In the context of high-performance computing, the integration of more computing capabilities with generic cores or dedicated accelerators for artificial intelligence (AI) application is raising more and more challenges. Due to the increasing costs of advanced nodes and the difficulties of shrinking analog and circuit input output signals (IOs), alternative architecture solutions to single die are becoming mainstream. Chiplet-based systems using 3D technologies enable modular and scalable architecture and technology partitioning. Nevertheless, there are still limitations due to chiplet integration on passive interposers—silicon or organic. In this article we present the first CMOS active interposer, integrating: 1) power management without any external components; 2) distributed interconnects enabling any chiplet-to-chiplet communication; and3) system infrastructure, design-for-test, and circuit IOs. The IntAct circuit prototype integrates six chiplets in FDSOI 28-nm technology, which are 3D-stacked onto this active interposer in 65-nm process, offering a total of 96 computing cores. Full scalability of the computing system is achieved using an innovative scalable cache-coherent memory hierarchy, enabled by distributed network-on-chips, with 3-Tbit/s/mm2 high bandwidth 3D-plug interfaces using 20- $\mu \text{m}$ pitch micro-bumps, 0.6-ns/mm low latency asynchronous interconnects, while the six chiplets are locally power-supplied with 156-mW/mm2 at 82%-peak-efficiency dc–dc converters through the active interposer. Thermal dissipation is studied showing the feasibility of such approach.

/pdf/intact-a-96-core-processor-with-six-chiplets-3d-stacked-on-hllhck4f45.pdf

IntAct: A 96-Core Processor With Six Chiplets 3D-Stacked on an Active Interposer With Distributed Interconnects and Integrated Power Management

Data-parallel problems demand ever growing floating-point (FP) operations per second under tight area- and energy-efficiency constraints. In this work, we present Manticore, a general-purpose, ultra-efficient chiplet-based architecture for data-parallel FP workloads. We have manufactured a prototype of the chiplet's computational core in Globalfoundries 22FDX process and demonstrate more than 5x improvement in energy efficiency on FP intensive workloads compared to CPUs and GPUs. The compute capability at high energy and area efficiency is provided by Snitch clusters containing eight small integer cores, each controlling a large FPU. The core supports two custom ISA extensions: The SSR extension elides explicit load and store instructions by encoding them as register reads and writes. The FREP extension decouples the integer core from the FPU allowing floating-point instructions to be issued independently. These two extensions allow the single-issue core to minimize its instruction fetch bandwidth and saturate the instruction bandwidth of the FPU, achieving FPU utilization above 90%, with more than 40% of core area dedicated to the FPU.

Manticore: A 4096-core RISC-V Chiplet Architecture for Ultra-efficient Floating-point Computing

The revolution of machine learning poses an unprecedented demand for computation resources, urging more transistors on a single monolithic chip, which is not sustainable in the Post-Moore era. The multichip integration with small functional dies, called chiplets, can reduce the manufacturing cost, improve the fabrication yield, and achieve die-level reuse for different system scales. DNN workload mapping and hardware design space exploration on such multichip systems are critical, but missing in the current stage. This work provides a hierarchical and analytical framework to describe the DNN mapping on a multichip accelerator and analyze the communication overhead. Based on this framework, we propose an automatic tool called NN-Baton with a pre-design flow and a post-design flow. The pre-design flow aims to guide the chiplet granularity exploration with given area and performance budgets for the target workload. The post-design flow focuses on the workload orchestration on different computation levels - package, chiplet, and core - in the hierarchy. Compared to Simba, NN-Baton generates mapping strategies that save 22.5%~44% energy under the same computation and memory configurations. The architecture exploration demonstrates that area is a decisive factor for the chiplet granularity. For a 2048-MAC system under a 2 mm2 chiplet area constraint, the 4-chiplet implementation with 4 cores and 16 lanes of 8-size vector-MAC is always the top-pick computation allocation across several benchmarks. In contrast, the optimal memory allocation policy in the hierarchy typically depends on the neural network models.

NN-baton: DNN workload orchestration and chiplet granularity exploration for multichip accelerators

Data-parallel problems demand ever growing floating-point (FP) operations per second under tight area- and energy-efficiency constraints. In this work, we present Manticore, a general-purpose, ultraefficient chiplet-based architecture for data-parallel FP workloads. We have manufactured a prototype of the chiplet’s computational core in Globalfoundries 22FDX process and demonstrate more than 5x improvement in energy efficiency on FP intensive workloads compared to CPUs and GPUs. The compute capability at high energy and area efficiency is provided in “Snitch: A tiny pseudo dual-issue processor for area and energy efficient execution of floating-point intensive workloads,” IEEE Trans. Comput., containing eight small integer cores, each controlling a large floating-point unit (FPU). The core supports two custom ISA extensions: The SSRs extension elides explicit load and store instructions by encoding them as register reads and writes (“Stream semantic registers: A lightweight RISC-V ISA extension achieving full compute utilization in single-issue cores,” IEEE Trans. Comput.). The floating-point repetition extension decouples the integer core from the FPU allowing floating-point instructions to be issued independently. These two extensions allow the single-issue core to minimize its instruction fetch bandwidth and saturate the instruction bandwidth of the FPU, achieving FPU utilization above 90%, with more than 40% of core area dedicated to the FPU.

Manticore: A 4096-Core RISC-V Chiplet Architecture for Ultraefficient Floating-Point Computing

This paper presents the integration of resonant clocking to multi-die architectures to synchronize individual chiplets connected through an active silicon interposer. The proposed inter-chiplet synchronization through the active silicon interposer rotary oscillator array (ASI-ROA) provides a unitary clock domain to the multiple die (i.e. multiple chiplets) in the package with a very low design overhead. System performance analysis is performed with parasitics-extracted, post-layout simulation models of two different sizes of representative heterogeneous multi-die architectures, each with varying number of RISC-V cores per die. Each RISC-V core of the multi-die package belongs to the unitary clock domain, designed with ASI-ROA to operate at a frequency of 2 GHz. The proposed architecture is investigated for robustness in frequency and skew across the multi-die system (MDS) with SPICE based simulations of post layout models, demonstrating variations of only 80 MHz for a 2 GHz target frequency. The power savings are upto 41% for the overall MDS, compared to an equivalent implementation with a contemporary ADPLL used to synchronize the multiple chiplets over the active interposer. The average clock skew of the completely resonant architecture presented in this work is 8.2 ps.

Resonant Clock Synchronization With Active Silicon Interposer for Multi-Die Systems

In the context of high-performance computing and big-data applications, the quest for performance requires modular, scalable, energy-efficient, low-cost manycore systems. Partitioning the system into multiple chiplets 3D-stacked onto large-scale interposers - organic substrate [1], 2.5D passive interposer [2] or silicon bridge [3] -leads to large modular architectures and cost reductions in advanced technologies by the Known Good Die (KGD) strategy and yield management. However, these approaches lack flexible efficient long-distance communications, smooth integration of heterogeneous chiplets, and easy integration of less-scalable analog functions, such as power management [4] and system IOs. To tackle these issues, this paper presents an active interposer integrating: i) a Switched Capacitor Voltage Regulator (SCVR) for on-chip power management; ii) flexible system interconnect topologies between all chiplets for scalable cache coherency support; iii) energy-efficient 3D-plugs for dense inter-layer communication; iv) a memory-IO controller and PHY for socket communication. The chip (Fig. 2.3.7) integrates 96 cores in 6 chiplets in 28nm FDSOI CMOS, 30-stacked in a face-to-face configuration using 20µm-pitch micro-bumps (µ-bumps) onto a 200 mm2 active interposer with 40µm-pitch Through Silicon Via (TSV) middle in a 65nm technology node. Even though complex functions are integrated, active-interposer yield is high thanks to the mature 65nm node and a reduced complexity (0.08transistors/µm2), with 30% of interposer area devoted to a SCVR variability-tolerant capacitors scheme.

2.3 A 220GOPS 96-Core Processor with 6 Chiplets 3D-Stacked on an Active Interposer Offering 0.6ns/mm Latency, 3Tb/s/mm 2 Inter-Chiplet Interconnects and 156mW/mm 2 @ 82%-Peak-Efficiency DC-DC Converters

Alexandre Arriordaz

Papers

IntAct: A 96-Core Processor With Six Chiplets 3D-Stacked on an Active Interposer With Distributed Interconnects and Integrated Power Management

2.3 A 220GOPS 96-Core Processor with 6 Chiplets 3D-Stacked on an Active Interposer Offering 0.6ns/mm Latency, 3Tb/s/mm 2 Inter-Chiplet Interconnects and 156mW/mm 2 @ 82%-Peak-Efficiency DC-DC Converters