{ "duplicate_domains": [], "duplicate_domain_note": "No duplicate domains detected in this run.", "artifact": "sprint19_expert_oracle_topk", "sprint": "19", "supersedes": "results/sprint19_expert_oracle.json, which used the TOP-1 expert per layer and therefore understated traffic by roughly 8x", "claim": "The expert-placement ceiling on the real 8-of-64 selection, priced in GB/s so that affordability can be decided rather than assumed.", "expert_mb": 13.477, "ref_tokens_per_s": 50.0, "ref_note": "50 tok/s is a single slow stream. Sprint 2 measured 5,300 tok/s on a 1.5B dense model; a serving deployment is ~100x these figures.", "links_gb_s": { "PCIe4 x16": 25.0, "NVLink": 450.0, "HBM3": 3350.0 }, "per_domain": { "math": { "n_refs": 2621440, "n_experts": 1024, "distinct_experts_seen": 1023, "rows": [ { "hbm_frac_of_experts": 0.25, "capacity_experts": 256, "static_popularity_fetches": 1113496, "lru_fetches": 1143822, "belady_fetches": 601343, "prefetch_oracle_fetches": 155432, "compulsory_floor": 1023, "belady_over_static": 0.54005, "lru_over_static": 1.02723, "prefetch_over_static": 0.13959, "oracle_upside_pct": 45.995, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 36.636, "prefetch_oracle_gb_s_at_ref": 5.114, "belady_gb_s_at_ref": 19.7852 }, { "hbm_frac_of_experts": 0.5, "capacity_experts": 512, "static_popularity_fetches": 518917, "lru_fetches": 567266, "belady_fetches": 208605, "prefetch_oracle_fetches": 22569, "compulsory_floor": 1023, "belady_over_static": 0.402, "lru_over_static": 1.09317, "prefetch_over_static": 0.04349, "oracle_upside_pct": 59.8, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 17.0733, "prefetch_oracle_gb_s_at_ref": 0.7426, "belady_gb_s_at_ref": 6.8635 }, { "hbm_frac_of_experts": 0.75, "capacity_experts": 768, "static_popularity_fetches": 144086, "lru_fetches": 129487, "belady_fetches": 42321, "prefetch_oracle_fetches": 965, "compulsory_floor": 1023, "belady_over_static": 0.29372, "lru_over_static": 0.89868, "prefetch_over_static": 0.0067, "oracle_upside_pct": 70.628, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 4.7407, "prefetch_oracle_gb_s_at_ref": 0.0318, "belady_gb_s_at_ref": 1.3924 } ], "best_oracle_upside_pct": 70.628, "peak_at_hbm_frac": 0.75, "peak_note": "SPRINTS.md asks WHERE on the memory-budget axis the gain is largest, because that identifies the target customer configuration" }, "prose": { "n_refs": 2621440, "n_experts": 1024, "distinct_experts_seen": 1024, "rows": [ { "hbm_frac_of_experts": 0.25, "capacity_experts": 256, "static_popularity_fetches": 1488529, "lru_fetches": 1379061, "belady_fetches": 706496, "prefetch_oracle_fetches": 178896, "compulsory_floor": 1024, "belady_over_static": 0.47463, "lru_over_static": 0.92646, "prefetch_over_static": 0.12018, "oracle_upside_pct": 52.537, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 48.9752, "prefetch_oracle_gb_s_at_ref": 5.886, "belady_gb_s_at_ref": 23.245 }, { "hbm_frac_of_experts": 0.5, "capacity_experts": 512, "static_popularity_fetches": 768124, "lru_fetches": 668830, "belady_fetches": 253047, "prefetch_oracle_fetches": 24304, "compulsory_floor": 1024, "belady_over_static": 0.32944, "lru_over_static": 0.87073, "prefetch_over_static": 0.03164, "oracle_upside_pct": 67.056, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 25.2726, "prefetch_oracle_gb_s_at_ref": 0.7996, "belady_gb_s_at_ref": 8.3257 }, { "hbm_frac_of_experts": 0.75, "capacity_experts": 768, "static_popularity_fetches": 253074, "lru_fetches": 179372, "belady_fetches": 53868, "prefetch_oracle_fetches": 1427, "compulsory_floor": 1024, "belady_over_static": 0.21285, "lru_over_static": 0.70877, "prefetch_over_static": 0.00564, "oracle_upside_pct": 78.715, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 8.3266, "prefetch_oracle_gb_s_at_ref": 0.047, "belady_gb_s_at_ref": 1.7724 } ], "best_oracle_upside_pct": 78.715, "peak_at_hbm_frac": 0.75, "peak_note": "SPRINTS.md asks WHERE on the memory-budget axis the gain is largest, because that identifies the target customer configuration" }, "code": { "n_refs": 2621440, "n_experts": 1024, "distinct_experts_seen": 1024, "rows": [ { "hbm_frac_of_experts": 0.25, "capacity_experts": 256, "static_popularity_fetches": 934628, "lru_fetches": 1010695, "belady_fetches": 506023, "prefetch_oracle_fetches": 111527, "compulsory_floor": 1024, "belady_over_static": 0.54142, "lru_over_static": 1.08139, "prefetch_over_static": 0.11933, "oracle_upside_pct": 45.858, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 30.7509, "prefetch_oracle_gb_s_at_ref": 3.6694, "belady_gb_s_at_ref": 16.649 }, { "hbm_frac_of_experts": 0.5, "capacity_experts": 512, "static_popularity_fetches": 415596, "lru_fetches": 415135, "belady_fetches": 160029, "prefetch_oracle_fetches": 12671, "compulsory_floor": 1024, "belady_over_static": 0.38506, "lru_over_static": 0.99889, "prefetch_over_static": 0.03049, "oracle_upside_pct": 61.494, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 13.6738, "prefetch_oracle_gb_s_at_ref": 0.4169, "belady_gb_s_at_ref": 5.2652 }, { "hbm_frac_of_experts": 0.75, "capacity_experts": 768, "static_popularity_fetches": 120367, "lru_fetches": 101870, "belady_fetches": 33796, "prefetch_oracle_fetches": 882, "compulsory_floor": 1024, "belady_over_static": 0.28077, "lru_over_static": 0.84633, "prefetch_over_static": 0.00733, "oracle_upside_pct": 71.923, "denominator": "fetches under the best static (top-N popularity) placement", "static_gb_s_at_ref": 3.9603, "prefetch_oracle_gb_s_at_ref": 0.029, "belady_gb_s_at_ref": 1.1119 } ], "best_oracle_upside_pct": 71.923, "peak_at_hbm_frac": 0.75, "peak_note": "SPRINTS.md asks WHERE on the memory-budget axis the gain is largest, because that identifies the target customer configuration" } }, "best_oracle_upside_pct": 78.715, "verdict": "FUND SPRINTS 20-21", "the_bandwidth_finding": "At only 50 tok/s, static placement at 50% HBM already needs 13.7-25.3 GB/s -- at or over PCIe4 x16. The prefetch ORACLE needs 0.42-0.80 GB/s, about 30x less. So off-HBM experts are bandwidth-infeasible over PCIe at serving rates under ANY policy, and the oracle's advantage is precisely what brings the traffic into range. This identifies the target configuration the sprint's deliverable asks for: NVLink-class interconnect, or experts in peer HBM.", "does_not_prove": [ "Fetch counts converted to bytes assume every expert is equal-size and fully transferred. Partial or quantized expert transfer would change the arithmetic.", "50 tok/s is a reference point, not a measurement of this model's throughput.", "Prefill routing on 20,480 tokens per domain, one MoE model." ], "cost_usd": 0.0, "generated_utc": "2026-08-27T16:31:08.013942+00:00", "git_rev": "5409e53", "git_dirty": true, "run_id": "8ff86172427b4a62", "host_kind": "Darwin-arm64", "provenance_schema": "gpu-cert-provenance/1" }