After completing my presentation and poster session at DAC 2026, attending many sessions, and meeting numerous researchers, I certainly learned a great deal academically. More importantly, however, the experience allowed me to step back and look at the path I am taking from a broader perspective.
1. Interaction: All it takes to introduce yourself is the courage to start a conversation
Approaching someone who did not know me and with whom I had no prior connection required more courage than I had expected. I was afraid that my request for a conversation might feel intrusive or that the other person might judge me negatively.
Before the conference, I spent a considerable amount of time memorizing how to introduce myself and researching the areas, research directions, and recent work of the professors I hoped to meet.
However, once I actually approached them and introduced myself, I realized that what I needed was not an impressive technique or exceptional communication skill. What mattered was the substance of what I had done and studied, together with the courage required for that single moment of starting the conversation.
Trying to present myself more impressively than what I had genuinely accomplished only made me anxious and consumed even more of my time.
Reaching out to new people is still difficult for me. However, this experience gave me the courage to be more active instead of waiting for the perfect setting or opportunity to appear.
2. Presentation: A good presentation needs one central idea that can hold every detail, and that is extremely difficult
Although I prepared diligently for my presentation, I still noticed many areas where I could have done better. Watching other talks also made the difference between an effective presentation and an ineffective one very clear.
The Qualcomm keynote discussed below began with a broad, general problem. It then identified the components and bottlenecks within that problem, explained the opportunities that could emerge from them, and finally returned to the original big picture.
Although the presentation targeted a very general idea, the examples and explanations supporting it were highly concrete. For instance, when discussing re-placement, the speaker showed an actual chip layout.
I believe the same approach can be applied when presenting an individual research project. The problem addressed by a single research project is much more specific and smaller in scope. Still, if the same structure can be applied within that context, the presentation can avoid being overwhelmed by details without becoming overly abstract.
Achieving that balance is extremely difficult, and it is something I still have much to learn about.
3. Technical Insight: Ultimately, it is all about making better architectural decisions
My biggest technical takeaway was gaining a clearer view of the challenges currently facing the hardware industry and where my own research fits within that broader landscape. This theme appeared repeatedly across several sessions, particularly in the Qualcomm keynote.
Today, workloads and intelligent devices are rapidly diversifying. Workloads span radically different scales, from AR/VR to distributed LLM serving. Production environments such as data centers, edge AI systems, and autonomous vehicles also impose drastically different requirements.
To respond to this diversity, the industry faces the global challenge of quickly deploying a wide range of specialized, or entirely new products.
However, this is difficult. Meeting different requirements at different scales requires more than simple scaling. Reusing common IP or parameterizing an existing design is often insufficient. An appropriate architecture must be chosen for each requirement and scale.
The problem is that the validation process and feedback path for an architectural decision are extremely slow. As a result, each architectural decision is costly, and finding the optimal choice becomes exceedingly difficult.
There is a long path between a human architectural idea and its validation in silicon. Changing the RTL requires verification and SoC integration, followed by synthesis, timing closure, and many other steps that demand significant computation and sustained manual effort.
In other words, proposing an architectural idea can happen relatively quickly, but developing that idea into a silicon-ready product takes far more time. This makes it difficult to determine whether an idea that initially appears promising will actually prove effective.
This is similar to finding a path through a mountain covered in fog, or throwing a dart toward a target located extremely far away. A decision made now may have a significant impact, but its effect cannot be observed immediately. Reaching the dead end of a path or finally seeing where the dart lands requires a great deal of time and effort.
How can this difficult problem be addressed? I see three different approaches.
- Accelerate
Reach the target or the end of the path more quickly. This means accelerating the traditional silicon design flow algorithmically or using a specialized hardware platform. - Intermediate feedback
Place signposts along the mountain trail or predict the score by observing the trajectory of the dart. Instead of waiting for a fully accurate result from silicon, this approach provides feedback at an earlier stage using more limited information. - Automate
Minimize the human effort required throughout the process. Separate from accelerating the tool flow itself, this approach reduces bottlenecks that require direct human intervention.
These three approaches are orthogonal and can be applied simultaneously. They also share one common goal:
To reduce the cost of architectural decisions so that better decisions can be made more quickly.
The Qualcomm keynote described this goal as reducing the reasoning gap between human insight and silicon.
When viewed individually, the research projects presented at the conference seemed to address completely different problems. From a broader perspective, however, it was striking to see how many of them were moving toward this same goal.
My own research is closest to the second approach: intermediate feedback. By validating and evaluating designs with limited information and at a higher level of abstraction—before proceeding to final silicon implementation—it aims to support better architectural decisions at an earlier stage.
DAC reminded me once again how much I still do not know and how much I still have to learn. Nevertheless, I genuinely enjoy the process of gradually understanding and solving these problems. Moving forward, I hope not to become too absorbed in the problem directly in front of me, but to continue thinking about where my work belongs within the broader landscape as I find my own direction.