Recent developments concerning the rapid single-flux-quantum (RSFQ) circuit family are reviewed. Elementary cells in this circuit family can generate, pass, memorize, and reproduce picosecond voltage pulses with a nominally quantized area corresponding to transfer of a single magnetic flux quantum across a Josephson junction. Functionally, each cell can be viewed as a combination of a logic gate and an output latch (register) controlled by clock pulses, which are physically similar to the signal pulses. Hand-shaking style of local exchange by the clock pulses enables one to increase complexity of the LSI RSFQ systems without loss of operating speed. The simplest components of the RSFQ circuitry have been experimentally tested at clock frequencies exceeding 100 GHz, and an increase of the speed beyond 300 GHz is expected as a result of using an up-to-date fabrication technology. This review includes a discussion of possible future developments and applications of this novel, ultrafast digital technology. >

RSFQ logic/memory family: a new Josephson-junction technology for sub-terahertz-clock-frequency digital systems

We review the pros and cons of analog and digital computation. We propose that computation that is most efficient in its use of resources is neither analog computation nor digital computation but, rather, a mixture of the two forms. For maximum efficiency, the information and information-processing resources of the hybrid form must be distributed over many wires, with an optimal signal-to-noise ratio per wire. Our results suggest that it is likely that the brain computes in a hybrid fashion and that an underappreciated and important reason for the efficiency of the human brain, which consumes only 12 W, is the hybrid and distributed nature of its architecture.

/pdf/analog-versus-digital-extrapolating-from-electronics-to-3chqn5n8p8.pdf

Analog versus digital: extrapolating from electronics to neurobiology

The authors outline a design methodology for the realisation of digital filtering structures with significantly reduced numbers of elementary arithmetic operations. The directed acyclic graphs which result from the design algorithms completely describe the filter arithmetically and may be mapped directly onto hardware or software realisations. Vertex rearrangement, retiming and edge elimination techniques are presented which facilitate the generation of a logical graph with an efficient allocation of pipeline registers. An example of the technique is given for a bit-serial realisation employing a bit-level pipeline. >

Primitive operator digital filters

Digit-serial architectures are best suited for systems requiring moderate sample rate and where area and power consumption are critical. This paper presents a new approach for designing digit-serial/parallel finite field multipliers. This approach combines both array-type and parallel multiplication algorithms, where the digit-level array-type algorithm minimizes the latency for one multiplication operation and the parallel architecture inside of each digit cell reduces both the cycle-time as well as the switching activities, hence power consumption. By appropriately constraining the feasible primitive polynomials, the mod p(x) operation involved in finite field multiplication can be performed in a more efficient way. As a result, the computation delay and energy consumption of one finite field multiplication using the proposed digit-serial/parallel architectures are significantly less than of those obtained by folding the parallel semi-systolic multipliers. Furthermore, their energy-delay products are reduced by a even larger percentage. Therefore, the proposed digit-serial/parallel architectures are attractive for both low-energy and high-performance applications.

Low-Energy Digit-Serial/Parallel Finite Field Multipliers

This paper presents the Neural Cache architecture, which re-purposes cache structures to transform them into massively parallel compute units capable of running inferences for Deep Neural Networks. Techniques to do in-situ arithmetic in SRAM arrays, create efficient data mapping and reducing data movement are proposed. The Neural Cache architecture is capable of fully executing convolutional, fully connected, and pooling layers in-cache. The proposed architecture also supports quantization in-cache. Our experimental results show that the proposed architecture can improve inference latency by 18.3X over state-of-art multi-core CPU (Xeon E5), 7.7X over server class GPU (Titan Xp), for Inception v3 model. Neural Cache improves inference throughput by 12.4X over CPU (2.2X over GPU), while reducing power consumption by 50% over CPU (53% over GPU).

David Renshaw

Papers

VLSI Signal Processing: A Bit-Serial Approach