Mihir_Gupta
New member
With the recent Computex news from Cadence and NVIDIA validating Level-5 autonomous design frameworks via specialized agents, the industry is clearly shifting toward automated design. I'm looking at how we can apply a similar automated approach to the material physics layer of advanced packaging.
The data center power wall is hitting a boiling point, with hyperscalers like Microsoft scrambling to optimize AI workloads because electricity costs are out of control. A massive chunk of that waste is a 'data movement tax'—energy lost as heat just pushing bits over copper lines between HBM and the GPU.
To cut that power waste by 70%, the most suitable approach is to pack the chips incredibly close together. But if you do that on traditional plastic or organic boards, the heat makes them warp and the connections snap. Glass is looking like the most viable material flat and rigid enough to handle it, which is why companies like Intel are pushing hard into glass substrates right now.
Here is the actual bottleneck: designing for glass sucks right now. Because the thermal properties are so different from silicon, a layout engineer has to manually draw a path, run a 12-hour thermal simulation, see where it cracks, tweak the drawing, and rerun it. It’s a painfully slow, manual back-and-forth loop that destroys manufacturing yields.
To answer the inevitable question about why a single AI model won't work and why it needs a pipelined setup: Designing for glass introduces conflicting physics problems that happen in sequence. A single model can’t handle both the electrical routing math and the structural multi-physics math simultaneously they are entirely different computing problems.
Our conceptual model relies on a clean, two-step pipeline to automate this loop rather than replacing standard physics solvers:
For example, the first model handles the initial chiplet placement and routing paths.
The second, physics-informed model is built specifically to predict thermal stress and micro-cracking on glass.
The second model instantly flags structural errors back to the first to adjust lines before anyone ever kicks off a heavy, traditional sign-off tool like Calibre or Ansys. My goal is to build a tool that automates that tedious layout-simulation-tweak loop so engineers can design high-yield glass modules in days instead of months.
For those handling physical verification or advanced packaging, where do you see the biggest structural hurdles in trying to automate this kind of multi-physics iteration loop? Am I missing something fundamental about why automating this specific iteration loop is harder than it looks?
The data center power wall is hitting a boiling point, with hyperscalers like Microsoft scrambling to optimize AI workloads because electricity costs are out of control. A massive chunk of that waste is a 'data movement tax'—energy lost as heat just pushing bits over copper lines between HBM and the GPU.
To cut that power waste by 70%, the most suitable approach is to pack the chips incredibly close together. But if you do that on traditional plastic or organic boards, the heat makes them warp and the connections snap. Glass is looking like the most viable material flat and rigid enough to handle it, which is why companies like Intel are pushing hard into glass substrates right now.
Here is the actual bottleneck: designing for glass sucks right now. Because the thermal properties are so different from silicon, a layout engineer has to manually draw a path, run a 12-hour thermal simulation, see where it cracks, tweak the drawing, and rerun it. It’s a painfully slow, manual back-and-forth loop that destroys manufacturing yields.
To answer the inevitable question about why a single AI model won't work and why it needs a pipelined setup: Designing for glass introduces conflicting physics problems that happen in sequence. A single model can’t handle both the electrical routing math and the structural multi-physics math simultaneously they are entirely different computing problems.
Our conceptual model relies on a clean, two-step pipeline to automate this loop rather than replacing standard physics solvers:
For example, the first model handles the initial chiplet placement and routing paths.
The second, physics-informed model is built specifically to predict thermal stress and micro-cracking on glass.
The second model instantly flags structural errors back to the first to adjust lines before anyone ever kicks off a heavy, traditional sign-off tool like Calibre or Ansys. My goal is to build a tool that automates that tedious layout-simulation-tweak loop so engineers can design high-yield glass modules in days instead of months.
For those handling physical verification or advanced packaging, where do you see the biggest structural hurdles in trying to automate this kind of multi-physics iteration loop? Am I missing something fundamental about why automating this specific iteration loop is harder than it looks?