Commit a2d1e10
docs(articles): add TPM stress-test empirical-ceiling article
Empirical stress test of SmartRouter at full system load (all four groups
driven in parallel). Reports realized vs theoretical TPM/RPM at two request
sizes (~600 tok/req: 275K TPM, RPM-bound; ~5500 tok/req: 822K TPM, TPM-bound)
and identifies three structural bottlenecks: vision missing fallback chain,
SambaNova p50 latency exploding to 100s under load, and Gemini being
severely under-weighted in score routing.
Bilingual zh + en, matching existing article schema.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>1 parent 3030547 commit a2d1e10
1 file changed
Lines changed: 216 additions & 0 deletions
0 commit comments