Search

In recent years, artificial intelligence (AI) technology has proliferated rapidly and widely into application areas such as speech recognition, health care, and autonomous driving. To increase the capabilities of AI more powerful systems are needed to process a larger amount of data. This requirement has made domain-specific accelerators, such as GPUs and TPUs, popular; as they can provide orders of magnitude higher performance than state-of-the-art CPUs. However, these accelerators can only operate at their peak performance when they get the necessary data from memory as quickly as it is processed: requiring off-chip memory with a high bandwidth and a large capacity [1]. HBM has thus far met the bandwidth and capacity requirement [2] –[6], but recent AI technologies such as recurrent neural networks require an even higher bandwidth than HBM [7]–[8]. While a further increase in off-chip bandwidth can be accomplished by various techniques, it is often limited by power constraints at the chip or system level [9]. Hence, it is essential to decrease demand for off-chip bandwidth with unconventional architectures: such as processing-in-memory. In this paper, we present function-In-memory DRAM (FIMDRAM) that integrates a 16-wide single-instruction multiple-data engine within the memory banks and that exploits bank-level parallelism to provide $4 \times $ higher processing bandwidth than an off-chip memory solution. Second, we show techniques that do not require any modification to conventional memory controllers and their command protocols, which make FIMDRAM more practical for quick industry adoption. Finally, we conclude this paper with circuit- and system-level evaluations of our fabricated FIMDRAM.

This paper proposes an determination algorithm of amplify-and-forward (AF) relay precoders and user receivers to minimize sum mean-squared-error (SMSE) for multi-relay aided multi-user (MRAMU) multiple-input multiple-output (MIMO) network. This algorithm is advantageous in that it can be executed with only local channel state information (CSI) at relays, meaning that a relay is aware of 2nd hop CSI of only the channel from itself to all users. Block diagonalization (BD) precoder at base station (BS) can make it unnecessary that 2nd hop CSI should be shared among relays. Given fixed BD precoder at BS, relay precoders and user receivers are designed as solution of optimization problem. We attempt to achieve suboptimal structures of relay precoders and user receivers based on sequential iteration and Karush-Kuhn-Tucker (KKT) theorem. Simulation results verify that the proposed algorithm provides the reduced SMSE compared to conventional simple AF (SAF) and minimum mean-squared-error (MMSE) relaying.

This paper investigates the comparison of two different channelization schemes of relay links in the relay aided cellular network for different transmit power and the number of relays. One is the orthogonal relay subchannel allocation (ORRSA) which assigns each relay links to mobiles with orthogonal subchannels, leading to no interference between relays in the same cell. The other is the overlapped relay subchannel allocation (OVRSA) that maximizes spectral efficiency by allowing each relay to share same subchannels among them. Simulation results show that ORRSA provides slightly better cell throughput over OVRSA in multiple cell environment and both schemes of being investigated show cell throughput increases as the number of RSs per cell increases. Finally, it is observed that RS transmit power has little influence on the cell throughput performance.

In this paper, a distributed mobile relay power control (DMRPC) scheme for maximizing individual cell throughput is proposed for mobile relay station (MRS) aided system. In the system with DMRPC, the power levels of MRSs are controlled by individual cell without cell cooperation and signaling overhead, and the appropriate power levels of MRSs are selected to maximize the cell throughput of their individual serving cell. It is demonstrated by numerical simulation that DMRPC provides the better cell throughput performance than either the full power MRS aided system or conventional system without MRS does. Moreover, it is shown that the MRS aided system with DMRPC outperforms the full power MRS aided system in average cell edge throughput. Finally, it is also verified that MRS aided systems with DMRPC and the conventional system without MRS have almost identical average cell edge throughput performance, while full power MRS aided systems show worse performance in cell edge throughput performance than the system without MRS.

The paper proposes novel carrier frequency offset estimation algorithm that exploits scattered pilot symbols used for channel estimation of OFDM systems. Carefully designed pilot symbols can be incorporated to form null subcarriers by trivial manipulation of the received OFDM signals so that null subcarrier based maximum likelihood CFO estimation method can be applied, whereas conventional maximum likelihood CFO estimation using pilot symbols jointly estimates CFO and channel impulse response iteratively which is impractical. Simulation results show satisfactory performance for frequency tracking stage. The paper also provides reduced order polynomial rooting method for simple and fast CFO estimation.

Ocean Scanning Multi-spectral Imager (OSMI) is a payload on the KOrean Multi-Purpose SATellite (KOMPSAT) launched 21 December 1999. OSMI will perform worldwide ocean color monitoring for the study of biological oceanography. OSMI acquires image data of 6 spectral bands at a time by performing whisk-broom scan with a ground sample distance (GSD) of 1 km and a ground swath width of 800 km during 20% duty cycle. For flexible mission OSMI has on-orbit spectral band selectability and programmable gain and offset. OSMI performs solar calibration and dark calibration for on-orbit instrument calibration. The post-launch radiometric responsivity of OSMI is analyzed at various gain settings in the viewpoint of the instrument developer for the instrument calibration. This analysis is carried out based on the laboratory performance measurement results for the primary spectral bands of OSMI and by using linear curve fitting method with the OSMI on-board calibration data acquired during the early operation phase. This analysis is considered to be useful for the OSMI image data calibration and user's understanding about OSMI image quality.