- Home
- Documents
*VLSI Implementation of Hard- and Soft-Output Sphere ... studer/papers/11VLSISOC-book.pdf¢ ...*

prev

next

out of 27

View

0Download

0

Embed Size (px)

VLSI Implementation of Hard- and Soft-Output Sphere Decoding for Wide-Band MIMO Systems

Christoph Studer1, Markus Wenk2, and Andreas Burg3

1Dept. of Electrical and Computer Engineering, Rice University, Houston, TX 77005, USA; e-mail: studer@rice.edu

2Dept. of Information Technology and Electrical Engineering, ETH Zurich, 8092 Zurich, Switzerland; e-mail: mawenk@iis.ee.ethz.ch

3School of Engineering, EPF Lausanne, 1015 Lausanne, Switzerland; e-mail: andreas.burg@epfl.ch

Abstract. Multiple-input multiple-output (MIMO) technology in com- bination with orthogonal frequency-division multiplexing (OFDM) is the key to meet the demands for data rate and link reliability of modern wide- band wireless communication systems, such as IEEE 802.11n or 3GPP- LTE. The full potential of such systems can, however, only be achieved by high-performance data-detection algorithms, which typically exhibit pro- hibitive computational complexity. Hard-output sphere decoding (SD) and soft-output single tree-search (STS) SD are promising means for re- alizing high-performance MIMO detection and have been demonstrated to enable efficient implementations in practical systems. In this chapter, we consider the design and optimization of register transfer-level imple- mentations of hard-output SD and soft-output STS-SD with minimum area-delay product, which are well-suited for wide-band MIMO systems. We explain in detail the design, implementation, and optimization of VLSI architectures and present corresponding implementation results for 130 nm CMOS technology. The reported implementations significantly outperform the area-delay product of previously reported hard-output SD and soft-output STS-SD implementations.

Key words: VLSI implementation, MIMO-OFDM communication sys- tems, sphere decoding (SD), single tree-search (STS) SD algorithm.

1 Introduction

The evolution of data rate and quality-of-service in modern wide-band wireless communication systems is fueled by novel physical-layer technologies providing high spectral efficiency and excellent link reliability. Multiple-input multiple- output (MIMO) technology [1, 2], which employs multiple antennas at both ends of the wireless link, in combination with spatial multiplexing, orthogo- nal frequency-division multiplexing (OFDM), and channel coding is believed to be the key for reliable, high-speed, and bandwidth-efficient data transmission.

2 Christoph Studer, Markus Wenk, and Andreas Burg

Therefore, MIMO-OFDM technology is incorporated in many modern wide-band wireless communication standards, such as IEEE 802.11n [3] or 3GPP-LTE [4].

In such systems, data detection, i.e., the separation of the multiplexed data streams, is (besides channel decoding) typically among the main implementa- tion challenges in terms of computational complexity and power consumption. Therefore, corresponding efficient VLSI implementations are the key to enable high-performance, low-power, and low-cost user equipment. The performance of MIMO technology critically depends on the employed data-detection algo- rithm and corresponding high-performance methods usually entail very high complexity. In particular, a straightforward implementation of hard-output maximum-likelihood (ML) detection and soft-output a-posteriori probability (APP) detection—both providing excellent error-rate performance—requires to exhaustively test all possible transmit symbols, which results in prohibitive com- plexity, even for moderate data rates and in deep submicron technologies.

The sphere-decoding (SD) algorithm [5–11] is known to be a promising means for efficient hard-output ML and soft-output APP detection. The key idea of SD is to transform MIMO detection into a tree-search problem, which can then be solved efficiently through a branch-and-bound procedure. The drawback of this approach lies in the fact that the decoding effort—measured in terms of the number of nodes to be examined during the tree search—depends on the instantaneous channel and noise realization. In the worst-case, the number of visited nodes, which typically corresponds to the number of clock cycles required for detection in VLSI [11,12], is equivalent to that of an exhaustive search [13]. Since on-chip storage and higher-layer requirements limit the processing latency that may be inferred to support the processing of received data, the worst- case complexity of SD renders its application in real-world systems extremely challenging. This challenge can be mastered by limiting the maximum decoding effort by means of early termination of the decoding process [11, 14, 15]. This approach, however, leads to a trade-off between the maximum decoding effort and the performance of the MIMO detector. Therefore, a universally applicable VLSI architecture for SD-based MIMO detection suitable for wide-band MIMO wireless communication systems must provide a robust solution allowing for the smooth adjustment of this trade-off while minimizing the required silicon area for a given minimum performance requirement.

1.1 Contributions

In this chapter, we describe an SD-based detector architecture for wide-band MIMO communication systems and detail corresponding design and implemen- tation trade-offs of hard- and soft-output SD. To this end, we first review the hard-output SD algorithm and the soft-output single tree-search (STS) SD al- gorithm. Then, a VLSI architecture suitable for efficient data detection in wide- band MIMO systems is presented and we argue that the optimization target for the parallelly-operating SD cores corresponds to minimizing the area-delay product, which differs fundamentally from minimizing the area or maximizing the throughput, as it would be the case for narrow-band systems. To arrive at SD

VLSI Implementation of Sphere Decoding for Wide-Band MIMO Systems 3

architectures that minimize the area-delay product, we start with the VLSI im- plementations for hard-output SD [12] and soft-output STS-SD [11] and propose a variety of optimizations, which improve (i.e., lower) the area-delay product of the detectors. In particular, we propose a low-complexity approximation to the Schnorr-Euchner (SE) enumeration scheme and employ pipeline interleaving, which enables us to achieve the desired design goals. We finally present imple- mentation results for 130 nm CMOS technology and perform a comparison to previously reported implementations of SD.

1.2 Outline of the Chapter

The remainder of this chapter is organized as follows. In Section 2, the MIMO system model is introduced and the employed hard-output and soft-output STS- SD algorithms are reviewed. In Section 3, we develop a receiver architecture suitable for wide-band MIMO systems and analyze the optimization goals for the SD-core implementations. The VLSI architectures for hard-output SD and soft- output STS-SD along with corresponding optimization techniques are detailed in Section 4 and Section 5, respectively. VLSI-implementation results are presented in Section 6 and we conclude in Section 7.

1.3 Notation

Matrices are set in boldface capital letters, column-vectors in boldface lowercase letters. The superscripts T and H stand for transposition and conjugate trans- position, respectively. The real and imaginary part of a complex-valued number x are denoted by

4 Christoph Studer, Markus Wenk, and Andreas Burg

channel encoder

MIMO mapper

MIMO transmitter

MIMO receiver

MIMO detector

channel decoder

Fig. 1. Coded wide-band MIMO communication system.

constellation of size 2Q and k = 1, . . . , T designates the OFDM-tone index; the maximum number of OFDM carriers corresponds to T . Each transmit vector s[k] is associated with MTQ binary values xi,b,k ∈ {0, 1}, i = 1, . . . ,MT, b = 1, . . . , Q corresponding to the bth bit of the ith entry (i.e., spatial stream) of s[k]. The baseband input-output relation of the wireless MIMO channel for each OFDM tone is given by

y[k] = H[k]s[k] + n[k] (1)

where H[k] stands for the MR ×MT complex-valued channel matrix on OFDM tone k, y[k] is the MR-dimensional received vector, and n[k] is MR-dimensional i.i.d. zero-mean complex Gaussian distributed noise with variance N0 per entry. We assume E[s[k]s[k]H ] = 1MT IMT in the following.

In the receiver, a hard-output MIMO detector computes estimates ŝ[k] for the transmit vector, which are then used to generate binary-valued estimates x̂ for the coded bit-stream x. If a soft-output MIMO detector is used, reliability information in the form of log-likelihood ratios (LLRs) Li,b,k for each coded bit xi,b,k is generated instead. For both detection schemes we assume coherent detection, i.e., the channel matrices H[k], k = 1, . . . , T , and the noise variance N0 are perfectly known by the receiver. Finally, the MIMO receiver generates estimates for the information bits b̂ using the channel decoder, which operates either on the basis of the de-interleaved (denoted by

∏−1 in Fig. 1) bit stream x̂ for hard-output MIMO detectors or on the de-interleaved sequence of LLRs Li,b,k generated by the soft-output MIMO detector. Since the MIMO detector can treat the OFDM tones independently of each other, the tone index k is omitted in the remainder of the chapter.

VLSI Implementation of Sphere Decoding for Wide-Band MIMO Systems 5

Fig. 2. MIMO detection reformulated as a tree-search problem for MT = 3 spatial streams and QPSK modulation.

2.2 ML Detection using the Sphere-Decoding Algorithm

Hard-output MIMO detection using the ML-detection rule maximizes the prob- ability of detecting the correct transmitted vector s. The ML rule for the input- output relation (1) co