pub const PERF_JSON: &str = "{\n \"schemaVersion\": 1,\n \"generatedAt\": \"2026-08-06T00:00:00.000Z\",\n \"classes\": [\n {\n \"id\": \"nvidia-8gb-class\",\n \"label\": \"NVIDIA 8GB class (RTX 3060 Ti / 4060)\",\n \"vendor\": \"nvidia\",\n \"kind\": \"discrete\",\n \"memBandwidthGBps\": 360,\n \"efficiency\": 0.6,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.6\n },\n \"minBytes\": 6442450944,\n \"maxBytes\": 10737418240,\n \"sources\": {\n \"bandwidth\": \"NVIDIA spec sheets: RTX 3060 Ti 448 GB/s, RTX 4060 Ti 288 GB/s, RTX 4060 272 GB/s; bracket representative approx 360 GB/s\",\n \"efficiency\": \"Back-computed from llama.cpp Q4_K_M 7B decode approx 49 tok/s (weights approx 4.4 GB): 49 * 4.4 / 360 approx 0.60\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.6,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference\"\n }\n }\n }\n },\n {\n \"id\": \"nvidia-12gb-class\",\n \"label\": \"NVIDIA 12GB class (RTX 3060 12GB / 4070 / 4070 Ti)\",\n \"vendor\": \"nvidia\",\n \"kind\": \"discrete\",\n \"memBandwidthGBps\": 432,\n \"efficiency\": 0.62,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.62\n },\n \"minBytes\": 10737418240,\n \"maxBytes\": 15032385536,\n \"sources\": {\n \"bandwidth\": \"NVIDIA spec sheets: RTX 3060 12GB 360 GB/s, RTX 4070 / 4070 Ti 504 GB/s; bracket midpoint approx 432 GB/s\",\n \"efficiency\": \"Back-computed from llama.cpp Q4_K_M 7B decode approx 61 tok/s (weights approx 4.4 GB): 61 * 4.4 / 432 approx 0.62\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.62,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference\"\n }\n }\n }\n },\n {\n \"id\": \"nvidia-16gb-class\",\n \"label\": \"NVIDIA 16GB class (RTX 4070 Ti Super / 4080)\",\n \"vendor\": \"nvidia\",\n \"kind\": \"discrete\",\n \"memBandwidthGBps\": 640,\n \"efficiency\": 0.65,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.65\n },\n \"minBytes\": 15032385536,\n \"maxBytes\": 21474836480,\n \"sources\": {\n \"bandwidth\": \"NVIDIA spec sheets: cards in this 14-20 GiB bracket span 288 GB/s (RTX 4060 Ti 16GB) to 736 GB/s (RTX 4080 Super); representative approx 640 GB/s\",\n \"efficiency\": \"Back-computed from llama.cpp Q4_K_M 7B decode approx 94 tok/s (weights approx 4.4 GB): 94 * 4.4 / 640 approx 0.65\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.65,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference\"\n }\n }\n }\n },\n {\n \"id\": \"nvidia-24gb-class\",\n \"label\": \"NVIDIA 24GB class (RTX 3090 / 4090)\",\n \"vendor\": \"nvidia\",\n \"kind\": \"discrete\",\n \"memBandwidthGBps\": 950,\n \"efficiency\": 0.68,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.68\n },\n \"minBytes\": 21474836480,\n \"maxBytes\": 30064771072,\n \"sources\": {\n \"bandwidth\": \"NVIDIA spec sheets: RTX 3090 936 GB/s, RTX 4090 1008 GB/s; bracket midpoint approx 950 GB/s\",\n \"efficiency\": \"Back-computed from llama.cpp Q4_K_M 7B decode approx 147 tok/s (weights approx 4.4 GB): 147 * 4.4 / 950 approx 0.68\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.68,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference\"\n }\n }\n }\n },\n {\n \"id\": \"apple-silicon-base\",\n \"label\": \"Apple silicon base (M1/M2/M3/M4)\",\n \"vendor\": \"apple\",\n \"kind\": \"unified\",\n \"memBandwidthGBps\": 100,\n \"efficiency\": 0.7,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.7\n },\n \"minBytes\": 1073741824,\n \"maxBytes\": 25769803776,\n \"sources\": {\n \"bandwidth\": \"Apple spec: M2/M3/M4 base approx 100-120 GB/s unified (M1 base 68 GB/s); representative approx 100 GB/s. Size-based bracket approximation\",\n \"efficiency\": \"Back-computed from llama.cpp Metal Q4_K_M 7B decode approx 16 tok/s (weights approx 4.4 GB): 16 * 4.4 / 100 approx 0.70\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.7,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/ggml-org/llama.cpp/discussions/4167\"\n }\n }\n }\n },\n {\n \"id\": \"apple-silicon-pro\",\n \"label\": \"Apple silicon Pro (M-series Pro)\",\n \"vendor\": \"apple\",\n \"kind\": \"unified\",\n \"memBandwidthGBps\": 200,\n \"efficiency\": 0.7,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.7\n },\n \"minBytes\": 25769803776,\n \"maxBytes\": 42949672960,\n \"sources\": {\n \"bandwidth\": \"Apple spec: M1/M2 Pro 200 GB/s, M3 Pro 150 GB/s, M4 Pro 273 GB/s; representative approx 200 GB/s. Size-based bracket approximation\",\n \"efficiency\": \"Back-computed from llama.cpp Metal Q4_K_M 7B decode approx 32 tok/s (weights approx 4.4 GB): 32 * 4.4 / 200 approx 0.70\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.7,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/ggml-org/llama.cpp/discussions/4167\"\n }\n }\n }\n },\n {\n \"id\": \"apple-silicon-max\",\n \"label\": \"Apple silicon Max (M-series Max)\",\n \"vendor\": \"apple\",\n \"kind\": \"unified\",\n \"memBandwidthGBps\": 400,\n \"efficiency\": 0.7,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.7\n },\n \"minBytes\": 42949672960,\n \"maxBytes\": 103079215104,\n \"sources\": {\n \"bandwidth\": \"Apple spec: M1/M2 Max 400 GB/s, M3 Max 300-400 GB/s, M4 Max 546 GB/s; representative approx 400 GB/s. Size-based bracket approximation\",\n \"efficiency\": \"Back-computed from llama.cpp Metal Q4_K_M 7B decode approx 64 tok/s (weights approx 4.4 GB): 64 * 4.4 / 400 approx 0.70\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.7,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/ggml-org/llama.cpp/discussions/4167\"\n }\n }\n }\n },\n {\n \"id\": \"apple-silicon-ultra\",\n \"label\": \"Apple silicon Ultra (M-series Ultra)\",\n \"vendor\": \"apple\",\n \"kind\": \"unified\",\n \"memBandwidthGBps\": 800,\n \"efficiency\": 0.7,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.7\n },\n \"minBytes\": 103079215104,\n \"maxBytes\": 274877906944,\n \"sources\": {\n \"bandwidth\": \"Apple spec: M1/M2/M3 Ultra 800 GB/s unified. Size-based bracket approximation\",\n \"efficiency\": \"Back-computed from llama.cpp Metal Q4_K_M 7B decode approx 127 tok/s (weights approx 4.4 GB): 127 * 4.4 / 800 approx 0.70\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.7,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/ggml-org/llama.cpp/discussions/4167\"\n }\n }\n }\n },\n {\n \"id\": \"cpu-desktop-ddr\",\n \"label\": \"CPU-only desktop (dual-channel DDR4/DDR5)\",\n \"vendor\": \"none\",\n \"kind\": \"cpu\",\n \"memBandwidthGBps\": 50,\n \"efficiency\": 0.5,\n \"efficiencyByBackend\": {\n \"llamacpp\": 0.5\n },\n \"minBytes\": 1073741824,\n \"maxBytes\": 1099511627776,\n \"sources\": {\n \"bandwidth\": \"Dual-channel DDR4-3200 approx 51 GB/s, DDR5-5600 approx 90 GB/s theoretical; conservative sustained approx 50 GB/s\",\n \"efficiency\": \"Back-computed from llama.cpp CPU Q4_K_M 7B decode approx 5.7 tok/s (weights approx 4.4 GB): 5.7 * 4.4 / 50 approx 0.50\",\n \"efficiencyByBackend\": {\n \"llamacpp\": {\n \"value\": 0.5,\n \"trustTier\": \"session-verified\",\n \"basisBytesPerToken\": 4400000000,\n \"url\": \"https://github.com/ggerganov/llama.cpp\"\n }\n }\n }\n }\n ]\n}\n";Expand description
Curated throughput dataset bundled into every release.