Despite continuous efforts in layout automation for full-custom circuits, including analog/mixed-signal (AMS) designs, automated layout tools have not yet been widely adopted in current industrial full-custom design flows due to the high circuit complexity and sensitivity to layout parasitics. Nevertheless, the strict design rules and grid-based restrictions in nanometer-scale FinFET nodes limit the degree of freedom in full-custom layout design and thus reduce the gap between automation tools and human experts. This paper presents AutoCRAFT, an automatic layout generator targeting region-based layouts for advanced FinFET-based full-custom circuits. AutoCRAFT uses specialized place-and-route (P&R) algorithms to handle various design constraints while adhering to typical FinFET layout styles. Verified by comprehensive post-layout analyses, AutoCRAFT has achieved promising preliminary results in generating sign-off quality layouts for industrial benchmarks.

Analog/mixed-signal (AMS) computation can be more energy efficient than digital approaches for deep learning inference, but incurs an accuracy penalty from precision loss. Prior AMS approaches focus on small networks/datasets, which can maintain accuracy even with 2b precision. We analyze applicability of AMS approaches to larger networks by proposing a generic AMS error model, implementing it in an existing training framework, and investigating its effect on ImageNet classification with ResNet-50. We demonstrate significant accuracy recovery by exposing the network to AMS error during retraining, and we show that batch normalization layers are responsible for this accuracy recovery. We also introduce an energy model to predict the requirements of high-accuracy AMS hardware running large networks and use it to show that for ADC-dominated designs, there is a direct tradeoff between energy efficiency and network accuracy. Our model predicts that achieving < 0.4% accuracy loss on ResNet-50 with AMS hardware requires a computation energy of at least ~300 fJ/MAC. Finally, we propose methods for improving the energy-accuracy tradeoff.

This article consists of a collection of slides from the author's conference presentation.

Package-level integration using multi-chip-modules (MCMs) is a promising approach for building large-scale systems. Compared to a large monolithic die, an MCM combines many smaller chiplets into a larger system, substantially reducing fabrication and design costs. Current MCMs typically only contain a handful of coarse-grained large chiplets due to the high area, performance, and energy overheads associated with inter-chiplet communication. This work investigates and quantifies the costs and benefits of using MCMs with fine-grained chiplets for deep learning inference, an application area with large compute and on-chip storage requirements. To evaluate the approach, we architected, implemented, fabricated, and tested Simba, a 36-chiplet prototype MCM system for deep-learning inference. Each chiplet achieves 4 TOPS peak performance, and the 36-chiplet MCM package achieves up to 128 TOPS and up to 6.1 TOPS/W. The MCM is configurable to support a flexible mapping of DNN layers to the distributed compute and storage units. To mitigate inter-chiplet communication overheads, we introduce three tiling optimizations that improve data locality. These optimizations achieve up to 16% speedup compared to the baseline layer mapping. Our evaluation shows that Simba can process 1988 images/s running ResNet-50 with batch size of one, delivering inference latency of 0.50 ms.

High speed chip-to-chip communication links suffer from attenuation and inter-symbol interference (ISI) causing channels to behave like transmission lines. Typical equalizer structures are digitally-controlled FIR filters that consume tens of milliwatts in power. This paper presents the design and simulation results of an active inductor equalizer for ground referenced signaling (GRS) receivers. Providing a tunable complex pole to the receiver transfer function allows for ISI cancellation with little power overhead. The proposed system equalizes at 25 Gb/s and consumes 470 μW. Compared to previous work, this system achieves a reduction in FOM between 48× and 218× in mW per Gb/s.

Many techniques have been proposed to optimize digital system timing. Each technique can be advantageous in particular applications, however they are most often applied individually rather than concurrently. The framework presented here allows for concurrent timing optimization using retiming, intentional clock skew, and wave pipelining for latch-based designed systems with single or multi-phase clocking. This optimization is formulated as a mixed integer linear program. Our integrated framework also includes a new optimization technique called resynchronization which allows for the insertion of latches in the shortest paths and thus avoids race conditions. Our work has been applied to several designs and is able to significantly reduce the clock period.

Wave-pipelining is a timing methodology used in digital systems to achieve maximal rate operation. Using this technique, new data are applied to the inputs of a combinational block before the previous outputs are available, thus effectively pipelining the combinational logic and maximizing the utilization of the logic without inserting registers. This paper presents a timing constraint formulation for the correct clocking of wave-pipelined systems. Both single- and multiple-stage systems including feedback are considered. Based on the formulation of this paper, several important new results are presented relating to performance limits of wave-pipelined circuits. These results include the specification of distinct and disjoint regions of valid operation dependent on the clock period, intentional clock skew, and the global clock latency. Also, implications and motivations for the use of accurate delay models and exact timing analysis in the determination of combinational logic delays are given, and an analogous relationship between the multi-stage system and the single-stage system in terms of performance limits is shown. The minimum clock period is obtained by clock skew optimization formulated as a linear program. In addition, important special cases are examined and their relative performance limits are analyzed.>

The authors present the organization of an interleaved wrap-around memory system for a partitionable parallel/pipeline architecture with P pipes of L processors each. The architecture is designed to efficiently support real-time image processing and computer vision algorithms, especially those requiring global data operations. The interleaved memory system makes the architecture highly scalable in that L and P can be chosen to optimize performance for particular problems and reconfigurable in that, once L and P are fixed, problems of any size can still be mapped onto the architecture. The authors demonstrate techniques and methods for mapping computational structures to the architecture by considering the case of the 1-D butterfly network (1DBN). Since many other computational structures can be mapped to 1DBN, this gives a firm application base for the architecture. The authors also demonstrate methods for scheduling and controlling the memory system.>

A high-performance partitionable parallel/pipeline architecture (P/sup 3/A) that is capable of real-time image processing is discussed. The architecture consists of P disjoint pipes of L processors each, connected together through a novel wraparound memory. Many different problem classes, including shuffle-exchange, butterfly, and tree algorithms, can be easily mapped into P/sup 3/A. The power of the architecture lies in its ability to exploit both the spatial and temporal aspects of concurrency balancing parallelism and pipelining.>