Getting ready for an AI-centric semiconductor world

Just how much and fast AI is transforming semiconductor design is an area for debate. Regardless, changes are already happening, and to be successful in a more AI-centric world, organisations involved in the semiconductor supply chain must understand and plan for some significant challenges.

For instance, AI depends on the scale of Cloud computing, which is becoming increasingly expensive. In addition, AI GPUs are very costly, and the increasing cost of data centre resources is leading to escalating overheads. AI comes at a literal price.

Conversely, while AI designs are currently expensive and proprietary, expect to see a move away from these towards third-party solutions – standard AI processing units – which can be used as building blocks. When this happens, extremely efficient AI IP supply chain management will be vital to control complexity and scale, while also managing security and preventing IP leakage. These are already challenges in the semiconductor industry and will be exponentially more so if not addressed.

AI also depends on curating large amounts of test data to train AI models, but this data needs to be reviewed and cleaned to avoid pollution. New data sets need to be onboarded in a measured way, plus secondary data sets need to be weighted to influence AI outcomes appropriately. For many organisations, this level of sophisticated data management is new territory.

So, how can they address all these challenges? One method is to move away from a project-based approach towards IP-centric design closely coupled with a scalable file level data management tool, something that many of the world’s leading semiconductor organisations have or are already doing. IP-centric design creates a centralised environment for better IP management, control, and visibility.

IP-centric design practices enable IP to be traced and reused in a controlled way, which is particularly important for high-value IP such as AI GPUs, cores, and low-power components. IP data leakage can also be reduced. By implementing access control, organisations can choose who has access to what IP and have visibility of users’ actions. IP can even be geofenced. For instance, if an ordinarily authorised user briefly works remotely in a country where access to specific IP is restricted, they cannot access that IP.

All these techniques and features are typically implemented using integrated product lifecycle management (IPLM) tools, which – depending on the solution – may also include features to manage AI data sets more efficiently, such as indexing, cleaning, organising, curating, and versioning.

Also, consider implementing Software Bills of Materials (SBOMs), an increasing requirement for many complex systems. This approach allows systems integrators to have deeper traceability analysis across open-source software, accessing information buried deep in the deliverables from software providers. SBOMs and hardware bills of materials (HBOMs) can also be integrated into one environment.

BOMs, together with a more IP-centric approach, can lead towards more efficient and controlled AI-focused semiconductor design, giving organisations a solid base on which they can build. No one can accurately predict exactly how much or fast AI will transform the industry, but it is certainly will, so now is the time to get those foundations in place.