Prospector / data /analysis /inference_speed.json
haiqian-mit's picture
Prospector v0.2.3 release
b284f5b verified
Raw History Blame Contribute Delete
20 kB
{
"replay_floor_s_per_turn": 0.05,
"latency": {
"tsmc": [
{
"arm": "our 4B stack, vLLM on an H200",
"run": "spd4-4b-h200--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 3.2,
"turns_common": 7.0,
"p25_s": 3.6,
"p50_s": 5.6,
"p75_s": 8.0,
"turns_per_episode": 12.8,
"s_per_turn": 0.44,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "our 4B stack, vLLM on an L40S",
"run": "spd4-4b-l40s--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 10.2,
"turns_common": 7.0,
"p25_s": 11.3,
"p50_s": 18.2,
"p75_s": 22.2,
"turns_per_episode": 13.7,
"s_per_turn": 1.38,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "our 4B stack, MLX on a MacBook M4 Pro",
"run": "spd4-4b-mac-mlx--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 30.0,
"turns_common": 7.0,
"p25_s": 30.4,
"p50_s": 50.4,
"p75_s": 79.8,
"turns_per_episode": 13.6,
"s_per_turn": 4.22,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gpt-5-mini + our tools",
"run": "spd5-gpt5mini-native--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 36.3,
"turns_common": 16.5,
"p25_s": 31.3,
"p50_s": 39.5,
"p75_s": 53.4,
"turns_per_episode": 17.3,
"s_per_turn": 2.29,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gpt-5.6-sol + our tools",
"run": "spd5-gpt56sol-native--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 29.0,
"turns_common": 12.5,
"p25_s": 24.4,
"p50_s": 31.3,
"p75_s": 35.9,
"turns_per_episode": 14.1,
"s_per_turn": 2.28,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gpt-6-astra + our tools",
"run": "spd5-gpt6astra-label--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 14.1,
"turns_common": 6.0,
"p25_s": 10.1,
"p50_s": 14.7,
"p75_s": 21.4,
"turns_per_episode": 7.7,
"s_per_turn": 2.32,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gemini-3.7-flash + our tools",
"run": "spd5-gemini37flash-label--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 23.7,
"turns_common": 9.5,
"p25_s": 25.6,
"p50_s": 29.8,
"p75_s": 37.9,
"turns_per_episode": 14.7,
"s_per_turn": 2.18,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gemini-3.8-flash + our tools",
"run": "gemini38flash-agent-t20-label--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 21.2,
"turns_common": 9.0,
"p25_s": 34.1,
"p50_s": 36.9,
"p75_s": 38.4,
"turns_per_episode": 16.6,
"s_per_turn": 1.92,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "claude-opus-5.5 + our tools",
"run": "opus55-agent-t20-3tool--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 26.2,
"turns_common": 6.0,
"p25_s": 21.8,
"p50_s": 38.3,
"p75_s": 56.0,
"turns_per_episode": 7.7,
"s_per_turn": 4.78,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "deepseek-v4-pro + our tools",
"run": "dsv4pro-agent-t20-3tool--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 49.2,
"turns_common": 12.5,
"p25_s": 66.1,
"p50_s": 78.0,
"p75_s": 87.8,
"turns_per_episode": 18.2,
"s_per_turn": 4.06,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "qwen3.5-397b-a17b + our tools",
"run": "qwen397-agent-t20-3tool--tsmc_core_agentic",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": 45.8,
"turns_common": 12.5,
"p25_s": 47.7,
"p50_s": 57.1,
"p75_s": 66.3,
"turns_per_episode": 18.1,
"s_per_turn": 2.92,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gemini-3.7-flash, closed book",
"run": "spd2-gemini-cb--tsmc_core__closed_book",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 5.9,
"p50_s": 7.5,
"p75_s": 10.3,
"turns_per_episode": 1.0,
"s_per_turn": 7.54,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gpt-5.6-sol, closed book",
"run": "spd2-gpt56sol-cb--tsmc_core__closed_book",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 5.9,
"p50_s": 8.6,
"p75_s": 21.3,
"turns_per_episode": 1.0,
"s_per_turn": 8.58,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gpt-6-astra, closed book",
"run": "spd2-gpt6astra-cb--tsmc_core__closed_book",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 5.8,
"p50_s": 7.8,
"p75_s": 11.4,
"turns_per_episode": 1.0,
"s_per_turn": 7.83,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gemini-3.8-flash, closed book",
"run": "spd2-gemini38-cb--tsmc_core__closed_book",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 5.7,
"p50_s": 7.9,
"p75_s": 14.5,
"turns_per_episode": 1.0,
"s_per_turn": 7.92,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "gpt-5-mini, closed book",
"run": "spd5-gpt5mini-cb--tsmc_core__closed_book",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 5.5,
"p50_s": 9.0,
"p75_s": 12.0,
"turns_per_episode": 1.0,
"s_per_turn": 8.97,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "claude-opus-5.5, closed book",
"run": "spd5-opus55-cb--tsmc_core__closed_book",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 8.5,
"p50_s": 9.4,
"p75_s": 11.9,
"turns_per_episode": 1.0,
"s_per_turn": 9.42,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "qwen3.5-397b-a17b, closed book",
"run": "spd5-qwen397-cb--tsmc_core__closed_book",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 51.3,
"p50_s": 123.2,
"p75_s": 168.0,
"turns_per_episode": 1.0,
"s_per_turn": 123.19,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
},
{
"arm": "deepseek-v4-pro, closed book",
"run": "spd5-dsv4pro-cb--tsmc_core__closed_book",
"protocol": "single",
"n": 15,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 10.5,
"p50_s": 24.8,
"p75_s": 146.0,
"turns_per_episode": 1.0,
"s_per_turn": 24.78,
"s_per_question_amortised": null,
"protocol_source": "summary.json"
}
]
},
"common_correct": {
"tsmc": [
"hard-0001",
"medium-0004"
]
},
"throughput": [
{
"arm": "our policy, 4B stack, local vLLM on a GH200",
"run": "flash20cap1024-s0--tsmc_core_agentic",
"protocol": "batched",
"n": 195,
"worked": 1.0,
"p50_worked_s": null,
"p50_common_s": null,
"turns_common": null,
"p25_s": 65.9,
"p50_s": 119.9,
"p75_s": 147.3,
"turns_per_episode": 13.3,
"s_per_turn": 8.23,
"s_per_question_amortised": 0.89,
"protocol_source": "summary.json"
}
],
"pool": {
"l40s": [
{
"arm": "width 1 (control)",
"run": "poollat-l40s-w1--tsmc_core_agentic",
"width": 1,
"n_questions": 15,
"n_episodes": 15,
"p25_s": 11.7,
"p50_s": 18.1,
"p75_s": 20.5,
"questions_per_min": 3.32,
"turns_slowest_draw": 13.6,
"s_per_turn": 1.45
},
{
"arm": "width 8",
"run": "poollat-l40s-w8--tsmc_core_agentic",
"width": 8,
"n_questions": 15,
"n_episodes": 120,
"p25_s": 45.7,
"p50_s": 59.8,
"p75_s": 71.6,
"questions_per_min": 1.0,
"turns_slowest_draw": 18.4,
"s_per_turn": 3.51
},
{
"arm": "width 17",
"run": "poollat-l40s-w17--tsmc_core_agentic",
"width": 17,
"n_questions": 15,
"n_episodes": 255,
"p25_s": 66.6,
"p50_s": 88.2,
"p75_s": 97.3,
"questions_per_min": 0.68,
"turns_slowest_draw": 19.9,
"s_per_turn": 3.92
},
{
"arm": "width 34 (the shipped pool)",
"run": "poollat-l40s-w34--tsmc_core_agentic",
"width": 34,
"n_questions": 15,
"n_episodes": 510,
"p25_s": 85.7,
"p50_s": 109.5,
"p75_s": 133.6,
"questions_per_min": 0.55,
"turns_slowest_draw": 20.0,
"s_per_turn": 5.35
}
],
"gh200": [
{
"arm": "greedy, shipped policy (Figure 7 protocol)",
"run": "flash20cap1024-s0--tsmc_core_agentic/latency",
"width": 1,
"n_questions": 15,
"n_episodes": 15,
"p25_s": 3.3,
"p50_s": 5.5,
"p75_s": 7.0,
"questions_per_min": 10.93,
"turns_slowest_draw": 11.8,
"s_per_turn": 0.5
},
{
"arm": "greedy (Figure 7 protocol)",
"run": "poollat-gh200-greedy--tsmc_core_agentic",
"width": 1,
"n_questions": 15,
"n_episodes": 15,
"p25_s": 4.7,
"p50_s": 5.9,
"p75_s": 7.0,
"questions_per_min": 10.17,
"turns_slowest_draw": 13.5,
"s_per_turn": 0.5
},
{
"arm": "width 1, sampled",
"run": "poollat-gh200-w1--tsmc_core_agentic",
"width": 1,
"n_questions": 15,
"n_episodes": 15,
"p25_s": 4.4,
"p50_s": 6.0,
"p75_s": 7.3,
"questions_per_min": 9.99,
"turns_slowest_draw": 13.5,
"s_per_turn": 0.56
},
{
"arm": "width 8",
"run": "poollat-gh200-w8--tsmc_core_agentic",
"width": 8,
"n_questions": 15,
"n_episodes": 120,
"p25_s": 13.3,
"p50_s": 19.8,
"p75_s": 23.9,
"questions_per_min": 3.02,
"turns_slowest_draw": 17.7,
"s_per_turn": 1.35
},
{
"arm": "width 17",
"run": "poollat-gh200-w17--tsmc_core_agentic",
"width": 17,
"n_questions": 15,
"n_episodes": 255,
"p25_s": 23.4,
"p50_s": 28.7,
"p75_s": 35.5,
"questions_per_min": 2.09,
"turns_slowest_draw": 18.8,
"s_per_turn": 1.72
},
{
"arm": "width 34 (the shipped pool)",
"run": "poollat-gh200-w34--tsmc_core_agentic",
"width": 34,
"n_questions": 15,
"n_episodes": 510,
"p25_s": 28.1,
"p50_s": 41.4,
"p75_s": 44.3,
"questions_per_min": 1.45,
"turns_slowest_draw": 19.8,
"s_per_turn": 1.85
}
],
"mac (4-bit)": [
{
"arm": "width 1",
"run": "poollat-q4-w1--tsmc_core_agentic",
"width": 1,
"n_questions": 15,
"n_episodes": 15,
"p25_s": 15.1,
"p50_s": 25.7,
"p75_s": 35.0,
"questions_per_min": 2.34,
"turns_slowest_draw": 15.1,
"s_per_turn": 1.77
},
{
"arm": "width 4",
"run": "poollat-q4-w4--tsmc_core_agentic",
"width": 4,
"n_questions": 15,
"n_episodes": 60,
"p25_s": 71.1,
"p50_s": 80.9,
"p75_s": 95.2,
"questions_per_min": 0.74,
"turns_slowest_draw": 18.0,
"s_per_turn": 4.08
},
{
"arm": "width 8, no prefix reuse",
"run": "poollat-q4nc-w8--tsmc_core_agentic",
"width": 8,
"n_questions": 15,
"n_episodes": 120,
"p25_s": 245.0,
"p50_s": 539.3,
"p75_s": 784.9,
"questions_per_min": 0.11,
"turns_slowest_draw": 19.1,
"s_per_turn": 27.01
}
]
},
"cost": {
"gpu_usd_per_hour": {
"gh200": 2.29
},
"api_usd_per_mtok": {
"gemini-3.7-flash": [
0.75,
3.75
],
"gpt-5.6-sol": [
4.0,
20.0
],
"gpt-6-astra": [
10.0,
50.0
],
"gemini-3.8-flash": [
0.75,
3.75
],
"gpt-5-mini": [
0.25,
2.0
],
"claude-opus-5.5": [
4.0,
20.0
],
"qwen3.5-397b-a17b": [
0.55,
3.5
],
"deepseek-v4-pro": [
0.21,
0.42
]
},
"quoted": {
"gemini-3.7-flash": "2026-09-16",
"gpt-5.6-sol": "2026-09-16",
"gpt-6-astra": "2026-09-16",
"gemini-3.8-flash": "2026-09-16",
"gpt-5-mini": "2026-10-03",
"claude-opus-5.5": "2026-10-02",
"qwen3.5-397b-a17b": "2026-10-02",
"deepseek-v4-pro": "2026-10-02"
},
"ours": [
{
"arm": "greedy, shipped policy (Figure 7 protocol)",
"card": "gh200",
"protocol": "single",
"s_per_question": 5.5,
"usd_per_1k": 3.5,
"questions_per_usd": 285.8,
"gated_acc": 0.839
},
{
"arm": "greedy (Figure 7 protocol)",
"card": "gh200",
"protocol": "single",
"s_per_question": 5.9,
"usd_per_1k": 3.75,
"questions_per_usd": 266.4,
"gated_acc": 0.733
},
{
"arm": "width 8",
"card": "gh200",
"protocol": "single",
"s_per_question": 19.8,
"usd_per_1k": 12.6,
"questions_per_usd": 79.4,
"gated_acc": 0.78
},
{
"arm": "width 17",
"card": "gh200",
"protocol": "single",
"s_per_question": 28.7,
"usd_per_1k": 18.26,
"questions_per_usd": 54.8,
"gated_acc": 0.79
},
{
"arm": "width 34 (the shipped pool)",
"card": "gh200",
"protocol": "single",
"s_per_question": 41.4,
"usd_per_1k": 26.34,
"questions_per_usd": 38.0,
"gated_acc": 0.79
},
{
"arm": "our policy, 4B stack, local vLLM on a GH200 (bulk)",
"card": "gh200",
"protocol": "batched",
"s_per_question": 0.89,
"usd_per_1k": 0.57,
"questions_per_usd": 1766.4,
"gated_acc": 0.839
}
],
"api": [
{
"arm": "gpt-5-mini, closed book",
"model": "gpt-5-mini",
"prefill_tok": 68,
"generated_tok_floor": 99,
"usd_per_1k": 0.21,
"usd_per_1k_at_observed_output": 2.66,
"questions_per_usd": 4651.2,
"gated_acc": 0.205
},
{
"arm": "claude-opus-5.5, closed book",
"model": "claude-opus-5.5",
"prefill_tok": 68,
"generated_tok_floor": 266,
"usd_per_1k": 5.59,
"usd_per_1k_at_observed_output": 26.69,
"questions_per_usd": 178.8,
"gated_acc": 0.226
},
{
"arm": "qwen3.5-397b-a17b, closed book",
"model": "qwen3.5-397b-a17b",
"prefill_tok": 68,
"generated_tok_floor": 207,
"usd_per_1k": 0.76,
"usd_per_1k_at_observed_output": 4.66,
"questions_per_usd": 1312.5,
"gated_acc": 0.154
},
{
"arm": "deepseek-v4-pro, closed book",
"model": "deepseek-v4-pro",
"prefill_tok": 68,
"generated_tok_floor": 101,
"usd_per_1k": 0.06,
"usd_per_1k_at_observed_output": 0.57,
"questions_per_usd": 17636.7,
"gated_acc": 0.2
},
{
"arm": "gemini-3.7-flash, closed book",
"model": "gemini-3.7-flash",
"prefill_tok": 66,
"generated_tok_floor": 93,
"usd_per_1k": 0.4,
"usd_per_1k_at_observed_output": 5.0,
"questions_per_usd": 2511.0,
"gated_acc": 0.2
},
{
"arm": "gpt-5.6-sol, closed book",
"model": "gpt-5.6-sol",
"prefill_tok": 66,
"generated_tok_floor": 53,
"usd_per_1k": 1.32,
"usd_per_1k_at_observed_output": 26.68,
"questions_per_usd": 755.3,
"gated_acc": 0.226
},
{
"arm": "gpt-6-astra, closed book",
"model": "gpt-6-astra",
"prefill_tok": 66,
"generated_tok_floor": 65,
"usd_per_1k": 3.91,
"usd_per_1k_at_observed_output": 66.71,
"questions_per_usd": 255.8,
"gated_acc": 0.21
},
{
"arm": "gemini-3.8-flash, closed book",
"model": "gemini-3.8-flash",
"prefill_tok": 66,
"generated_tok_floor": 89,
"usd_per_1k": 0.38,
"usd_per_1k_at_observed_output": 5.0,
"questions_per_usd": 2609.3,
"gated_acc": 0.19
},
{
"arm": "gpt-5-mini + our tools",
"model": "gpt-5-mini",
"run": "gpt5mini-agent-t20-3tool-native-f16k--tsmc_core_agentic",
"source": "provider usage",
"prefill_tok": 33779,
"generated_tok": 2470,
"reasoning_tok": 1802,
"cached_prefill_share": 0.35,
"usd_per_1k": 13.39,
"questions_per_usd": 74.7,
"gated_acc": 0.549
},
{
"arm": "gpt-5.6-sol + our tools",
"model": "gpt-5.6-sol",
"run": "gpt56sol-agent-t20-3tool-native--tsmc_core_agentic",
"source": "provider usage",
"prefill_tok": 31663,
"generated_tok": 1045,
"reasoning_tok": 673,
"cached_prefill_share": 0.79,
"usd_per_1k": 147.56,
"questions_per_usd": 6.8,
"gated_acc": 0.785
},
{
"arm": "gpt-6-astra + our tools",
"model": "gpt-6-astra",
"run": "gpt6astra-agent-t20-label--tsmc_core_agentic",
"source": "provider usage",
"prefill_tok": 18750,
"generated_tok": 445,
"reasoning_tok": 86,
"cached_prefill_share": 0.72,
"usd_per_1k": 209.76,
"questions_per_usd": 4.8,
"gated_acc": 0.877
},
{
"arm": "gemini-3.7-flash + our tools",
"model": "gemini-3.7-flash",
"run": "gemini37flash-agent-t20-3tool-label--tsmc_core_agentic",
"source": "provider usage",
"prefill_tok": 28278,
"generated_tok": 1532,
"reasoning_tok": 693,
"cached_prefill_share": 0.01,
"usd_per_1k": 26.95,
"questions_per_usd": 37.1,
"gated_acc": 0.687
},
{
"arm": "gemini-3.8-flash + our tools",
"model": "gemini-3.8-flash",
"run": "gemini38flash-agent-t20-label--tsmc_core_agentic",
"source": "provider usage",
"prefill_tok": 36570,
"generated_tok": 2431,
"reasoning_tok": 1224,
"cached_prefill_share": 0.02,
"usd_per_1k": 36.54,
"questions_per_usd": 27.4,
"gated_acc": 0.641
},
{
"arm": "claude-opus-5.5 + our tools",
"model": "claude-opus-5.5",
"run": "opus55-agent-t20-3tool--tsmc_core_agentic",
"source": "provider usage",
"prefill_tok": 18413,
"generated_tok": 1154,
"reasoning_tok": 427,
"cached_prefill_share": 0.0,
"usd_per_1k": 96.73,
"questions_per_usd": 10.3,
"gated_acc": 0.846
},
{
"arm": "deepseek-v4-pro + our tools",
"model": "deepseek-v4-pro",
"run": "dsv4pro-agent-t20-3tool--tsmc_core_agentic",
"source": "provider usage",
"prefill_tok": 41651,
"generated_tok": 4702,
"reasoning_tok": 4112,
"cached_prefill_share": 0.61,
"usd_per_1k": 10.72,
"questions_per_usd": 93.3,
"gated_acc": 0.59
},
{
"arm": "qwen3.5-397b-a17b + our tools",
"model": "qwen3.5-397b-a17b",
"run": "qwen397-agent-t20-3tool--tsmc_core_agentic",
"source": "provider usage",
"prefill_tok": 44483,
"generated_tok": 2883,
"reasoning_tok": 2400,
"cached_prefill_share": 0.89,
"usd_per_1k": 34.56,
"questions_per_usd": 28.9,
"gated_acc": 0.605
}
]
},
"pool_not_measured": [
"poollat-h200-w1--tsmc_core_agentic",
"poollat-h200-w17--tsmc_core_agentic",
"poollat-h200-w34--tsmc_core_agentic",
"poollat-h200-w8--tsmc_core_agentic"
]
}