AMD EPYC 8534P

AMD EPYC 8534P 64-Core testing with a AMD Cinnabar (RCB1009C BIOS) and ASPEED on Ubuntu 23.10 via the Phoronix Test Suite.

Compare your own system(s) to this result file with the Phoronix Test Suite by running the command: phoronix-test-suite benchmark 2401286-NE-AMDEPYC8520
Jump To Table - Results

View

Do Not Show Noisy Results
Do Not Show Results With Incomplete Data
Do Not Show Results With Little Change/Spread
List Notable Results
Show Result Confidence Charts

Limit displaying results to tests within:

CPU Massive 2 Tests
HPC - High Performance Computing 5 Tests
Large Language Models 2 Tests
Machine Learning 5 Tests
Python Tests 3 Tests

Statistics

Show Overall Harmonic Mean(s)
Show Overall Geometric Mean
Show Geometric Means Per-Suite/Category
Show Wins / Losses Counts (Pie Chart)
Normalize Results
Remove Outliers Before Calculating Averages

Graph Settings

Force Line Graphs Where Applicable
Convert To Scalar Where Applicable
Prefer Vertical Bar Graphs

Multi-Way Comparison

Condense Multi-Option Tests Into Single Result Graphs

Table

Show Detailed System Result Table

Run Management

Highlight
Result
Hide
Result
Result
Identifier
View Logs
Performance Per
Dollar
Date
Run
  Test
  Duration
a
January 27
  1 Hour, 30 Minutes
b
January 28
  1 Hour, 29 Minutes
c
January 28
  4 Hours, 26 Minutes
Invert Hiding All Results Option
  2 Hours, 29 Minutes

Only show results where is faster than
Only show results matching title/arguments (delimit multiple options with a comma):
Do not show results matching title/arguments (delimit multiple options with a comma):


AMD EPYC 8534POpenBenchmarking.orgPhoronix Test SuiteAMD EPYC 8534P 64-Core @ 2.30GHz (64 Cores / 128 Threads)AMD Cinnabar (RCB1009C BIOS)AMD Device 14a46 x 32 GB DRAM-4800MT/s Samsung M321R4GA0BB0-CQKMG3201GB Micron_7450_MTFDKCB3T2TFSASPEED2 x Broadcom NetXtreme BCM5720 PCIeUbuntu 23.106.5.0-5-generic (x86_64)GNOME ShellX Server 1.21.1.7GCC 13.2.0ext4640x480ProcessorMotherboardChipsetMemoryDiskGraphicsNetworkOSKernelDesktopDisplay ServerCompilerFile-SystemScreen ResolutionAMD EPYC 8534P BenchmarksSystem Logs- Transparent Huge Pages: madvise- --build=x86_64-linux-gnu --disable-vtable-verify --disable-werror --enable-bootstrap --enable-cet --enable-checking=release --enable-clocale=gnu --enable-default-pie --enable-gnu-unique-object --enable-languages=c,ada,c++,go,d,fortran,objc,obj-c++,m2 --enable-libphobos-checking=release --enable-libstdcxx-debug --enable-libstdcxx-time=yes --enable-link-serialization=2 --enable-multiarch --enable-multilib --enable-nls --enable-objc-gc=auto --enable-offload-defaulted --enable-offload-targets=nvptx-none=/build/gcc-13-FTCNCZ/gcc-13-13.2.0/debian/tmp-nvptx/usr,amdgcn-amdhsa=/build/gcc-13-FTCNCZ/gcc-13-13.2.0/debian/tmp-gcn/usr --enable-plugin --enable-shared --enable-threads=posix --host=x86_64-linux-gnu --program-prefix=x86_64-linux-gnu- --target=x86_64-linux-gnu --with-abi=m64 --with-arch-32=i686 --with-build-config=bootstrap-lto-lean --with-default-libstdcxx-abi=new --with-gcc-major-version-only --with-multilib-list=m32,m64,mx32 --with-target-system-zlib=auto --with-tune=generic --without-cuda-driver -v - Scaling Governor: acpi-cpufreq performance (Boost: Enabled) - CPU Microcode: 0xaa00212 - a: OpenJDK Runtime Environment (build 11.0.20+8-post-Ubuntu-1ubuntu1)- Python 3.11.5- gather_data_sampling: Not affected + itlb_multihit: Not affected + l1tf: Not affected + mds: Not affected + meltdown: Not affected + mmio_stale_data: Not affected + retbleed: Not affected + spec_rstack_overflow: Mitigation of safe RET + spec_store_bypass: Mitigation of SSB disabled via prctl + spectre_v1: Mitigation of usercopy/swapgs barriers and __user pointer sanitization + spectre_v2: Mitigation of Enhanced / Automatic IBRS IBPB: conditional STIBP: always-on RSB filling PBRSB-eIBRS: Not affected + srbds: Not affected + tsx_async_abort: Not affected

abcResult OverviewPhoronix Test Suite100%100%101%101%Llama.cppSVT-AV1SpeedbPyTorchLlamafileQuicksilverTensorFlowY-CruncherNeural Magic DeepSparse

AMD EPYC 8534Ptensorflow: CPU - 512 - ResNet-50quicksilver: CTS2pytorch: CPU - 16 - Efficientnet_v2_lpytorch: CPU - 512 - Efficientnet_v2_lquicksilver: CORAL2 P2llama-cpp: llama-2-70b-chat.Q5_0.ggufpytorch: CPU - 512 - ResNet-152pytorch: CPU - 16 - ResNet-152pytorch: CPU - 1 - Efficientnet_v2_ldeepsparse: BERT-Large, NLP Question Answering - Synchronous Single-Streamdeepsparse: BERT-Large, NLP Question Answering - Synchronous Single-Streamdeepsparse: BERT-Large, NLP Question Answering - Asynchronous Multi-Streamdeepsparse: BERT-Large, NLP Question Answering - Asynchronous Multi-Streampytorch: CPU - 1 - ResNet-152llamafile: mistral-7b-instruct-v0.2.Q8_0 - CPUspeedb: Update Randspeedb: Read Rand Write Randspeedb: Read While Writingspeedb: Rand Readpytorch: CPU - 16 - ResNet-50pytorch: CPU - 512 - ResNet-50quicksilver: CORAL2 P1deepsparse: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Synchronous Single-Streamdeepsparse: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Synchronous Single-Streamdeepsparse: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Asynchronous Multi-Streamdeepsparse: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Asynchronous Multi-Streamdeepsparse: NLP Document Classification, oBERT base uncased on IMDB - Asynchronous Multi-Streamdeepsparse: NLP Document Classification, oBERT base uncased on IMDB - Asynchronous Multi-Streamdeepsparse: NLP Token Classification, BERT base uncased conll2003 - Asynchronous Multi-Streamdeepsparse: NLP Token Classification, BERT base uncased conll2003 - Asynchronous Multi-Streamdeepsparse: NLP Document Classification, oBERT base uncased on IMDB - Synchronous Single-Streamdeepsparse: NLP Document Classification, oBERT base uncased on IMDB - Synchronous Single-Streamdeepsparse: BERT-Large, NLP Question Answering, Sparse INT8 - Asynchronous Multi-Streamdeepsparse: BERT-Large, NLP Question Answering, Sparse INT8 - Asynchronous Multi-Streamdeepsparse: NLP Token Classification, BERT base uncased conll2003 - Synchronous Single-Streamdeepsparse: NLP Token Classification, BERT base uncased conll2003 - Synchronous Single-Streamllamafile: wizardcoder-python-34b-v1.0.Q6_K - CPUdeepsparse: BERT-Large, NLP Question Answering, Sparse INT8 - Synchronous Single-Streamdeepsparse: BERT-Large, NLP Question Answering, Sparse INT8 - Synchronous Single-Streamdeepsparse: CV Segmentation, 90% Pruned YOLACT Pruned - Asynchronous Multi-Streamdeepsparse: CV Segmentation, 90% Pruned YOLACT Pruned - Asynchronous Multi-Streamdeepsparse: CV Segmentation, 90% Pruned YOLACT Pruned - Synchronous Single-Streamdeepsparse: CV Segmentation, 90% Pruned YOLACT Pruned - Synchronous Single-Streamdeepsparse: NLP Text Classification, DistilBERT mnli - Asynchronous Multi-Streamdeepsparse: NLP Text Classification, DistilBERT mnli - Asynchronous Multi-Streamdeepsparse: NLP Text Classification, DistilBERT mnli - Synchronous Single-Streamdeepsparse: NLP Text Classification, DistilBERT mnli - Synchronous Single-Streamdeepsparse: CV Detection, YOLOv5s COCO - Synchronous Single-Streamdeepsparse: CV Detection, YOLOv5s COCO - Synchronous Single-Streamdeepsparse: ResNet-50, Sparse INT8 - Synchronous Single-Streamdeepsparse: ResNet-50, Sparse INT8 - Synchronous Single-Streamdeepsparse: CV Detection, YOLOv5s COCO, Sparse INT8 - Synchronous Single-Streamdeepsparse: CV Detection, YOLOv5s COCO, Sparse INT8 - Synchronous Single-Streamdeepsparse: CV Detection, YOLOv5s COCO - Asynchronous Multi-Streamdeepsparse: CV Detection, YOLOv5s COCO - Asynchronous Multi-Streamdeepsparse: CV Detection, YOLOv5s COCO, Sparse INT8 - Asynchronous Multi-Streamdeepsparse: CV Detection, YOLOv5s COCO, Sparse INT8 - Asynchronous Multi-Streamtensorflow: CPU - 16 - ResNet-50deepsparse: ResNet-50, Sparse INT8 - Asynchronous Multi-Streamdeepsparse: ResNet-50, Sparse INT8 - Asynchronous Multi-Streamdeepsparse: ResNet-50, Baseline - Asynchronous Multi-Streamdeepsparse: ResNet-50, Baseline - Asynchronous Multi-Streamdeepsparse: CV Classification, ResNet-50 ImageNet - Asynchronous Multi-Streamdeepsparse: CV Classification, ResNet-50 ImageNet - Asynchronous Multi-Streamdeepsparse: ResNet-50, Baseline - Synchronous Single-Streamdeepsparse: ResNet-50, Baseline - Synchronous Single-Streamdeepsparse: CV Classification, ResNet-50 ImageNet - Synchronous Single-Streamdeepsparse: CV Classification, ResNet-50 ImageNet - Synchronous Single-Streamllama-cpp: llama-2-13b.Q4_0.ggufsvt-av1: Preset 4 - Bosphorus 4Kpytorch: CPU - 1 - ResNet-50tensorflow: CPU - 1 - ResNet-50llamafile: llava-v1.5-7b-q4 - CPUllama-cpp: llama-2-7b.Q4_0.ggufy-cruncher: 1Bsvt-av1: Preset 8 - Bosphorus 4Ksvt-av1: Preset 4 - Bosphorus 1080psvt-av1: Preset 12 - Bosphorus 4Ksvt-av1: Preset 13 - Bosphorus 4Ky-cruncher: 500Msvt-av1: Preset 8 - Bosphorus 1080psvt-av1: Preset 12 - Bosphorus 1080psvt-av1: Preset 13 - Bosphorus 1080pabc103.3163200006.026.04162500003.414.3814.679.3831.082432.1619695.375245.582616.8614.2735526425065281538102028770222336.6336.44213000004.9412202.163122.19131439.7734882.432635.6549885.823635.536733.728829.640546.4063688.567533.859829.52625.515.614163.9831508.89362.393322.641544.1242101.6471313.98727.3198136.46696.59151.49121.2115822.756.605151.2523150.7692211.5649149.3881213.597653.658.47793766.130166.9993476.836766.9655477.00145.3282187.45295.3237187.619916.656.69745.125.9221.9427.959.94667.3217.115192.633195.9044.897129.024506.246599.849103.2163600006.046.05162600003.414.7914.559.5930.993832.2534694.56345.635816.4414.2934551525141961528421929844556536.1337.28213300004.9419202.111322.17971441.0421882.882935.6944883.462235.606633.726229.642846.6414685.106433.792829.58415.4615.681963.7036512.169962.159422.703643.9996101.7295313.80127.3167136.52816.5866151.55951.2107823.4536.5703152.04150.1556212.3922148.779214.547153.618.51343749.031567.0028476.678466.9522477.28265.3168187.86945.328187.480216.726.7745.605.9821.8927.949.91568.42317.63193.581191.9274.914133.524505.536599.42103.19162933336.026.03161933333.4114.5514.679.6130.851632.4023694.926445.531516.5814.2735139425080821544242029833220336.6836.71213166675.0360198.369122.15501442.4037883.114835.5891882.972135.647433.802329.575946.5963685.737333.739229.63135.5015.595864.0546510.483862.144522.674344.0566101.7894313.75287.2910137.01576.5609152.14051.2141821.13176.4896153.9350150.4754211.9097149.2086213.872853.628.50263754.337266.9723477.056866.9459477.25315.3279187.46275.4112184.658116.636.72145.335.9421.9726.509.92967.77817.565193.425194.4714.918130.345503.508603.599OpenBenchmarking.org

TensorFlow

This is a benchmark of the TensorFlow deep learning framework using the TensorFlow reference benchmarks (tensorflow/benchmarks with tf_cnn_benchmarks.py). Note with the Phoronix Test Suite there is also pts/tensorflow-lite for benchmarking the TensorFlow Lite binaries if desired for complementary metrics. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgimages/sec, More Is BetterTensorFlow 2.12Device: CPU - Batch Size: 512 - Model: ResNet-50cba20406080100SE +/- 0.01, N = 3103.19103.20103.30

Quicksilver

Quicksilver is a proxy application that represents some elements of the Mercury workload by solving a simplified dynamic Monte Carlo particle transport problem. Quicksilver is developed by Lawrence Livermore National Laboratory (LLNL) and this test profile currently makes use of the OpenMP CPU threaded code path. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgFigure Of Merit, More Is BetterQuicksilver 20230818Input: CTS2cab4M8M12M16M20MSE +/- 8819.17, N = 31629333316320000163600001. (CXX) g++ options: -fopenmp -O3 -march=native

PyTorch

This is a benchmark of PyTorch making use of pytorch-benchmark [https://github.com/LukasHedegaard/pytorch-benchmark]. Currently this test profile is catered to CPU-based testing. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 16 - Model: Efficientnet_v2_lacb246810SE +/- 0.01, N = 36.026.026.04MIN: 5.57 / MAX: 6.13MIN: 5.54 / MAX: 6.14MIN: 5.63 / MAX: 6.16

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 512 - Model: Efficientnet_v2_lcab246810SE +/- 0.02, N = 36.036.046.05MIN: 5.42 / MAX: 6.21MIN: 5.57 / MAX: 6.16MIN: 5.62 / MAX: 6.16

Quicksilver

Quicksilver is a proxy application that represents some elements of the Mercury workload by solving a simplified dynamic Monte Carlo particle transport problem. Quicksilver is developed by Lawrence Livermore National Laboratory (LLNL) and this test profile currently makes use of the OpenMP CPU threaded code path. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgFigure Of Merit, More Is BetterQuicksilver 20230818Input: CORAL2 P2cab3M6M9M12M15MSE +/- 6666.67, N = 31619333316250000162600001. (CXX) g++ options: -fopenmp -O3 -march=native

Llama.cpp

Llama.cpp is a port of Facebook's LLaMA model in C/C++ developed by Georgi Gerganov. Llama.cpp allows the inference of LLaMA and other supported models in C/C++. For CPU inference Llama.cpp supports AVX2/AVX-512, ARM NEON, and other modern ISAs along with features like OpenBLAS usage. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgTokens Per Second, More Is BetterLlama.cpp b1808Model: llama-2-70b-chat.Q5_0.ggufabc0.76731.53462.30193.06923.8365SE +/- 0.01, N = 33.403.403.411. (CXX) g++ options: -std=c++11 -fPIC -O3 -pthread -march=native -mtune=native -lopenblas

PyTorch

This is a benchmark of PyTorch making use of pytorch-benchmark [https://github.com/LukasHedegaard/pytorch-benchmark]. Currently this test profile is catered to CPU-based testing. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 512 - Model: ResNet-152acb48121620SE +/- 0.06, N = 314.3814.5514.79MIN: 12.94 / MAX: 14.47MIN: 13.35 / MAX: 14.76MIN: 13.56 / MAX: 14.89

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 16 - Model: ResNet-152bac48121620SE +/- 0.08, N = 314.5514.6714.67MIN: 14.35 / MAX: 14.64MIN: 14.48 / MAX: 14.78MIN: 13.21 / MAX: 14.86

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 1 - Model: Efficientnet_v2_labc3691215SE +/- 0.04, N = 39.389.599.61MIN: 8.84 / MAX: 9.52MIN: 9.48 / MAX: 9.67MIN: 8.95 / MAX: 9.78

Neural Magic DeepSparse

This is a benchmark of Neural Magic's DeepSparse using its built-in deepsparse.benchmark utility and various models from their SparseZoo (https://sparsezoo.neuralmagic.com/). Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: BERT-Large, NLP Question Answering - Scenario: Synchronous Single-Streamabc714212835SE +/- 0.02, N = 331.0830.9930.85

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: BERT-Large, NLP Question Answering - Scenario: Synchronous Single-Streamabc816243240SE +/- 0.02, N = 332.1632.2532.40

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: BERT-Large, NLP Question Answering - Scenario: Asynchronous Multi-Streamacb150300450600750SE +/- 0.19, N = 3695.38694.93694.56

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: BERT-Large, NLP Question Answering - Scenario: Asynchronous Multi-Streamcab1020304050SE +/- 0.07, N = 345.5345.5845.64

PyTorch

This is a benchmark of PyTorch making use of pytorch-benchmark [https://github.com/LukasHedegaard/pytorch-benchmark]. Currently this test profile is catered to CPU-based testing. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 1 - Model: ResNet-152bca48121620SE +/- 0.05, N = 316.4416.5816.86MIN: 14.99 / MAX: 16.55MIN: 16.34 / MAX: 16.81MIN: 14.83 / MAX: 17.02

Llamafile

Mozilla's Llamafile allows distributing and running large language models (LLMs) as a single file. Llamafile aims to make open-source LLMs more accessible to developers and users. Llamafile supports a variety of models, CPUs and GPUs, and other options. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgTokens Per Second, More Is BetterLlamafile 0.6Test: mistral-7b-instruct-v0.2.Q8_0 - Acceleration: CPUacb48121620SE +/- 0.02, N = 314.2714.2714.29

Speedb

Speedb is a next-generation key value storage engine that is RocksDB compatible and aiming for stability, efficiency, and performance. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgOp/s, More Is BetterSpeedb 2.7Test: Update Randombca80K160K240K320K400KSE +/- 2346.60, N = 33455153513943552641. (CXX) g++ options: -O3 -march=native -pthread -fno-builtin-memcmp -fno-rtti -lpthread

OpenBenchmarking.orgOp/s, More Is BetterSpeedb 2.7Test: Read Random Write Randomacb500K1000K1500K2000K2500KSE +/- 12093.31, N = 32506528250808225141961. (CXX) g++ options: -O3 -march=native -pthread -fno-builtin-memcmp -fno-rtti -lpthread

OpenBenchmarking.orgOp/s, More Is BetterSpeedb 2.7Test: Read While Writingbac3M6M9M12M15MSE +/- 188799.32, N = 31528421915381020154424201. (CXX) g++ options: -O3 -march=native -pthread -fno-builtin-memcmp -fno-rtti -lpthread

OpenBenchmarking.orgOp/s, More Is BetterSpeedb 2.7Test: Random Readacb60M120M180M240M300MSE +/- 222389.57, N = 32877022232983322032984455651. (CXX) g++ options: -O3 -march=native -pthread -fno-builtin-memcmp -fno-rtti -lpthread

PyTorch

This is a benchmark of PyTorch making use of pytorch-benchmark [https://github.com/LukasHedegaard/pytorch-benchmark]. Currently this test profile is catered to CPU-based testing. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 16 - Model: ResNet-50bac816243240SE +/- 0.18, N = 336.1336.6336.68MIN: 34.96 / MAX: 36.54MIN: 35.5 / MAX: 37.04MIN: 30.63 / MAX: 37.51

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 512 - Model: ResNet-50acb918273645SE +/- 0.29, N = 336.4436.7137.28MIN: 35.51 / MAX: 36.77MIN: 35.24 / MAX: 37.75MIN: 35.91 / MAX: 37.74

Quicksilver

Quicksilver is a proxy application that represents some elements of the Mercury workload by solving a simplified dynamic Monte Carlo particle transport problem. Quicksilver is developed by Lawrence Livermore National Laboratory (LLNL) and this test profile currently makes use of the OpenMP CPU threaded code path. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgFigure Of Merit, More Is BetterQuicksilver 20230818Input: CORAL2 P1acb5M10M15M20M25MSE +/- 34801.02, N = 32130000021316667213300001. (CXX) g++ options: -fopenmp -O3 -march=native

Neural Magic DeepSparse

This is a benchmark of Neural Magic's DeepSparse using its built-in deepsparse.benchmark utility and various models from their SparseZoo (https://sparsezoo.neuralmagic.com/). Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Synchronous Single-Streamcba1.13312.26623.39934.53245.6655SE +/- 0.0306, N = 35.03604.94194.9412

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Synchronous Single-Streamcba4080120160200SE +/- 1.20, N = 3198.37202.11202.16

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Asynchronous Multi-Streamabc510152025SE +/- 0.01, N = 322.1922.1822.16

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Asynchronous Multi-Streamabc30060090012001500SE +/- 0.78, N = 31439.771441.041442.40

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Asynchronous Multi-Streamcba2004006008001000SE +/- 0.36, N = 3883.11882.88882.43

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Asynchronous Multi-Streamcab816243240SE +/- 0.02, N = 335.5935.6535.69

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Asynchronous Multi-Streamabc2004006008001000SE +/- 0.60, N = 3885.82883.46882.97

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Asynchronous Multi-Streamabc816243240SE +/- 0.06, N = 335.5435.6135.65

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Synchronous Single-Streamcab816243240SE +/- 0.03, N = 333.8033.7333.73

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Synchronous Single-Streamcab714212835SE +/- 0.02, N = 329.5829.6429.64

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Asynchronous Multi-Streambca1122334455SE +/- 0.10, N = 346.6446.6046.41

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Asynchronous Multi-Streambca150300450600750SE +/- 1.46, N = 3685.11685.74688.57

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Synchronous Single-Streamabc816243240SE +/- 0.03, N = 333.8633.7933.74

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Synchronous Single-Streamabc714212835SE +/- 0.02, N = 329.5329.5829.63

Llamafile

Mozilla's Llamafile allows distributing and running large language models (LLMs) as a single file. Llamafile aims to make open-source LLMs more accessible to developers and users. Llamafile supports a variety of models, CPUs and GPUs, and other options. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgTokens Per Second, More Is BetterLlamafile 0.6Test: wizardcoder-python-34b-v1.0.Q6_K - Acceleration: CPUbac1.23752.4753.71254.956.1875SE +/- 0.02, N = 35.465.505.50

Neural Magic DeepSparse

This is a benchmark of Neural Magic's DeepSparse using its built-in deepsparse.benchmark utility and various models from their SparseZoo (https://sparsezoo.neuralmagic.com/). Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Synchronous Single-Streambac48121620SE +/- 0.03, N = 315.6815.6115.60

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Synchronous Single-Streambac1428425670SE +/- 0.14, N = 363.7063.9864.05

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Asynchronous Multi-Streambca110220330440550SE +/- 1.07, N = 3512.17510.48508.89

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Asynchronous Multi-Streamcba1428425670SE +/- 0.16, N = 362.1462.1662.39

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Synchronous Single-Streambca510152025SE +/- 0.02, N = 322.7022.6722.64

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Synchronous Single-Streambca1020304050SE +/- 0.05, N = 344.0044.0644.12

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: NLP Text Classification, DistilBERT mnli - Scenario: Asynchronous Multi-Streamcba20406080100SE +/- 0.11, N = 3101.79101.73101.65

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: NLP Text Classification, DistilBERT mnli - Scenario: Asynchronous Multi-Streamcba70140210280350SE +/- 0.43, N = 3313.75313.80313.99

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: NLP Text Classification, DistilBERT mnli - Scenario: Synchronous Single-Streamabc246810SE +/- 0.0130, N = 37.31987.31677.2910

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: NLP Text Classification, DistilBERT mnli - Scenario: Synchronous Single-Streamabc306090120150SE +/- 0.24, N = 3136.47136.53137.02

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: CV Detection, YOLOv5s COCO - Scenario: Synchronous Single-Streamabc246810SE +/- 0.0014, N = 36.59006.58666.5609

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: CV Detection, YOLOv5s COCO - Scenario: Synchronous Single-Streamabc306090120150SE +/- 0.03, N = 3151.49151.56152.14

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: ResNet-50, Sparse INT8 - Scenario: Synchronous Single-Streamcab0.27320.54640.81961.09281.366SE +/- 0.0028, N = 31.21411.21151.2107

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: ResNet-50, Sparse INT8 - Scenario: Synchronous Single-Streamcab2004006008001000SE +/- 1.85, N = 3821.13822.75823.45

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Synchronous Single-Streamabc246810SE +/- 0.0019, N = 36.60506.57036.4896

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Synchronous Single-Streamabc306090120150SE +/- 0.04, N = 3151.25152.04153.94

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: CV Detection, YOLOv5s COCO - Scenario: Asynchronous Multi-Streamacb306090120150SE +/- 0.10, N = 3150.77150.48150.16

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: CV Detection, YOLOv5s COCO - Scenario: Asynchronous Multi-Streamacb50100150200250SE +/- 0.12, N = 3211.56211.91212.39

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Asynchronous Multi-Streamacb306090120150SE +/- 0.14, N = 3149.39149.21148.78

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Asynchronous Multi-Streamacb50100150200250SE +/- 0.26, N = 3213.60213.87214.55

TensorFlow

This is a benchmark of the TensorFlow deep learning framework using the TensorFlow reference benchmarks (tensorflow/benchmarks with tf_cnn_benchmarks.py). Note with the Phoronix Test Suite there is also pts/tensorflow-lite for benchmarking the TensorFlow Lite binaries if desired for complementary metrics. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgimages/sec, More Is BetterTensorFlow 2.12Device: CPU - Batch Size: 16 - Model: ResNet-50bca1224364860SE +/- 0.04, N = 353.6153.6253.65

Neural Magic DeepSparse

This is a benchmark of Neural Magic's DeepSparse using its built-in deepsparse.benchmark utility and various models from their SparseZoo (https://sparsezoo.neuralmagic.com/). Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: ResNet-50, Sparse INT8 - Scenario: Asynchronous Multi-Streambca246810SE +/- 0.0299, N = 38.51348.50268.4779

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: ResNet-50, Sparse INT8 - Scenario: Asynchronous Multi-Streambca8001600240032004000SE +/- 13.07, N = 33749.033754.343766.13

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: ResNet-50, Baseline - Scenario: Asynchronous Multi-Streambac1530456075SE +/- 0.04, N = 367.0067.0066.97

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: ResNet-50, Baseline - Scenario: Asynchronous Multi-Streambac100200300400500SE +/- 0.22, N = 3476.68476.84477.06

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: CV Classification, ResNet-50 ImageNet - Scenario: Asynchronous Multi-Streamabc1530456075SE +/- 0.02, N = 366.9766.9566.95

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: CV Classification, ResNet-50 ImageNet - Scenario: Asynchronous Multi-Streamacb100200300400500SE +/- 0.10, N = 3477.00477.25477.28

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: ResNet-50, Baseline - Scenario: Synchronous Single-Streamacb1.19882.39763.59644.79525.994SE +/- 0.0047, N = 35.32825.32795.3168

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: ResNet-50, Baseline - Scenario: Synchronous Single-Streamacb4080120160200SE +/- 0.17, N = 3187.45187.46187.87

OpenBenchmarking.orgms/batch, Fewer Is BetterNeural Magic DeepSparse 1.6Model: CV Classification, ResNet-50 ImageNet - Scenario: Synchronous Single-Streamcba1.21752.4353.65254.876.0875SE +/- 0.0772, N = 35.41125.32805.3237

OpenBenchmarking.orgitems/sec, More Is BetterNeural Magic DeepSparse 1.6Model: CV Classification, ResNet-50 ImageNet - Scenario: Synchronous Single-Streamcba4080120160200SE +/- 2.60, N = 3184.66187.48187.62

Llama.cpp

Llama.cpp is a port of Facebook's LLaMA model in C/C++ developed by Georgi Gerganov. Llama.cpp allows the inference of LLaMA and other supported models in C/C++. For CPU inference Llama.cpp supports AVX2/AVX-512, ARM NEON, and other modern ISAs along with features like OpenBLAS usage. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgTokens Per Second, More Is BetterLlama.cpp b1808Model: llama-2-13b.Q4_0.ggufcab48121620SE +/- 0.04, N = 316.6316.6516.721. (CXX) g++ options: -std=c++11 -fPIC -O3 -pthread -march=native -mtune=native -lopenblas

SVT-AV1

This is a benchmark of the SVT-AV1 open-source video encoder/decoder. SVT-AV1 was originally developed by Intel as part of their Open Visual Cloud / Scalable Video Technology (SVT). Development of SVT-AV1 has since moved to the Alliance for Open Media as part of upstream AV1 development. SVT-AV1 is a CPU-based multi-threaded video encoder for the AV1 video format with a sample YUV video file. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgFrames Per Second, More Is BetterSVT-AV1 1.8Encoder Mode: Preset 4 - Input: Bosphorus 4Kacb246810SE +/- 0.038, N = 36.6976.7216.7701. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq

PyTorch

This is a benchmark of PyTorch making use of pytorch-benchmark [https://github.com/LukasHedegaard/pytorch-benchmark]. Currently this test profile is catered to CPU-based testing. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgbatches/sec, More Is BetterPyTorch 2.1Device: CPU - Batch Size: 1 - Model: ResNet-50acb1020304050SE +/- 0.20, N = 345.1245.3345.60MIN: 35.13 / MAX: 45.99MIN: 43.73 / MAX: 46.27MIN: 34.79 / MAX: 46.32

TensorFlow

This is a benchmark of the TensorFlow deep learning framework using the TensorFlow reference benchmarks (tensorflow/benchmarks with tf_cnn_benchmarks.py). Note with the Phoronix Test Suite there is also pts/tensorflow-lite for benchmarking the TensorFlow Lite binaries if desired for complementary metrics. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgimages/sec, More Is BetterTensorFlow 2.12Device: CPU - Batch Size: 1 - Model: ResNet-50acb1.34552.6914.03655.3826.7275SE +/- 0.04, N = 35.925.945.98

Llamafile

Mozilla's Llamafile allows distributing and running large language models (LLMs) as a single file. Llamafile aims to make open-source LLMs more accessible to developers and users. Llamafile supports a variety of models, CPUs and GPUs, and other options. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgTokens Per Second, More Is BetterLlamafile 0.6Test: llava-v1.5-7b-q4 - Acceleration: CPUbac510152025SE +/- 0.03, N = 321.8921.9421.97

Llama.cpp

Llama.cpp is a port of Facebook's LLaMA model in C/C++ developed by Georgi Gerganov. Llama.cpp allows the inference of LLaMA and other supported models in C/C++. For CPU inference Llama.cpp supports AVX2/AVX-512, ARM NEON, and other modern ISAs along with features like OpenBLAS usage. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgTokens Per Second, More Is BetterLlama.cpp b1808Model: llama-2-7b.Q4_0.ggufcba714212835SE +/- 0.02, N = 326.5027.9427.951. (CXX) g++ options: -std=c++11 -fPIC -O3 -pthread -march=native -mtune=native -lopenblas

Y-Cruncher

Y-Cruncher is a multi-threaded Pi benchmark capable of computing Pi to trillions of digits. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgSeconds, Fewer Is BetterY-Cruncher 0.8.3Pi Digits To Calculate: 1Bacb3691215SE +/- 0.006, N = 39.9469.9299.915

SVT-AV1

This is a benchmark of the SVT-AV1 open-source video encoder/decoder. SVT-AV1 was originally developed by Intel as part of their Open Visual Cloud / Scalable Video Technology (SVT). Development of SVT-AV1 has since moved to the Alliance for Open Media as part of upstream AV1 development. SVT-AV1 is a CPU-based multi-threaded video encoder for the AV1 video format with a sample YUV video file. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgFrames Per Second, More Is BetterSVT-AV1 1.8Encoder Mode: Preset 8 - Input: Bosphorus 4Kacb1530456075SE +/- 0.17, N = 367.3267.7868.421. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq

OpenBenchmarking.orgFrames Per Second, More Is BetterSVT-AV1 1.8Encoder Mode: Preset 4 - Input: Bosphorus 1080pacb48121620SE +/- 0.16, N = 317.1217.5717.631. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq

OpenBenchmarking.orgFrames Per Second, More Is BetterSVT-AV1 1.8Encoder Mode: Preset 12 - Input: Bosphorus 4Kacb4080120160200SE +/- 2.19, N = 3192.63193.43193.581. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq

OpenBenchmarking.orgFrames Per Second, More Is BetterSVT-AV1 1.8Encoder Mode: Preset 13 - Input: Bosphorus 4Kbca4080120160200SE +/- 1.23, N = 3191.93194.47195.901. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq

Y-Cruncher

Y-Cruncher is a multi-threaded Pi benchmark capable of computing Pi to trillions of digits. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgSeconds, Fewer Is BetterY-Cruncher 0.8.3Pi Digits To Calculate: 500Mcba1.10662.21323.31984.42645.533SE +/- 0.007, N = 34.9184.9144.897

SVT-AV1

This is a benchmark of the SVT-AV1 open-source video encoder/decoder. SVT-AV1 was originally developed by Intel as part of their Open Visual Cloud / Scalable Video Technology (SVT). Development of SVT-AV1 has since moved to the Alliance for Open Media as part of upstream AV1 development. SVT-AV1 is a CPU-based multi-threaded video encoder for the AV1 video format with a sample YUV video file. Learn more via the OpenBenchmarking.org test page.

OpenBenchmarking.orgFrames Per Second, More Is BetterSVT-AV1 1.8Encoder Mode: Preset 8 - Input: Bosphorus 1080pacb306090120150SE +/- 1.85, N = 3129.02130.35133.521. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq

OpenBenchmarking.orgFrames Per Second, More Is BetterSVT-AV1 1.8Encoder Mode: Preset 12 - Input: Bosphorus 1080pcba110220330440550SE +/- 4.17, N = 3503.51505.54506.251. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq

OpenBenchmarking.orgFrames Per Second, More Is BetterSVT-AV1 1.8Encoder Mode: Preset 13 - Input: Bosphorus 1080pbac130260390520650SE +/- 1.86, N = 3599.42599.85603.601. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq

83 Results Shown

TensorFlow
Quicksilver
PyTorch:
  CPU - 16 - Efficientnet_v2_l
  CPU - 512 - Efficientnet_v2_l
Quicksilver
Llama.cpp
PyTorch:
  CPU - 512 - ResNet-152
  CPU - 16 - ResNet-152
  CPU - 1 - Efficientnet_v2_l
Neural Magic DeepSparse:
  BERT-Large, NLP Question Answering - Synchronous Single-Stream:
    ms/batch
    items/sec
  BERT-Large, NLP Question Answering - Asynchronous Multi-Stream:
    ms/batch
    items/sec
PyTorch
Llamafile
Speedb:
  Update Rand
  Read Rand Write Rand
  Read While Writing
  Rand Read
PyTorch:
  CPU - 16 - ResNet-50
  CPU - 512 - ResNet-50
Quicksilver
Neural Magic DeepSparse:
  NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Synchronous Single-Stream:
    ms/batch
    items/sec
  NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  NLP Document Classification, oBERT base uncased on IMDB - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  NLP Token Classification, BERT base uncased conll2003 - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  NLP Document Classification, oBERT base uncased on IMDB - Synchronous Single-Stream:
    ms/batch
    items/sec
  BERT-Large, NLP Question Answering, Sparse INT8 - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  NLP Token Classification, BERT base uncased conll2003 - Synchronous Single-Stream:
    ms/batch
    items/sec
Llamafile
Neural Magic DeepSparse:
  BERT-Large, NLP Question Answering, Sparse INT8 - Synchronous Single-Stream:
    ms/batch
    items/sec
  CV Segmentation, 90% Pruned YOLACT Pruned - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  CV Segmentation, 90% Pruned YOLACT Pruned - Synchronous Single-Stream:
    ms/batch
    items/sec
  NLP Text Classification, DistilBERT mnli - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  NLP Text Classification, DistilBERT mnli - Synchronous Single-Stream:
    ms/batch
    items/sec
  CV Detection, YOLOv5s COCO - Synchronous Single-Stream:
    ms/batch
    items/sec
  ResNet-50, Sparse INT8 - Synchronous Single-Stream:
    ms/batch
    items/sec
  CV Detection, YOLOv5s COCO, Sparse INT8 - Synchronous Single-Stream:
    ms/batch
    items/sec
  CV Detection, YOLOv5s COCO - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  CV Detection, YOLOv5s COCO, Sparse INT8 - Asynchronous Multi-Stream:
    ms/batch
    items/sec
TensorFlow
Neural Magic DeepSparse:
  ResNet-50, Sparse INT8 - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  ResNet-50, Baseline - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  CV Classification, ResNet-50 ImageNet - Asynchronous Multi-Stream:
    ms/batch
    items/sec
  ResNet-50, Baseline - Synchronous Single-Stream:
    ms/batch
    items/sec
  CV Classification, ResNet-50 ImageNet - Synchronous Single-Stream:
    ms/batch
    items/sec
Llama.cpp
SVT-AV1
PyTorch
TensorFlow
Llamafile
Llama.cpp
Y-Cruncher
SVT-AV1:
  Preset 8 - Bosphorus 4K
  Preset 4 - Bosphorus 1080p
  Preset 12 - Bosphorus 4K
  Preset 13 - Bosphorus 4K
Y-Cruncher
SVT-AV1:
  Preset 8 - Bosphorus 1080p
  Preset 12 - Bosphorus 1080p
  Preset 13 - Bosphorus 1080p