smart memory block and a routing structure for inter block communication. In this article we show the benefits of this approach with an application example in the image processing field. We can achieve a 4X computational time reduction for an image processing algorithm (Summed Area Table) with respect to the best architecture present in the literature, even with a preliminary and not optimized version. Besides the adoption of massive parallelism to increase performance, new technologies to open the post-CMOS era are explored. Among them NanoMagnet Logic (NML) is particularly interesting for its ability to mix logic and memory in the same device. We present here the preliminary results of the NML implementation of the LIM architecture. We thus demonstrate that it is not only a good solution for a standard CMOS technology but can also exploit the potential of an emerging technology as NML