The VPR toolset has been widely used in FPGA architecture and CAD research, but has not evolved over the past decade. This article describes and illustrates the use of a new version of the toolset that includes four new features: first, it supports a broad range of single-driver routing architectures, which have superior architectural and electrical properties over the prior multidriver approach (and which is now employed in the majority of FPGAs sold). Second, it can now model, for placement and routing a heterogeneous selection of hard logic blocks. This is a key (but not final) step toward the incluion of blocks such as memory and multipliers. Third, we provide optimized electrical models for a wide range of architectures in different process technologies, including a range of area-delay trade-offs for each single architecture. Finally, to maintain robustness and support future development the release includes a set of regression tests for the software. To illustrate the use of the new features, we explore several architectural issues: the FPGA area efficiency versus logic block granularity, the effect of single-driver routing, and a simple use of the heterogeneity to explore the impact of hard multipliers on wiring track count.

Field-programmable gate arrays (FPGAs) are used in a variety of markets that have differing cost, performance and power consumption requirements. While it would be ideal to serve all these markets with a single FPGA family, the diversity in the needs of these markets means that generally more than one family is appropriate. Consequently, FPGA vendors have moved to provide a diverse set of families that sit at different points in the area-speed-power design space. This paper aims to understand the circuit and architectural design attributes of FPGAs that enable tradeoffs between area and speed, and to determine the magnitude of the possible tradeoffs. This will be useful for architects seeking to determine the number of device families in a suite of offerings, as well as the changes to make between families. We explore a broad range of architectures and circuit designs and developed a transistor sizing tool that automatically optimizes each design. In this paper, we describe this tool and demonstrate that it achieves results that are comparable to past work but with vastly less effort. We then use the designs produced by the tool to explore the range of tradeoffs possible. We find that through architecture and transistor sizing changes it is possible to usefully vary the area of an FPGA by a factor of 2.0 and the performance of an FPGA by a factor of 2.1. We also observe that the range of area and delay tradeoffs possible by varying only the transistor sizing of a single architecture is larger than the ranges observed in past architectural experiments. In addition to transistor size, we note that LUT size is one of the most useful parameters for trading off area and delay.

The VPR toolset [6, 7] has been widely used to perform FPGA architecture and CAD research, but has not evolved over the past decade to include many architectural features now present in modern FPGAs. This paper describes a new version of the toolset that includes four significant features: first, it now supports a broad range of single-driver routing architectures [29, 4, 16]. Single-driver routing has significantly different architectural and electrical properties from the multi-driver approach previously modelled, and is now employed in the majority of FPGAs sold. Second, the new release can now model a heterogeneous selection of hard logic blocks, which could include the hard memory and multipliers that are now ubiquitous in FPGAs. Third, we provide optimized electrical models of a wide range of architectures in different process technologies, including a range of area-delay tradeoffs for each single architecture. Prior releases of VPR did not publish even one architecture file with accurate resistance and capacitance parameters. Finally, to maintain robustness and to support future development the release includes a set of regression tests to check functionality and quality of result of the output of the tools.

The creation of an FPGA requires extensive transistor-level design. This is necessary for both the final design, and during architecture exploration, when many different logic and routing architectures are considered. For such explorations, it is not feasible to spend significant amounts of time on transistor-level design. This paper presents an automated transistor sizing tool for FPGA architecture exploration that uses a two-phased approach - a coarse rapid phase with simple modeling followed by refinement with much more accurate models. The output of the system is a design optimized towards a specific area-delay criterion. We compare the quality of our results to prior manual and partially automated approaches. Also, our tool has been used to produce hundreds of candidate architectures which we are releasing to support future high quality explorations.

Field-programmable gate arrays (FPGAs) are used in a wide range of markets that have differing cost, performance and power consumption requirements. It would be advantageous if a single device family could serve these varied needs but the economics of catering to this wide distribution of market demands suggest more than one family is appropriate. Consequently, FPGA vendors have moved to provide a more diverse set of families that sit at different points in the area-speed-power design space.

This paper presents experimental measurements of the differences between a 90-nm CMOS field programmable gate array (FPGA) and 90-nm CMOS standard-cell application-specific integrated circuits (ASICs) in terms of logic density, circuit speed, and power consumption for core logic. We are motivated to make these measurements to enable system designers to make better informed choices between these two media and to give insight to FPGA makers on the deficiencies to attack and, thereby, improve FPGAs. We describe the methodology by which the measurements were obtained and show that, for circuits containing only look-up table-based logic and flip-flops, the ratio of silicon area required to implement them in FPGAs and ASICs is on average 35. Modern FPGAs also contain "hard" blocks such as multiplier/accumulators and block memories. We find that these blocks reduce this average area gap significantly to as little as 18 for our benchmarks, and we estimate that extensive use of these hard blocks could potentially lower the gap to below five. The ratio of critical-path delay, from FPGA to ASIC, is roughly three to four with less influence from block memory and hard multipliers. The dynamic power consumption ratio is approximately 14 times and, with hard blocks, this gap generally becomes smaller

This paper presents experimental measurements of the differences between a 90nm CMOS FPGA and 90nm CMOS Standard Cell ASICs in terms of logic density, circuit speed and power consumption. We are motivated to make these measurements to enable system designers to make better informed hoices between these two media and to give insight to FPGA makers on the deficiencies to attack and thereby improve FPGAs. In the paper, we describe the methodology by which the measurements were obtained and we show that, for circuits containing only combinational logic and flip-flops, the ratio of silicon area required to implement them in FPGAs and ASICs is on average 40. Modern FPGAs also contain "hard" blocks such as multiplier/accumulators and block memories and we find that these blocks reduce this average area gap significantly to as little as 21. The ratio of critical path delay, from FPGA to ASIC, is roughly 3 to 4, with less influence from block memory and hard multipliers. The dynamic power onsumption ratio is approximately 12 times and, with hard blocks, this gap generally becomes smaller.

Creating a new FPGA is a challenging undertaking because of the significant effort that must be spent on circuit design, layout and verification. It currently takes approximately 50 to 200 person years from architecture definition to tape-out for a new FPGA family. Such a lengthy development time is necessary because the process is primarily done manually. Simplifying and shortening the design process would be advantageous since it could reduce the time to market for new FPGAs while also enhancing architecture explorations. One way to accomplish this is through automation and, in this paper, we describe our efforts to automate the entire process by making use of a previously developed set of tools that assist in the creation of the repeatable FPGA tile [25]. Our aim is to demonstrate the feasibility of a CAD flow that uses an input FPGA architecture description to generate a layout that can be sent for fabrication. We prove the feasibility of this proposition by actually designing and fabricating a complete FPGA. Initial functional testing of the FPGA appears promising but is inconclusive at this time. Through this architecture to layout process, we investigate the issues that are faced in the architecture selection, circuit design, layout and verification of such an automatically produced FPGA. We found that there are significant savings in design time. As well, we demonstrate that we can produce a layout using automated tools that is only 36% larger than a commercial FPGA device layout. Given the significant time savings and the relatively minor area penalty, we feel that this work demonstrates that automated layout of FPGAs is practical and advantageous.

Current high-performance applications are typically implemented on large-scale general-purpose distributed or multiprocessing systems often based on commodity microprocessors. Field-Programmable Gate Arrays (FPGAs) have now reached a level of sophistication that they too could be used for such applications. In this paper we explore the feasibility of using FPGAs to implement large-scale application-specific computations by way of a case study that implements a novel molecular dynamics system. The system has been designed such that it is scalable and parallelizable. On the Transmogrifier 3 (TM3), the system performs calculations on an 8,192 particle system in 37 seconds at 26 MHz. This implementation shows that by scaling to more modern parts running at 100 MHz, a speedup of over 20 x can be achieved compared to a state-of-the-art microprocessor. This can also be achieved at less cost, using less power and taking less space than a standard microprocessor-based system, while maintaining the computational precision required.

Current high-performance supercomputing applications are typically implemented on large-scale general-purpose distributed or multiprocessing systems often based on commodity microprocessors. FPGAs have now reached a level of sophistication that they too could be used for such applications. We explore the feasibility of using FPGAs to implement large-scale application-specific computations by way of a case study that implements a novel Molecular Dynamics system. The system has been designed such that it is scalable and parallelizable. On the Transmogrifier 3, the system performs calculations on an 8,192 particle system in 37 seconds at 26MHz. This implementation shows that by scaling to more modern parts running at 100MHz and using a better architecture, a speedup of over 20x can be achieved compared to a state-of-the-art microprocessor. This can also be achieved at less cost, using less power and taking less space than a standard microprocessor-based system, while maintaining the computational precision required.

The physical layout of modern commercial FPGAs is one of the last bastions of manual VLSI layout. Our recent work has automated the FPGA layout process from architectural description to mask-level layout of the repeated FPGA tile. Here we improve on that work using two approaches: 1) by making better choices for the grouping of circuitry into the cells used for the layout and 2) through better allocation of metal. The new groupings improve the FPGA tile area by between 10% and 14%. That, together with the superior metal layer allocation allows us to automatically lay out a very accurate capture of a Xilinx Virtex-E tile that is only 54% to 92% larger than the real thing. With two additional metal layers, our tile area is only 13% larger. In addition, we show that a standard cell implementation is 102% larger than the real Xilinx Virtex-E.