Sponsor Editorial

Broadcast snooping works at small agent counts. Past roughly 8 coherent agents, the math breaks: snoop traffic grows as the square of agent count, link bandwidth fills up fast, and every production SoC above that ceiling is directory-based. The interesting design problem isn't whether to use a directory. It's how to scale the directory itself.

Here's what the paper covers:
🔹 Why broadcast snooping breaks — the actual cost model, not the theoretical one, with measured numbers
 🔹 Sparse and coarse-vector directory formats — how production fabrics cut storage from intractable to tractable
 🔹 DMT, DCT, and DWT (CHI Rev E.b) — direct-transfer features that remove the home node from the data path
 🔹 Home-node distribution — how to turn one ordering-point bottleneck into many
 🔹 How the design choices change at 8, 32, 128, and 256 agents — same protocol, different configuration at every level
 🔹 Published evidence: AmpereOne (192 cores, 64 distributed home nodes), EPYC Turin (192 cores, probe filter), AMD MI300A (multi-die coherent), NVIDIA Grace Hopper (CHI across UCIe chip-to-chip)
 🔹 The multi-die frontier — C-NOC UCIe interface in the current product, multi-die coherency extensions on the roadmap
 🔹 Heterogeneity inside the coherent domain — fully coherent, I/O coherent, non-coherent, and what happens when you assign each block wrong
 🔹 Functional safety requirements — what ASIL-B through ASIL-D needs from a coherent fabric, and where the roadmap sits