Explorer

Exact fingerprints

• fp-h100-7b-a
• fp-h100-7b-b
• fp-mi300-7b-c

Dimensions

• gpuVendor: anyamdnvidia
• gpuModel: anyh100mi300
• kvPrecision: anybf16fp8

Fingerprints

• fp-h100-7b-a — demo/7b@rev-1 — KV fp8 — context 32768 — resident
• fp-h100-7b-b — demo/7b@rev-2 — KV bf16 — context 32768 — available
• fp-mi300-7b-c — demo/7b@rev-1 — KV fp8 — context 8192 — available

Benchmark records

• bench-1 (fp-h100-7b-a, iw-bench@r3): ttft 120 ms, decode 95 tps — measured — Pareto frontier
• bench-2 (fp-h100-7b-a, iw-bench@r3): ttft 140 ms, decode 88 tps — measured
• bench-3 (fp-h100-7b-b, iw-bench@r3): ttft 160 ms, decode 91 tps — predicted
• bench-4 (fp-mi300-7b-c, iw-bench@r2): ttft 200 ms, decode 70 tps — measured
• bench-5 (fp-mi300-7b-c, iw-bench@r2): ttft 220 ms, decode 64 tps — predicted

Distributions by fingerprint

• fp-h100-7b-a: 2 records (2 measured, 0 predicted), ttft 120–140 ms, median decode 95 tps
• fp-h100-7b-b: 1 records (0 measured, 1 predicted), ttft 160–160 ms, median decode 91 tps
• fp-mi300-7b-c: 2 records (1 measured, 1 predicted), ttft 200–220 ms, median decode 70 tps

Recommendations

• fp-h100-7b-a: frontier on both axes across its measured runs — confidence high (two measured runs, same suite revision) — regressions: none known
• fp-h100-7b-b: frontier on decode throughput, TTFT cost of the precision change — confidence medium (single predicted run, not yet measured) — regressions: ttft: +33% vs fp-h100-7b-a
• fp-mi300-7b-c: dominated on both axes; retained for AMD coverage — confidence low (one measured and one predicted run, mixed context) — regressions: decode throughput: -26% vs fp-h100-7b-a