Hugging Face's logo

Datasets:
LocalLLMAdvisor
/
local-llm-advisor-benchmarks

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
llama.cpp
benchmark
gguf
local-llm
inference
gpu
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
local-llm-advisor-benchmarks / own_measurements_multigpu.csv
LocalLLMAdvisor's picture
LocalLLMAdvisor
Add dual RTX 3090 multi-GPU rows (21, incl. Llama 3.3 70B Q4_K_M) as separate config
462ae82 verified
Raw History Blame Contribute Delete
10.9 kB
hw_id,device_raw,device_count,backend,model,quant,params_b,size_gib,fa,ngl,threads,ubatch,depth,pp_n,pp_tps,tg_n,tg_tps,build,build_num,n_samples,date,notes
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,16,512,0,512,4697.9,128,129.87,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 210.12; tg stddev 1.15"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,16,512,4096,512,4236.57,128,123.73,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 78.28; tg stddev 0.71"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Gemma 4 26B-A4B,Q8_0,25.81,25.0,1,99,16,512,16384,512,3551.94,128,118.72,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 26B.A4B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-Q8_0.gguf; pp stddev 37.93; tg stddev 0.76"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Gemma 4 31B,Q8_0,31.27,30.38,1,99,16,512,0,512,1205.75,128,24.4,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 31B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q8_0.gguf; pp stddev 26.01; tg stddev 0.02"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Gemma 4 31B,Q8_0,31.27,30.38,1,99,16,512,4096,512,1088.86,128,23.57,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 31B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q8_0.gguf; pp stddev 10.54; tg stddev 0.02"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Gemma 4 31B,Q8_0,31.27,30.38,1,99,16,512,16384,512,904.09,128,22.94,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'gemma4 31B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-Q8_0.gguf; pp stddev 14.24; tg stddev 0.02"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Llama 3.3 70B,Q4_K_M,70.55,39.59,1,99,16,512,0,512,555.05,128,18.98,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'Llama 3.3 70B'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Llama-3.3-70B-Instruct-GGUF/Llama-3.3-70B-Instruct-Q4_K_M.gguf; pp stddev 3.55; tg stddev 0.00"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Llama 3.3 70B,Q4_K_M,70.55,39.59,1,99,16,512,4096,512,511.21,128,18.43,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'Llama 3.3 70B'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Llama-3.3-70B-Instruct-GGUF/Llama-3.3-70B-Instruct-Q4_K_M.gguf; pp stddev 4.54; tg stddev 0.01"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Llama 3.3 70B,Q4_K_M,70.55,39.59,1,99,16,512,16384,512,412.76,128,17.03,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'Llama 3.3 70B'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Llama-3.3-70B-Instruct-GGUF/Llama-3.3-70B-Instruct-Q4_K_M.gguf; pp stddev 2.21; tg stddev 0.01"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,16,512,0,512,1794.88,128,34.18,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 74.97; tg stddev 0.01"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,16,512,4096,512,1723.45,128,33.24,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 7.87; tg stddev 0.02"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Mistral Small 3.2 24B,Q8_0,24.01,23.33,1,99,16,512,16384,512,1375.83,128,30.89,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'llama 13B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf; pp stddev 17.84; tg stddev 0.01"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3-Coder 30B-A3B,Q8_0,30.53,30.25,1,99,16,512,0,512,4465.69,128,162.6,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q8_0.gguf; pp stddev 178.00; tg stddev 1.87"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3-Coder 30B-A3B,Q8_0,30.53,30.25,1,99,16,512,4096,512,3671.17,128,150.5,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q8_0.gguf; pp stddev 23.04; tg stddev 1.32"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3-Coder 30B-A3B,Q8_0,30.53,30.25,1,99,16,512,16384,512,2269.97,128,124.32,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen3moe 30B.A3B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q8_0.gguf; pp stddev 31.60; tg stddev 0.51"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,16,512,0,512,1453.87,128,28.45,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 27.19; tg stddev 0.01"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,16,512,4096,512,1387.23,128,28.24,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 13.14; tg stddev 0.03"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3.6 27B,Q8_0,27.78,26.62,1,99,16,512,16384,512,1214.05,128,27.53,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35 27B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q8_0.gguf; pp stddev 6.03; tg stddev 0.03"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3.6 35B-A3B,Q8_0,35.95,34.36,1,99,16,512,0,512,4000.71,128,147.85,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-Q8_0.gguf; pp stddev 244.87; tg stddev 1.02"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3.6 35B-A3B,Q8_0,35.95,34.36,1,99,16,512,4096,512,3745.95,128,146.67,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-Q8_0.gguf; pp stddev 124.92; tg stddev 1.42"
2x-rtx-3090,NVIDIA GeForce RTX 3090,2,CUDA,Qwen3.6 35B-A3B,Q8_0,35.95,34.36,1,99,16,512,16384,512,3278.13,128,140.36,3e4dfa8f3,11658,3,2026-10-18,"own run, llama-bench from llama.cpp b11658 (commit 3e4dfa8f3), CUDA 12.8 build, driver 595.71.05, vast.ai, CPU AMD Ryzen Threadripper 2950X 16-Core Processor, RAM 125 GiB; llama-bench model_type 'qwen35moe 35B.A3B Q8_0'; -r 3; 2 GPUs, llama.cpp layer split (default); GGUF unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-Q8_0.gguf; pp stddev 52.68; tg stddev 1.28"