In most 3D work to date, people have looked at two situations: 1) a case in which power density is not a problem, and the parts of a processor and/or entire processors can be stacked atop each other, and 2) a case in which power density is limited, and storage is stacked atop processors. In this paper, we consider the case in which power density is a limitation, yet we stack processors atop processors. We also will discuss some of the physical limitations today that render many of the good ideas presented in other work impractical, and what would be required in the technology to make them feasible. In the high-performance regime, circuits are not designed to be “power efficient;” they're designed to be fast. In power-efficient design, the speed and power of a processor should be nearly proportional. In the high-performance regime, the frequency is (ever progressingly) sublinear in power. Thus, when the power density is constrained - as it is in high-performance machines, there may be opportunities to selectively exploit parallelism in workloads by running processor-on-processor systems at the same power, yet at much greater than half speed.

No abstract available.

The impact of pipeline length on both the power and performance of a microprocessor is explored both by theory and by simulation. A theory is presented for a range of power/performance metrics, BIPS m /W . The theory shows that the more important power is to the metric, the shorter the optimum pipeline length that results. For typical parameters neither BIPS/W nor BIPS 2 /W yield an optimum, i.e., a non-pipelined design is optimal. For BIPS 3 /W the optimum, averaged over all 55 workloads studied, occurs at a 22.5 FO4 design point, a 7 stage pipeline, but this value is highly dependent on the assumed growth in latch count with pipeline depth. As dynamic power grows, the optimal design point shifts to shorter pipelines. Clock gating pushes the optimum to deeper pipelines. Surprisingly, as leakage power grows, the optimum is also found to shift to deeper pipelines. The optimum pipeline depth varies for different classes of workloads: SPEC95 and SPEC2000 integer applications, traditional (legacy) database and on-line transaction processing applications, modern (e. g. web) applications, and floating point applications.

The impact of pipeline length on both the power and performance of a microprocessor is explored both theoretically and by simulation. A theory is presented for a wide range of power/performance metrics, BIPS/sup m//W. The theory shows that the more important power is to the metric, the shorter the optimum pipeline length that results. For typical parameters neither BIPS/W nor BIPS/sup 2//W yield an optimum, i.e., a non-pipelined design is optimal. For BIPS/sup 3//W the optimum, averaged over all 55 workloads studied, occurs at a 22.5 FO4 design point, a 7 stage pipeline, but this value is highly dependent on the assumed growth in latch count with pipeline depth. As dynamic power grows, the optimal design point shifts to shorter pipelines. Clock gating pushes the optimum to deeper pipelines. Surprisingly, as leakage power grows, the optimum is also found to shift to deeper pipelines. The optimum pipeline depth varies for different classes of workloads: SPEC95 and SPEC2000 integer applications, traditional (legacy) database and on-line transaction processing applications, modern (e.g. Web) applications, and floating point applications.

The impact of pipeline length on the performance of a microprocessor is explored both theoretically and by simulation. An analytical theory is presented that shows two opposing architectural parameters affect the optimal pipeline length: the degree of instruction level parallelism (superscalar) decreases the optimal pipeline length, while the lack of pipeline stalls increases the optimal pipeline length. This theory is tested by analyzing the optimal pipeline length for 35 applications representing three classes of workloads. Trace tapes are collected from SPEC95 and SPEC2000 applications, traditional (legacy) database and online transaction processing (OLTP) applications, and modern applications primarily written in Java and C++. The results show that there is a clear and significant difference in the optimal pipeline length between the SPEC workloads and both the legacy and modern applications. The SPEC applications, written in C, optimize to a shorter pipeline length than the legacy applications, largely written in assembler language, with relatively little overlap in the two distributions. Additionally, the optimal pipeline length distribution for the C++ and Java workloads overlaps with the legacy applications, suggesting similar workload characteristics. These results are explored across a wide range of superscalar processors, both in-order and out-of-order.

Neurons with threshold controlled synapses are easier to implement in VLSI technology than the more commonly studied multiplicative-type synapses. In this paper I derive a backpropagation algorithm which is suitable for networks using this type of neuron. The decision surface obtained from this type of network is composed out of elementary hyperoctahedra centered on each point in decision space. Simulations of a simple two-layer feedforward network are used to show that a network with one hidden layer can learn the logical AND, OR, and XOR functions, and in addition solve the eight-bit parity problem and the four-bit problem.