A chip architect is designing a heterogeneous 5 nm SoC with four functional blocks: CPU, GPU, L3 cache, and I/O. The GPU block is thermally limited — hitting 95°C at target density — so the team is considering a $3M cooling upgrade from server air to direct liquid cooling.
Aethermor shows a better answer exists inside the chip itself.
"We have a 50 W power budget and our GPU block is at the thermal limit. Should we upgrade cooling, or is there a way to get more throughput without changing hardware?"
Most architects would expect:
- More aggressive cooling → more thermal headroom → more compute density
- All blocks are similarly utilized
- The only way to add throughput is to add cooling capacity
Running the SoC bottleneck case study reveals:
| Block | Density | Temperature | Headroom |
|---|---|---|---|
| CPU (CMOS) | High | At limit | ~0% |
| GPU (CMOS) | High | At limit | ~0% |
| L3 Cache | Low | Well below limit | 26× headroom |
| I/O | Low | Well below limit | 15× headroom |
The cache and I/O blocks are using less than 5% of their thermal budget.
By redistributing compute density from the over-utilized GPU block to the under-utilized cache and I/O blocks, Aethermor finds:
| Strategy | Throughput Gain | Cost |
|---|---|---|
| Upgrade to liquid cooling | Marginal | $3M retrofit |
| Redistribute density across blocks | ~47% | Free |
| Switch CPU to adiabatic logic | Additional density gain | Design effort |
Don't upgrade cooling. Redistribute compute.
The 47% throughput gain from reallocation is free — it requires no hardware changes, no cooling upgrades, and no additional power. It works because heat transfer is a local phenomenon: a thermally underutilized block has capacity that cannot help a distant overheated block, but compute workload can be moved there.
This insight is difficult to get from forward-only thermal tools:
- HotSpot shows per-block temperatures but doesn't compute headroom maps or optimal reallocation in a single call
- COMSOL requires manual post-processing to extract headroom percentages
- Manual analysis requires setting up separate runs for each redistribution scenario
Aethermor's thermal_headroom_map() and optimize_power_distribution()
answer this question in one function call.
python benchmarks/case_study_soc_bottleneck.pySee also: benchmarks/case_study_soc_bottleneck.py