
Modern computing systems must handle increasingly diverse and demanding workloads. Coarse-grained reconfigurable arrays (CGRAs) provide an effective solution by offering flexible datapaths that can be reconfigured at runtime for new applications, while achieving higher performance than CPUs and FPGAs through higher-level operations and spatial parallelism. Recent research has explored many CGRA architectures focusing on flexibility, power efficiency, and performance. However, most of these efforts remain small-scale and limited to RTL exploration, with few advancing to logic synthesis and even fewer achieving silicon implementations. As a result, the scalability, physical design (PD) feasibility, and post-silicon performance of such architectures remain unclear. This paper presents SPEC, a CGRA architecture and hierarchical physical design workflow that enables scalable, physically feasible systems with reliable post-silicon performance. Architecturally, SPEC uses a column-structured tile organization and places global signals and supporting cells inside the macro boundary, forming an abutted, channel-less design that enables compact and scalable physical layouts. This allows the SPEC to use a strictly hierarchical workflow, decomposing the physical design into tile-, array-, and SoC-level stages. With hardened tiles organized hierarchically, even large arrays containing up to 4,096 tiles can complete the physical design flow with clean inter-tile timing in approximately one day on a 72-core server, and yield better timing and smaller areas. In contrast, the conventional thin-channel design spends substantial time optimizing inter-tile timing and eventually times out. To evaluate post-silicon application frequencies, we validate SPEC using three CGRA prototypes implemented with SPEC and report their measured operating frequencies. To quantify the impact of our compiler-aware timing constraints, we compare a representative prototype with a simulated variant in which the constraints are removed. The resulting 9% to 25% degradation in STA frequency demonstrates that SPEC not only supports scalable physical implementation but also enables compiler-mapped applications to operate at high frequencies.
Wireless connectivity is a prerequisite for the artificial-intelligence services that now drive the semiconductor industry, and the radio-frequency (RF) transceivers that realize this connectivity must satisfy standard-driven performance targets under commercial constraints on cost, die area, and process availability. This paper reviews transceiver design considerations for commercial wireless products. The fundamental relations that link signal quality, channel bandwidth, spatial multiplexing, and link budget are first revisited, and the dominant circuit non-idealities are introduced. Commercial considerations are then explained: the importance of long-lasting process nodes, the diminishing power, performance, area, and cost returns from most advanced nodes, and RF properties of active and passive components are discussed, followed by impact due to choice of transceiver partition and interface. Three recently reported transceivers designed for commercial products—a 12-nm CMOS 4G/5G frequency-range-1 (FR1) transceiver with multiple-input multiple-output (MIMO) support, a 28/39-GHz dual-polarized 5G FR2 phased-array transceiver paired with its IF transceiver, and a 55-nm CMOS tri-band switchable Wi-Fi 7 transceiver—are then reviewed as case studies that illustrate how these considerations are resolved in practice.
The impulse sensitivity function (ISF) framework for the analysis of phase noise is revisited in light of rigorous phase-reduced models (PRMs) of oscillators originally developed in the applied mathematics literature. Starting from a general class of nonlinear oscillator models, we present a systematic methodology for deriving phase-reduced representations and computing the associated ISFs with accuracy and computational efficiency. In particular, the proposed formulation establishes a direct connection between ISFs and the phase sensitivity function (PSF) of the oscillator, which is a more fundamental quantity within the phase-reduction framework, and shows that ISFs can be expressed as certain linear combinations of its components. The proposed approach is applied to three canonical examples, including the Van der Pol–Duffing oscillator and standard Class-B and Class-D differential LC oscillators. Across all phase-noise comparisons presented, the maximum absolute deviation between the proposed approach and SpectreRF is 0.4 dB. Notably, the upconversion of flicker noise to phase noise in standard differential LC CMOS oscillator topologies is explained through the incremental Groszkowski effect and AM–PM conversion, both naturally captured through the oscillator’s phase geometry, without the need for additional phenomenological constructs. This provides a unified, physically grounded, and mathematically rigorous framework for the analysis and efficient computation of phase noise in practical oscillator designs.