Three-dimensional packaging technologies are critical for enabling ultra-compact, massively parallel processors (MPPs) for embedded applications. Through-water optical interconnect has been proposed as a useful technology for building ultra-compact MPPs since it provides a simplified mechanism for interconnecting stacked multichip substrates. This paper presents the offset cube, a new network topology designed to exploit the packaging benefits of through-wafer optical interconnect in ultra-compact MPP systems. We validate the offset cube's topological efficiency by developing deadlock-free adaptive routing protocols with modest virtual channel requirements (only two virtual channels per link needed for full adaptivity). A preliminary analysis of router complexity suggests these protocols can be efficiently implemented in hardware. We also present a 3D mesh embedding for the offset cube. Network simulations show the offset cube performs comparably to a bidirectional 3D mesh of equal size under uniform, hot-spot, and trace-driven traffic loads. While the offset cube is not proposed as a general replacement for the mesh topology it leverages the benefits of through-wafer optical interconnect more effectively than a mesh by completely eliminating chip-to-chip wires for data signals. Hence, the offset cube is an effective topology for interconnecting ultra-compact MCM-level MPP systems.

This paper describes Pica, a fine-grain, message-passing architecture designed to efficiently support high-throughput, low-memory parallel applications, such as image processing, object recognition, and data compression. By specializing the processor and reducing local memory (4,096 36-bit words), multiple nodes can be implemented on a single chip. This allows high-performance systems for high-throughput applications to be realized at lower cost. The architecture minimizes overhead for basic parallel operations. An operand-addressed context cache and round-robin task manager support fast task swapping. Fixed-sized activation contexts simplify storage management. Word-tag synchronization bits provide low-cost synchronization. Several applications have been developed for this architecture, including thermal relaxation, matrix multiplication, JPEG image compression, and Positron Emission Tomography image reconstruction. These applications have been executed using an instrumented instruction-level simulator. The results of these experiments and an evaluation of Pica's architectural features are presented.

The rapid advance of VLSI and packaging technologies has a significant impact on system architecture. In this paper, an analytical model is used to explore the design space of interconnection networks for a 4,096 node processing system incorporating multi-node chips packaged on a single MCM substrate. Possible designs are evaluated for a two-level interconnect with separate k-ary n-cube networks for intrachip and interchip communication. An analysis of the impact several architectural and technological parameters have on the optimal network implementation (based on average no-load latency) is presented.

Recent architectural and technological advances have led to the feasibility of a new class of massively parallel processing systems based on a fine-grain, message-passing computational model. These machines provide a new alternative for the development of fast, cost-efficient Maximum Likelihood-Expectation Maximization (ML-EM) algorithmic formulations. As an important first step in determining the potential performance benefits to be gathered from such formulations, we have developed an ML-EM algorithm suitable for the high-communications, low-memory (HCLM) execution model supported by this new class of machines. Evaluation of this algorithm indicates a normalized least-square error comparable to, or better than, that obtained via a sequential ray-driven ML-EM formulation and an effective speedup in execution time (as determined via discrete-event simulation of the Pica multiprocessor system currently under development at the Georgia Institute of Technology) of well over two orders of magnitude compared to current ray-driven sequential ML-EM formulations on high-end workstations. Thus, the HCLM algorithmic formulation may provide ML-EM reconstructions within clinical time-frames.