--- license: mit language: - en tags: - benchmarking - llama-cpp - llm-inference - local-llm - gpu - quantization task_categories: - tabular-regression - tabular-classification pretty_name: Poor Paul's Benchmark Results --- # Poor Paul's Benchmark Results Community-submitted LLM inference benchmark data from real consumer, prosumer, and small-business hardware. Each row is one normalized benchmark result: a specific model × quantization × hardware × configuration combination with measured performance metrics. **Use this to:** - Compare throughput, latency, and power efficiency across GPU models and quantizations - Study how concurrency and context length scale on different hardware - Build leaderboards, dashboards, and data-driven GPU purchasing decisions - Query via AI: any MCP client connected to `mcp.poorpaul.dev` can answer questions from this data directly --- ## Files | File | Description | | ---------------------- | ------------------------------------------------------------------ | | `data/results_*.jsonl` | Raw benchmark submissions, one file per run, appended continuously | | `llms.txt` | Machine-readable summary for LLM context injection | --- ## Quick start ```python import pandas as pd from datasets import load_dataset # Stream all rows (recommended for large queries) ds = load_dataset("paulplee/ppb-results", streaming=True, split="train") df = pd.DataFrame(iter(ds)) # Filter to throughput rows for a specific GPU tput = df[ (df["gpu_name"] == "NVIDIA GeForce RTX 5090") & (df["runner_type"] == "llama-bench") ][["model_base", "quant", "n_ctx", "throughput_tok_s"]] ``` Or via DuckDB directly from HuggingFace: ```sql SELECT gpu_name, model_base, quant, concurrent_users, AVG(throughput_tok_s) AS mean_tok_s, COUNT(*) AS n FROM read_parquet('hf://datasets/paulplee/ppb-results/data/*.jsonl') WHERE runner_type = 'llama-bench' GROUP BY ALL ORDER BY mean_tok_s DESC; ``` --- ## Schema (v0.9.0) All columns are present on every row. Fields that do not apply to a given runner are `null`. ### Model identity | Column | Type | Description | | ------------- | ------------ | ---------------------------------------------------------------------------- | | `run_type` | string | `quantitative`, `qualitative`, or `all` | | `model` | string | Full model path (e.g. `unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf`) | | `model_base` | string | Base model name without quant suffix (e.g. `Qwen3.5-9B`) | | `quant` | string | Quantization format (e.g. `Q4_K_M`, `Q8_0`, `BF16`) | | `model_org` | string\|null | HuggingFace organisation (e.g. `unsloth`); null for local paths | | `model_repo` | string\|null | Full HF `org/repo` string; null for local paths | | `runner_type` | string | Benchmark backend: `llama-bench`, `llama-server`, or `llama-server-loadtest` | ### LLM engine | Column | Type | Description | | -------------------- | ------------ | -------------------------------------------------------- | | `llm_engine_name` | string\|null | Inference engine (e.g. `llama.cpp`) | | `llm_engine_version` | string\|null | Engine version with build hash (e.g. `b5063 (58ab80c3)`) | ### Hardware | Column | Type | Description | | ------------------------ | ------------ | ----------------------------------------------------------------------- | | `gpu_name` | string\|null | Primary GPU model name | | `gpu_vram_gb` | float\|null | Primary GPU VRAM in GB | | `gpu_driver` | string\|null | GPU driver version | | `gpu_count` | int | Number of GPUs used | | `gpu_names` | string\|null | Comma-joined list of all GPU names (multi-GPU runs) | | `gpu_total_vram_gb` | float\|null | Total VRAM across all GPUs | | `unified_memory` | bool\|null | `true` for Apple Silicon — GPU and CPU share the same memory pool | | `gpu_compute_capability` | string\|null | CUDA compute capability (e.g. `"9.0"` for Blackwell); null for non-CUDA | | `gpu_pcie_gen` | int\|null | PCIe generation (e.g. `5`); null for unified-memory platforms | | `gpu_pcie_width` | int\|null | PCIe link width in lanes (e.g. `16`); null for unified-memory platforms | | `gpu_power_limit_w` | float\|null | Configured TDP limit in Watts (from NVML); null for non-NVIDIA | | `backends` | string\|null | Compute backend with version (e.g. `CUDA 13.0`, `Metal`, `CPU`) | | `cpu_model` | string\|null | CPU model name | ### Benchmark configuration | Column | Type | Description | | ------------------ | ------------ | ------------------------------------------------------------------------------------------------------------------------------- | | `n_ctx` | int\|null | Context window size in tokens | | `n_batch` | int\|null | Batch size for prompt processing | | `split_mode` | string\|null | Multi-GPU split strategy (`layer`, `row`, `none`); null for single-GPU | | `tensor_split` | string\|null | Per-GPU VRAM weight string (e.g. `"1,1"`); null for single-GPU | | `concurrent_users` | int\|null | Number of simulated parallel users. For `llama-server-loadtest`, **each row is one concurrency level** from the measured curve. | ### Workload | Column | Type | Description | | ---------------- | ------------ | -------------------------------------------------------------- | | `task_type` | string\|null | Workload category (e.g. `text-generation`, `context-rot-niah`) | | `prompt_dataset` | string\|null | Prompt source (e.g. `sharegpt-v3`); null for `llama-bench` | | `num_prompts` | int\|null | Prompts sent per run; null for `llama-bench` | | `n_predict` | int\|null | Max tokens generated per prompt; null for `llama-bench` | ### Performance — throughput | Column | Type | Description | | ------------------- | ----------- | ------------------------------------------------------------------- | | `throughput_tok_s` | float\|null | Tokens per second (primary throughput metric) | | `vram_cliff_tokens` | int\|null | Largest `n_ctx` that loaded without OOM during pre-flight discovery | ### Performance — power | Column | Type | Description | | ------------- | ----------- | ------------------------------- | | `avg_power_w` | float\|null | Average GPU power draw in Watts | | `max_power_w` | float\|null | Peak GPU power draw in Watts | ### Performance — thermal | Column | Type | Description | | ------------------- | ----------- | ---------------------------- | | `avg_gpu_temp_c` | float\|null | Average GPU temperature (°C) | | `max_gpu_temp_c` | float\|null | Peak GPU temperature (°C) | | `avg_cpu_temp_c` | float\|null | Average CPU temperature (°C) | | `max_cpu_temp_c` | float\|null | Peak CPU temperature (°C) | | `avg_fan_speed_rpm` | float\|null | Average fan speed (RPM) | | `max_fan_speed_rpm` | float\|null | Peak fan speed (RPM) | ### Performance — user experience (server runners only) | Column | Type | Description | | ------------- | ----------- | -------------------------------- | | `avg_ttft_ms` | float\|null | Average Time-To-First-Token (ms) | | `p50_ttft_ms` | float\|null | Median TTFT (ms) | | `p99_ttft_ms` | float\|null | 99th-percentile TTFT (ms) | | `avg_itl_ms` | float\|null | Average Inter-Token Latency (ms) | | `p50_itl_ms` | float\|null | Median ITL (ms) | | `p99_itl_ms` | float\|null | 99th-percentile ITL (ms) | ### Qualitative evaluation Populated when `run_type` is `qualitative` or `all`. All null for pure `quantitative` runs. | Column | Type | Description | | -------------------------------- | ------------ | ----------------------------------------------------------------------- | | `context_rot_score` | float\|null | Mean accuracy across all (length × depth) long-context recall cases | | `context_rot_accuracy_by_length` | string\|null | JSON `{haystack_length: accuracy}` map | | `context_rot_accuracy_by_depth` | string\|null | JSON `{depth_pct: accuracy}` map | | `tool_selection_accuracy` | float\|null | Fraction of cases with correct tool name selected | | `parameter_accuracy` | float\|null | Fraction of cases with all required arguments matching ground truth | | `parameter_hallucination_rate` | float\|null | Fraction of cases with invented arguments not in schema | | `parse_success_rate` | float\|null | Fraction of cases with parseable tool-call JSON | | `overall_tool_accuracy` | float\|null | Geometric mean of tool selection × parameter accuracy | | `knowledge_accuracy_mean` | float\|null | Mean fraction of factual claims judged consistent with common knowledge | | `knowledge_accuracy_std` | float\|null | Standard deviation of per-prompt knowledge-accuracy scores | | `answer_relevancy_mean` | float\|null | Mean judge-rated response relevancy (0–1) | | `coherence_mean` | float\|null | Mean judge-rated coherence (0–1) | | `quality_composite_score` | float\|null | Mean of knowledge accuracy, relevancy, and coherence | | `memory_accuracy` | float\|null | LongMemEval recall accuracy (0–1); null when MT-Bench mode was used | | `mt_bench_score` | float\|null | MT-Bench score (1–10 scale); null when LongMemEval mode was used | | `cases_evaluated` | int\|null | Number of evaluation cases that completed | | `cases_skipped_context` | int\|null | Cases skipped because context exceeded `vram_cliff_tokens` | ### Blob columns | Column | Type | Description | | -------------- | ------------ | ------------------------------------------------------------------------ | | `qualitative` | string\|null | Full qualitative result payload as JSON string | | `quantitative` | string\|null | Full quantitative result payload as JSON string | | `meta` | string\|null | Reproducibility hints (e.g. `quality_prompts_cache_hash`) as JSON string | ### OS / system context | Column | Type | Description | | ------------------- | ------------ | ------------------------------------------- | | `os_system` | string\|null | OS family: `Linux`, `Darwin`, `Windows` | | `os_release` | string\|null | Kernel / OS release string | | `os_machine` | string\|null | CPU architecture (e.g. `x86_64`, `arm64`) | | `os_distro` | string\|null | Distribution name (e.g. `Ubuntu`, `macOS`) | | `os_distro_version` | string\|null | Distribution version (e.g. `24.04`, `15.5`) | | `cpu_cores` | int\|null | Number of logical CPU cores | | `ram_total_gb` | float\|null | Total system RAM in GB | ### Submission metadata | Column | Type | Description | | -------------- | ------------ | ---------------------------------------------------- | | `submitter` | string\|null | Optional public display name of the contributor | | `timestamp` | string\|null | ISO 8601 UTC time the benchmark run produced the row | | `submitted_at` | string\|null | ISO 8601 UTC time the row was uploaded | ### Provenance and deduplication | Column | Type | Description | | --------------------- | ------------ | --------------------------------------------------------------------------- | | `schema_version` | string | Schema version at time of flattening (`0.9.0`) | | `benchmark_version` | string | PPB software version that produced the row | | `suite_run_id` | string\|null | UUID shared by all rows from the same `ppb` invocation | | `submission_id` | string\|null | UUID assigned during upload | | `row_id` | string | UUID uniquely identifying this row | | `machine_fingerprint` | string | SHA-256 of hardware profile fields (anonymous machine identity) | | `run_fingerprint` | string | SHA-256 of benchmark configuration + machine fingerprint | | `result_fingerprint` | string | SHA-256 of run identity + measured metrics — uniquely identifies one result | | `source_file_sha256` | string\|null | SHA-256 of the source JSONL file | ### Extensibility | Column | Type | Description | | ------ | ------------ | ---------------------------------------------------------------- | | `tags` | string\|null | Free-form JSON string for arbitrary metadata from the suite TOML | --- ## Null value guide Many columns are runner-specific. Expected nulls by runner type: | Column group | `llama-bench` | `llama-server` | `llama-server-loadtest` | | -------------------------------------------- | --------------------- | --------------------- | ----------------------------- | | TTFT / ITL metrics | null | populated | populated | | `prompt_dataset`, `num_prompts`, `n_predict` | null | populated | populated | | `concurrent_users` | null | populated | populated (one row per level) | | `gpu_pcie_gen`, `gpu_pcie_width` | null on Apple Silicon | null on Apple Silicon | null on Apple Silicon | | `unified_memory` | null on NVIDIA | null on NVIDIA | null on NVIDIA | | Qualitative columns | null | null | null | Qualitative columns are populated only when `run_type` is `qualitative` or `all`. --- ## Deduplication Use fingerprints to control for duplicates in analysis: ```python # Exact duplicate rows (same result, same machine, same run) df.drop_duplicates(subset=["result_fingerprint"], inplace=True) # Latest run per (gpu, model, quant, n_ctx, concurrent_users) config latest = ( df.sort_values("timestamp") .drop_duplicates(subset=["run_fingerprint"], keep="last") ) ``` --- ## Ecosystem | | | | ------------------ | ---------------------------------------------------------------------------------------------------------------- | | **Benchmark tool** | [poor-pauls-benchmark](https://github.com/paulplee/poor-pauls-benchmark) — run benchmarks and contribute results | | **MCP server** | [ppb-mcp](https://github.com/paulplee/ppb-mcp) — lets any MCP-compatible LLM client query this dataset directly | | **Analytics** | [poorpaul.dev/insights](https://poorpaul.dev/insights) — leaderboard and visual analysis | Connect any MCP client to `https://mcp.poorpaul.dev/mcp` to query this data conversationally. --- ## Contributing results 1. Clone [poor-pauls-benchmark](https://github.com/paulplee/poor-pauls-benchmark) 2. Configure `suites/my_gpu.toml` with your hardware and models 3. Run: `uv run ppb.py all suites/my_gpu.toml` 4. Results are pushed here automatically — no PR required No hardware contribution is too small. Every GPU tier that's missing from this dataset is a blind spot for the community. --- ## License Dataset content: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — contributions are attributed to their submitters. Tooling: MIT — see the [benchmark repository](https://github.com/paulplee/poor-pauls-benchmark). Third-party evaluation data included in rows: - BFCL v4 evaluation cases © UC Berkeley, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) - MT-Bench questions © LMSYS, [MIT licence](https://github.com/lm-sys/FastChat/blob/main/LICENSE) - ShareGPT prompts under their [original licence](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered)