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
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:
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'GROUPBYALLORDERBY 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.
