As a semiconductor product guy turned cloud software marketing hack, it’s been gratifying to witness my two worlds collide. Though I exited the realm of “mA”s and “dB”s several years ago, I still miss it and maintain huge respect for the wizardry of circuit design. Also, truth be told, I’ve missed the jargon. Perhaps it’s just me, but I never got tired of theatrically demanding “Bump the die!” and “Extract the parasitics!”. Anyway, having shifted my career towards cloud-centric enterprise software, I thought I’d seen my last circuit diagram.
As it turns out, I was wrong.
It just so happens that the semiconductor design industry is making a tectonic shift towards the public cloud. Even just a few years ago, the cloud was viewed with derision and/or fear by many in the IC design world…but that’s now changing. At Elastifile, I’ve been privileged to witness this dramatic shift in attitudes firsthand. Today, IC design firms are embracing the public cloud…and here are five key reasons why:
1. The bursty nature of IC design – IC design projects typically progress through several well-established phases (e.g. Design, Layout, Tapeout, Fabrication, etc.) and the IT requirements for each phase differ greatly. For example, simulations performed during Design often stress IT compute and storage resources. While a chip is out for Fabrication, however, those same resources may sit essentially idle. This uneven workload constitutes a problem for semiconductor IT, as CAPEX-centric IT forecasting becomes a no-win situation. If infrastructure is sized based on average utilization, IT bottlenecks will occur during critical peak loads. However, sizing to handle peak load means costly resources will frequently go underutilized. The problem is amplified for independent design houses (IDHs) performing outsourced semiconductor IP development, since they typically handle fewer simultaneous projects than a semiconductor manufacturer…and thus benefit less from load-smoothing across projects in different phases. Fortunately, public cloud integration offers an effective solution. By elastically bursting workloads to public cloud infrastructure, semiconductor design firms can expand/contract their IT environment dynamically, ensuring a consistent, close match between requirements and resources.
2. Future-proofed access to new hardware and services (e.g. machine learning) – In addition to its more mundane (though still crucial) advantages for IT performance, scale, and elasticity, the public cloud also offers unique hardware and services that would be difficult-to-impossible to deploy and manage in-house. Machine learning (ML) provides a great example of this. As semiconductor design firms (especially those managing complex fabs) strive to optimize quality, cost, and time-to-market in a viciously competitive industry, ML frameworks like TensorFlow will be leveraged to inject intelligence into every facet of their operations. The broad integration of programmatic, experience-based decision-making promises to improve everything from simulation accuracy to fabrication efficiency. Public clouds are a natural home for ML workflows as they can provide 1) a centralized location for data aggregation (for ML model training), 2) specialized hardware resources that are unavailable on-premises (e.g. Google’s TPUs), and 3) plug-and-play APIs to make integration easy. Machine learning is just one example, however. Generally speaking, the public cloud represents the future-proofed IT destination for the latest-and-greatest of everything.
3. Shrinking transistors – As semiconductor process geometries get smaller, more transistors can fit within a given physical area. This increases the amount of functionality that a space-constrained IC can deliver which, in turn, increases the complexity of the corresponding design and verification simulations. In addition, these tightly-packed IC layouts often give rise to unintended electrical and magnetic effects (parasitics!) that also need to be accounted for. As a result, design teams need more and more compute power to ensure that they can execute sufficiently robust simulations in a reasonable amount of time. When seeking massively scalable compute, there’s no better destination than the public cloud.
4. EDA tool diversification – Some EDA tool vendors offer their own hosted clouds, but these proprietary envirionments raise serious concerns regarding vendor lock-in. Today, most semiconductor design firms leverage an assortment of EDA tools from multiple vendors (e.g. Cadence, Synopsys, Mentor Graphics). For example, a design team may choose to leverage Cadence tools for front-end design…yet they may prefer Synopsys tools for back-end verification. As a result, modern design flows require flexible IT environments, where tools can be mixed, matched, and swapped as desired. Public clouds are great fits for ensuring the required agility, as they natively deliver flexible, “EDA tool vendor”-agnostic solutions.
5. Increased collaboration across teams and sites – Modern semiconductor designs are increasingly modular in nature. This modularity allows work to be efficiently distributed across multiple teams and, often, across multiple geographies. For example, when developing a multimedia ASIC (e.g. to support a mobile device), a US-based design team might be responsible for audio and video functionality…while a UK-based design team is responsible for power management…and while a China-based design team is responsible for sensor signal conditioning. While the time-to-market benefits of such parallelization are clear, efficiently synchronizing and integrating distributed design environments can be a major challenge for semiconductor IT. By leveraging the public cloud as a global hub for data management, distributed design teams can 1) rely on a single, scalable, “source of truth” for common blocks and tools and 2) simplify the integration process by merging data to a unified, globally-accessible location.