To Appear in IEEE Computer – Draft – Subject to Change Open-source Hardware: Opportunities and Challenges Gagan Gupta 1,2, Tony Nowatzki 2, Vinay Gangadhar 2, and Karthikeyan Sankaralingam 2 1Microsoft Research 2 University of Wisconsin - Madison {gagang, tjn, vinay, karu}@cs.wisc.edu Abstract Hardware Stack Layers: Examples: Innovation in hardware is slowing due to rising costs of (3a) Consumer Novena, GoPro, Fetch Robotics, System Hardware Fitbit, Pebble Watch chip design and diminishing benefits from Moore’s law and Dennard scaling. Software innovation, on the other hand, is (3b) Enterprise Open Compute Project flourishing, helped in good measure by a thriving open-source System Hardware Open Power ecosystem. We believe that open source can similarly help TinyCircuits, (2) Circuit Boards hardware innovation, but has not yet due to several reasons. Arduino We identify these reasons and how the industry, academia, and (1) Semiconductor Our the hardware community at large can come together to address ??? Devices (IP, Cores, Tools) Focus them. 1. Introduction Figure 1: Open-source Success Stories in Hardware Advances in silicon technology and hardware architecture have been crucial in enabling new computing technologies and ap- ment [9]. This improves chances of success, fueling more plications. However, current trends, the slowing of Moore’s innovation in turn. law and Dennard scaling, are slowing silicon technology ad- vances. Meanwhile, hardware design has become ever more Open source has not pervaded the hardware industry in an complex and expensive, especially chip design. Consequently, analogous way. While open source has been fruitful at the innovation in hardware is slowing. By several accounts, the system hardware and circuit board levels (see Figure1), it has semiconductor industry’s growth is slowing and the number of been inconsequential at the semiconductor level for SoC and new hardware startups is dwindling [10]. Unfortunately this is FPGA design. This is where innovation is now needed the happening at a time when innovation is needed in many areas, most, but has stalled. We believe that open-source hardware e.g., big data processing, machine learning, augmented reality, (OSH) can drive SoC/FPGA innovation and industry growth, etc. by enabling wider community participation, low-cost develop- In contrast, innovation and revenue-growth in soft- ment, and quick adoption of new device-level technological ware is flourishing. According to Software & In- breakthroughs. formation Industry Association report, private firms Others have also argued that OSH can enable SoC/FPGA reported annual average growth of 55% for 2014 innovation [5], but no one has yet identified a practical path https://www.siia.net/Press/Software-Industry- ( forward. Hence here we address the questions: Why is OSH arXiv:1606.01980v2 [cs.AR] 11 Jun 2016 Revenue-Growth-Accelerating-and-Hiring- not as successful as OSS? How can we kick-start a vibrant /Expected-to-Jump-According-to-New-SIIA- OSH movement? OPEXEngine-Report). Google, Facebook, and Twit- ter are but a few mega-success stories of the last decade. Uber, In Section3 we first examine the benefits of open source Pinterest, and Airbnb are examples of some new successful to the software industry, and the differences and challenges startups. In 2013, software startups attracted fifteen-fold more to OSH. We then identify how academia, industry, and hob- investment than hardware startups [10]. byists can forge a substantive OSH movement, by leveraging We observe that software-based ventures aggressively lever- emerging technological trends, e.g., new FPGA platforms and age the thriving open-source software (OSS) ecosystem to stabilizing technology nodes, and emerging social trends, e.g., build products and services. For example, Facebook started easier access to industry resources and new academic OSH with PHP, Twitter and Shopify use Ruby on Rails, Uber uses efforts. Section4 shows how OSH can enable rapid innovation Node.js, and Pinterest uses Hadoop and Memcached, all open- in a hypothetical product for face recognition. Some readers source platforms. Using open-source technology helps in- may skip to Section4 for insight on how OSH can be useful novate faster, shorten time to market, and minimize invest- in practice, before returning to Section2. 1 2. Background Front-End Design Micro-architecture Open Source IP: Open Source Tools: We begin with brief background on two pieces that are likely RTL specification ISA enhancements Functionality test familiar to many readers: hardware design process and nascent Component (IP) RISC-V Cores Simulation Chisel OSH efforts. decisions OpenCores PyMTL MIAOW Verilator BERI MyHDL 2.1. Semiconductor Hardware Design Open Piton Icarus Verilog OpenSPARC Designing chips is complex and follows three broad steps as lowRISC outlined in Figure2: front-end design, back-end design, and fabrication. An array of EDA (Electronic Design Automa- Back-End Design tion) tools are used at each step. Synthesis GDS design, Open Source IP: Open Source Tools: Gate level netlist LVS & DRC Place & Route Timing Closure During front-end design, a solution is architected and is BaseJump ABC I/O placement Package Planning Open Circuit Design implemented, typically in RTL (Register Transfer Logic). The DFT Parasitic extraction YOSYS design may incorporate pre-existing components like memory VPR and bus controllers. Some components, e.g, SRAMs and I/O Lacks other back- Lacks cutting-edge pads, may only be behaviorally modeled in RTL. Once devel- end design IPs back-end design tools oped, the RTL is extensively tested and verified. Verification Fabrication often dominates the effort and cost of front-end design. Wafer fabrication Open Source IP: Open Source Tools: Back-end design transforms the RTL into a physical design. Yield engineering Final Chip & Post-silicon Deployment ready The RTL is first synthesized into a gate-level netlist. Physical validation counterparts of components that were only modeled in RTL, Lacks open source Lacks open components for source tools for e.g., the SRAMs and I/O pads, are incorporated in the netlist. fabrication fabrication The netlist is made testable. Then the gates are physically placed and wires connecting them are laid out. The physi- cal design must adhere to strict design guidelines to ensure Figure 2: Semiconductor Hardware Development integrity and manufacturability. Hence, the physical design Steps is also verified at different stages by performing timing, ther- mal, power, EM/IR (Electromigration and Voltage drop), ESD trollers, ALUs, floating-point units, USB controllers and Eth- (Electorstatic Discharge), parasitic, test vector, LVS (Layout ernet controllers. lowRISC ( www.lowrisc.org) is an open- vs Schematic), and DRC (Design Rule Checking) analysis. source SoC effort. Since back-end modules like SRAMs and Complex components, e.g., processor cores and SRAMs, I/Os are closely tied to a foundry and technology node, build- also referred to as IP (intellectual property), are typically de- ing these requires specialized knowledge that is often propri- signed by experts and reused. etary. Hence there are fewer back-end OSH modules. There In the final, fabrication step, masks are made from the are even fewer fabrication- and packaging-related OSH ef- physical design in the form of GDSII (Graphical Database forts, although BaseJump (bjump.org) is an example from System II) and the design is fabricated on silicon wafers. The academia. silicon wafer is cut into die pieces and tested. Functional die On occasion industry provides free IP for prototyping, e.g., is packaged and tested again, before being shipped. ARM provides Cortex-M0 microcontroller and foundries like Specialized EDA tools are needed in each step. For example, TSMC provide back-end IP, but not as modifiable open source. simulators are needed to verify the RTL, synthesis tools to EDA tools developed by the community include front-end generate the netlist, and place-and-route tools to place and lay tools like Verilator, new front-end languages like Chisel and out the design. Physical design verification tools are needed PyMTL, and back-end tools like Berkeley ABC, Open Circuit to check the design’s integrity. Fabrication and packaging are Design, YOSYS, and VPR. Because of the sophistication highly specialized processes, often performed by experts other of foundry design rules, back-end tools are targeted at older than the designers. technology nodes. OpenAccess is an industry-supported open- source program for EDA tools. 2.2. Open-source Hardware Efforts While not as widespread as OSS, prevailing OSH efforts target 3. Stimulating the Open-source Virtuous Cycle each step of the design process, as outlined in the last column Figure4 provides a brief history of OSS whose evolution and of Figure2. Front-end efforts are plentiful with full-fledged use is instructive to understand the path forward for OSH. In cores like RISC-V Rocket ( www.riscv.org), OpenRISC, our view, OSS works through a virtuous cycle where a com- BERI, Open Piton many-core processor, OpenSPARC, LEON munity of developers build platforms with no direct financial and GPUs like MIAOW (www.miaowgpu.org), Nyami, and incentive, and industry and startups use these to quickly build Nyuzi (www.nyuzi.org). The opencores.org repository products. In turn, industry contributes to the platforms and provides a wide assortment of modules, e.g., memory con- incentivises future platform growth (Figure3). 2 Table 1: Differences between