svt deepsparse AMD Ryzen Threadripper 7980X 64-Cores testing with a System76 Thelio Major (FA Z5 BIOS) and AMD Radeon Pro W7900 45GB on Ubuntu 23.10 via the Phoronix Test Suite.
HTML result view exported from: https://openbenchmarking.org/result/2403159-PTS-SVTDEEPS65&grs&sor .
svt deepsparse Processor Motherboard Chipset Memory Disk Graphics Audio Monitor Network OS Kernel Desktop Display Server OpenGL Compiler File-System Screen Resolution a b c d AMD Ryzen Threadripper 7980X 64-Cores @ 7.79GHz (64 Cores / 128 Threads) System76 Thelio Major (FA Z5 BIOS) AMD Device 14a4 4 x 32GB DRAM-4800MT/s Micron MTC20F1045S1RC48BA2 1000GB CT1000T700SSD5 AMD Radeon Pro W7900 45GB (1760/1124MHz) AMD Device 14cc DELL P2415Q Aquantia AQC113C NBase-T/IEEE + Realtek RTL8125 2.5GbE + Intel Wi-Fi 6 AX210/AX211/AX411 Ubuntu 23.10 6.5.0-25-generic (x86_64) GNOME Shell 45.2 X Server + Wayland 4.6 Mesa 23.2.1-1ubuntu3.1 (LLVM 15.0.7 DRM 3.54) GCC 13.2.0 ext4 1920x1080 OpenBenchmarking.org Kernel Details - Transparent Huge Pages: madvise Compiler Details - --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-XYspKM/gcc-13-13.2.0/debian/tmp-nvptx/usr,amdgcn-amdhsa=/build/gcc-13-XYspKM/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 Processor Details - Scaling Governor: amd-pstate-epp powersave (EPP: balance_performance) - CPU Microcode: 0xa108105 Python Details - Python 3.11.6 Security Details - 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
svt deepsparse svt-av1: Preset 12 - Bosphorus 4K svt-av1: Preset 8 - Bosphorus 1080p svt-av1: Preset 13 - Bosphorus 4K svt-av1: Preset 13 - Bosphorus 1080p svt-av1: Preset 12 - Bosphorus 1080p deepsparse: ResNet-50, Sparse INT8 - Asynchronous Multi-Stream deepsparse: ResNet-50, Sparse INT8 - Asynchronous Multi-Stream deepsparse: NLP Text Classification, DistilBERT mnli - Synchronous Single-Stream deepsparse: NLP Text Classification, DistilBERT mnli - Synchronous Single-Stream svt-av1: Preset 4 - Bosphorus 1080p svt-av1: Preset 4 - Bosphorus 4K deepsparse: Llama2 Chat 7b Quantized - Synchronous Single-Stream deepsparse: Llama2 Chat 7b Quantized - Synchronous Single-Stream svt-av1: Preset 8 - Bosphorus 4K deepsparse: BERT-Large, NLP Question Answering, Sparse INT8 - Synchronous Single-Stream deepsparse: BERT-Large, NLP Question Answering, Sparse INT8 - Synchronous Single-Stream deepsparse: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Synchronous Single-Stream deepsparse: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Synchronous Single-Stream deepsparse: ResNet-50, Baseline - Synchronous Single-Stream deepsparse: ResNet-50, Baseline - Synchronous Single-Stream deepsparse: CV Detection, YOLOv5s COCO, Sparse INT8 - Asynchronous Multi-Stream deepsparse: CV Detection, YOLOv5s COCO, Sparse INT8 - Asynchronous Multi-Stream deepsparse: NLP Document Classification, oBERT base uncased on IMDB - Asynchronous Multi-Stream deepsparse: Llama2 Chat 7b Quantized - Asynchronous Multi-Stream deepsparse: Llama2 Chat 7b Quantized - Asynchronous Multi-Stream deepsparse: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Asynchronous Multi-Stream deepsparse: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Asynchronous Multi-Stream deepsparse: CV Detection, YOLOv5s COCO, Sparse INT8 - Synchronous Single-Stream deepsparse: CV Detection, YOLOv5s COCO, Sparse INT8 - Synchronous Single-Stream deepsparse: CV Segmentation, 90% Pruned YOLACT Pruned - Asynchronous Multi-Stream deepsparse: NLP Document Classification, oBERT base uncased on IMDB - Asynchronous Multi-Stream deepsparse: BERT-Large, NLP Question Answering, Sparse INT8 - Asynchronous Multi-Stream deepsparse: BERT-Large, NLP Question Answering, Sparse INT8 - Asynchronous Multi-Stream deepsparse: ResNet-50, Sparse INT8 - Synchronous Single-Stream deepsparse: ResNet-50, Sparse INT8 - Synchronous Single-Stream deepsparse: NLP Token Classification, BERT base uncased conll2003 - Asynchronous Multi-Stream deepsparse: CV Classification, ResNet-50 ImageNet - Synchronous Single-Stream deepsparse: CV Segmentation, 90% Pruned YOLACT Pruned - Asynchronous Multi-Stream deepsparse: CV Classification, ResNet-50 ImageNet - Synchronous Single-Stream deepsparse: NLP Token Classification, BERT base uncased conll2003 - Asynchronous Multi-Stream deepsparse: NLP Token Classification, BERT base uncased conll2003 - Synchronous Single-Stream deepsparse: NLP Token Classification, BERT base uncased conll2003 - Synchronous Single-Stream deepsparse: NLP Text Classification, DistilBERT mnli - Asynchronous Multi-Stream deepsparse: NLP Text Classification, DistilBERT mnli - Asynchronous Multi-Stream deepsparse: CV Segmentation, 90% Pruned YOLACT Pruned - Synchronous Single-Stream deepsparse: CV Segmentation, 90% Pruned YOLACT Pruned - Synchronous Single-Stream deepsparse: CV Classification, ResNet-50 ImageNet - Asynchronous Multi-Stream deepsparse: NLP Document Classification, oBERT base uncased on IMDB - Synchronous Single-Stream deepsparse: NLP Document Classification, oBERT base uncased on IMDB - Synchronous Single-Stream deepsparse: CV Classification, ResNet-50 ImageNet - Asynchronous Multi-Stream deepsparse: ResNet-50, Baseline - Asynchronous Multi-Stream deepsparse: ResNet-50, Baseline - Asynchronous Multi-Stream a b c d 188.558 193.759 192.567 662.493 616.583 3.5332 2253.8602 8.8070 113.4701 27.476 9.892 11.1249 89.8704 94.509 89.7091 11.1350 252.3577 3.9594 170.3090 5.8648 61.7698 129.4408 365.3355 3.4391 2252.4627 887.4442 8.9970 11.4406 87.3635 39.5729 21.8474 396.5768 20.1516 1134.7704 0.8788 363.5336 5.8423 202.0153 170.9594 21.9224 58.0689 17.2184 188.3255 42.4509 29.7744 33.5681 29.1108 58.0191 17.2332 274.5853 29.1453 274.2576 195.907 194.769 193.286 653.688 615.540 3.5087 2269.5068 8.7347 114.4072 27.282 9.793 11.0159 90.7521 93.734 90.4983 11.0383 251.6102 3.9707 171.4934 5.8243 61.9978 128.9560 365.9832 3.4406 2251.6125 886.1847 9.0094 11.4159 87.5496 39.4685 21.8225 396.4446 20.1585 1131.6373 0.8810 364.4283 5.8302 202.5056 171.3099 21.9027 58.1442 17.1960 188.1505 42.4853 29.8179 33.5194 29.1324 58.0188 17.2332 274.3872 29.1215 274.4692 194.065 189.608 192.638 671.832 625.817 3.5324 2254.2228 8.7752 113.8761 27.518 9.889 11.0814 90.2253 93.930 89.8053 11.1223 251.2477 3.9766 171.4781 5.8245 61.9371 129.1012 365.5771 3.4298 2258.1407 884.6567 9.0252 11.4055 87.6317 39.4664 21.8606 397.5870 20.1007 1133.4956 0.8796 363.7623 5.8312 202.3677 171.2808 21.9041 58.1916 17.1825 188.2128 42.4756 29.7891 33.5525 29.1435 58.0207 17.2329 274.2789 29.1230 274.4696 195.858 195.612 187.625 670.138 621.442 3.5604 2236.9169 8.7026 114.8252 27.241 9.866 11.0145 90.7667 94.559 90.3875 11.0513 253.0264 3.9490 171.3561 5.8285 61.7579 129.4552 366.6823 3.4284 2259.2648 886.7758 9.0040 11.4067 87.6213 39.4525 21.7950 397.5652 20.1014 1132.3151 0.8806 363.5857 5.8281 202.4630 171.3647 21.9527 58.1383 17.1981 188.4415 42.4217 29.7882 33.5532 29.1083 58.0879 17.2130 274.5981 29.1232 274.4604 OpenBenchmarking.org
SVT-AV1 Encoder Mode: Preset 12 - Input: Bosphorus 4K OpenBenchmarking.org Frames Per Second, More Is Better SVT-AV1 2.0 Encoder Mode: Preset 12 - Input: Bosphorus 4K b d c a 40 80 120 160 200 SE +/- 2.54, N = 3 SE +/- 2.35, N = 3 SE +/- 1.67, N = 3 SE +/- 2.09, N = 5 195.91 195.86 194.07 188.56 1. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq
SVT-AV1 Encoder Mode: Preset 8 - Input: Bosphorus 1080p OpenBenchmarking.org Frames Per Second, More Is Better SVT-AV1 2.0 Encoder Mode: Preset 8 - Input: Bosphorus 1080p d b a c 40 80 120 160 200 SE +/- 1.92, N = 3 SE +/- 1.86, N = 3 SE +/- 2.31, N = 3 SE +/- 0.78, N = 3 195.61 194.77 193.76 189.61 1. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq
SVT-AV1 Encoder Mode: Preset 13 - Input: Bosphorus 4K OpenBenchmarking.org Frames Per Second, More Is Better SVT-AV1 2.0 Encoder Mode: Preset 13 - Input: Bosphorus 4K b c a d 40 80 120 160 200 SE +/- 2.02, N = 15 SE +/- 2.22, N = 3 SE +/- 2.20, N = 3 SE +/- 1.36, N = 3 193.29 192.64 192.57 187.63 1. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq
SVT-AV1 Encoder Mode: Preset 13 - Input: Bosphorus 1080p OpenBenchmarking.org Frames Per Second, More Is Better SVT-AV1 2.0 Encoder Mode: Preset 13 - Input: Bosphorus 1080p c d a b 140 280 420 560 700 SE +/- 5.18, N = 15 SE +/- 6.00, N = 15 SE +/- 3.32, N = 3 SE +/- 1.76, N = 3 671.83 670.14 662.49 653.69 1. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq
SVT-AV1 Encoder Mode: Preset 12 - Input: Bosphorus 1080p OpenBenchmarking.org Frames Per Second, More Is Better SVT-AV1 2.0 Encoder Mode: Preset 12 - Input: Bosphorus 1080p c d a b 140 280 420 560 700 SE +/- 4.75, N = 3 SE +/- 4.99, N = 3 SE +/- 8.04, N = 3 SE +/- 5.36, N = 3 625.82 621.44 616.58 615.54 1. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq
Neural Magic DeepSparse Model: ResNet-50, Sparse INT8 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: ResNet-50, Sparse INT8 - Scenario: Asynchronous Multi-Stream b c a d 0.8011 1.6022 2.4033 3.2044 4.0055 SE +/- 0.0132, N = 3 SE +/- 0.0151, N = 3 SE +/- 0.0104, N = 3 SE +/- 0.0258, N = 3 3.5087 3.5324 3.5332 3.5604
Neural Magic DeepSparse Model: ResNet-50, Sparse INT8 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: ResNet-50, Sparse INT8 - Scenario: Asynchronous Multi-Stream b c a d 500 1000 1500 2000 2500 SE +/- 8.66, N = 3 SE +/- 9.51, N = 3 SE +/- 6.75, N = 3 SE +/- 15.99, N = 3 2269.51 2254.22 2253.86 2236.92
Neural Magic DeepSparse Model: NLP Text Classification, DistilBERT mnli - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: NLP Text Classification, DistilBERT mnli - Scenario: Synchronous Single-Stream d b c a 2 4 6 8 10 SE +/- 0.0260, N = 3 SE +/- 0.0397, N = 3 SE +/- 0.0131, N = 3 SE +/- 0.0288, N = 3 8.7026 8.7347 8.7752 8.8070
Neural Magic DeepSparse Model: NLP Text Classification, DistilBERT mnli - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: NLP Text Classification, DistilBERT mnli - Scenario: Synchronous Single-Stream d b c a 30 60 90 120 150 SE +/- 0.34, N = 3 SE +/- 0.52, N = 3 SE +/- 0.17, N = 3 SE +/- 0.37, N = 3 114.83 114.41 113.88 113.47
SVT-AV1 Encoder Mode: Preset 4 - Input: Bosphorus 1080p OpenBenchmarking.org Frames Per Second, More Is Better SVT-AV1 2.0 Encoder Mode: Preset 4 - Input: Bosphorus 1080p c a b d 6 12 18 24 30 SE +/- 0.16, N = 3 SE +/- 0.22, N = 3 SE +/- 0.10, N = 3 SE +/- 0.12, N = 3 27.52 27.48 27.28 27.24 1. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq
SVT-AV1 Encoder Mode: Preset 4 - Input: Bosphorus 4K OpenBenchmarking.org Frames Per Second, More Is Better SVT-AV1 2.0 Encoder Mode: Preset 4 - Input: Bosphorus 4K a c d b 3 6 9 12 15 SE +/- 0.009, N = 3 SE +/- 0.028, N = 3 SE +/- 0.034, N = 3 SE +/- 0.014, N = 3 9.892 9.889 9.866 9.793 1. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq
Neural Magic DeepSparse Model: Llama2 Chat 7b Quantized - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: Llama2 Chat 7b Quantized - Scenario: Synchronous Single-Stream a c b d 3 6 9 12 15 SE +/- 0.07, N = 3 SE +/- 0.06, N = 3 SE +/- 0.00, N = 3 SE +/- 0.00, N = 3 11.12 11.08 11.02 11.01
Neural Magic DeepSparse Model: Llama2 Chat 7b Quantized - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: Llama2 Chat 7b Quantized - Scenario: Synchronous Single-Stream a c b d 20 40 60 80 100 SE +/- 0.58, N = 3 SE +/- 0.49, N = 3 SE +/- 0.01, N = 3 SE +/- 0.02, N = 3 89.87 90.23 90.75 90.77
SVT-AV1 Encoder Mode: Preset 8 - Input: Bosphorus 4K OpenBenchmarking.org Frames Per Second, More Is Better SVT-AV1 2.0 Encoder Mode: Preset 8 - Input: Bosphorus 4K d a c b 20 40 60 80 100 SE +/- 0.60, N = 3 SE +/- 0.33, N = 3 SE +/- 0.20, N = 3 SE +/- 0.31, N = 3 94.56 94.51 93.93 93.73 1. (CXX) g++ options: -march=native -mno-avx -mavx2 -mavx512f -mavx512bw -mavx512dq
Neural Magic DeepSparse Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Synchronous Single-Stream b d c a 20 40 60 80 100 SE +/- 0.29, N = 3 SE +/- 0.05, N = 3 SE +/- 0.14, N = 3 SE +/- 0.23, N = 3 90.50 90.39 89.81 89.71
Neural Magic DeepSparse Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Synchronous Single-Stream b d c a 3 6 9 12 15 SE +/- 0.04, N = 3 SE +/- 0.01, N = 3 SE +/- 0.02, N = 3 SE +/- 0.03, N = 3 11.04 11.05 11.12 11.14
Neural Magic DeepSparse Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Synchronous Single-Stream d a b c 60 120 180 240 300 SE +/- 1.88, N = 3 SE +/- 1.31, N = 3 SE +/- 0.19, N = 3 SE +/- 0.20, N = 3 253.03 252.36 251.61 251.25
Neural Magic DeepSparse Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Synchronous Single-Stream d a b c 0.8947 1.7894 2.6841 3.5788 4.4735 SE +/- 0.0293, N = 3 SE +/- 0.0204, N = 3 SE +/- 0.0031, N = 3 SE +/- 0.0030, N = 3 3.9490 3.9594 3.9707 3.9766
Neural Magic DeepSparse Model: ResNet-50, Baseline - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: ResNet-50, Baseline - Scenario: Synchronous Single-Stream b c d a 40 80 120 160 200 SE +/- 0.01, N = 3 SE +/- 0.16, N = 3 SE +/- 0.19, N = 3 SE +/- 0.32, N = 3 171.49 171.48 171.36 170.31
Neural Magic DeepSparse Model: ResNet-50, Baseline - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: ResNet-50, Baseline - Scenario: Synchronous Single-Stream b c d a 1.3196 2.6392 3.9588 5.2784 6.598 SE +/- 0.0004, N = 3 SE +/- 0.0053, N = 3 SE +/- 0.0066, N = 3 SE +/- 0.0111, N = 3 5.8243 5.8245 5.8285 5.8648
Neural Magic DeepSparse Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Asynchronous Multi-Stream d a c b 14 28 42 56 70 SE +/- 0.07, N = 3 SE +/- 0.09, N = 3 SE +/- 0.21, N = 3 SE +/- 0.07, N = 3 61.76 61.77 61.94 62.00
Neural Magic DeepSparse Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Asynchronous Multi-Stream d a c b 30 60 90 120 150 SE +/- 0.18, N = 3 SE +/- 0.18, N = 3 SE +/- 0.42, N = 3 SE +/- 0.15, N = 3 129.46 129.44 129.10 128.96
Neural Magic DeepSparse Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Asynchronous Multi-Stream a c b d 80 160 240 320 400 SE +/- 0.99, N = 3 SE +/- 0.63, N = 3 SE +/- 0.38, N = 3 SE +/- 0.31, N = 3 365.34 365.58 365.98 366.68
Neural Magic DeepSparse Model: Llama2 Chat 7b Quantized - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: Llama2 Chat 7b Quantized - Scenario: Asynchronous Multi-Stream b a c d 0.7741 1.5482 2.3223 3.0964 3.8705 SE +/- 0.0006, N = 3 SE +/- 0.0053, N = 3 SE +/- 0.0031, N = 3 SE +/- 0.0032, N = 3 3.4406 3.4391 3.4298 3.4284
Neural Magic DeepSparse Model: Llama2 Chat 7b Quantized - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: Llama2 Chat 7b Quantized - Scenario: Asynchronous Multi-Stream b a c d 500 1000 1500 2000 2500 SE +/- 0.34, N = 3 SE +/- 3.18, N = 3 SE +/- 1.96, N = 3 SE +/- 1.96, N = 3 2251.61 2252.46 2258.14 2259.26
Neural Magic DeepSparse Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Asynchronous Multi-Stream a d b c 200 400 600 800 1000 SE +/- 0.68, N = 3 SE +/- 0.73, N = 3 SE +/- 0.69, N = 3 SE +/- 0.15, N = 3 887.44 886.78 886.18 884.66
Neural Magic DeepSparse Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: NLP Text Classification, BERT base uncased SST2, Sparse INT8 - Scenario: Asynchronous Multi-Stream a d b c 3 6 9 12 15 SE +/- 0.0068, N = 3 SE +/- 0.0073, N = 3 SE +/- 0.0071, N = 3 SE +/- 0.0015, N = 3 8.9970 9.0040 9.0094 9.0252
Neural Magic DeepSparse Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Synchronous Single-Stream c d b a 3 6 9 12 15 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 SE +/- 0.02, N = 3 11.41 11.41 11.42 11.44
Neural Magic DeepSparse Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: CV Detection, YOLOv5s COCO, Sparse INT8 - Scenario: Synchronous Single-Stream c d b a 20 40 60 80 100 SE +/- 0.06, N = 3 SE +/- 0.08, N = 3 SE +/- 0.07, N = 3 SE +/- 0.16, N = 3 87.63 87.62 87.55 87.36
Neural Magic DeepSparse Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Asynchronous Multi-Stream a b c d 9 18 27 36 45 SE +/- 0.02, N = 3 SE +/- 0.04, N = 3 SE +/- 0.03, N = 3 SE +/- 0.04, N = 3 39.57 39.47 39.47 39.45
Neural Magic DeepSparse Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Asynchronous Multi-Stream c a b d 5 10 15 20 25 SE +/- 0.04, N = 3 SE +/- 0.08, N = 3 SE +/- 0.02, N = 3 SE +/- 0.03, N = 3 21.86 21.85 21.82 21.80
Neural Magic DeepSparse Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Asynchronous Multi-Stream c d a b 90 180 270 360 450 SE +/- 0.37, N = 3 SE +/- 0.27, N = 3 SE +/- 0.23, N = 3 SE +/- 0.87, N = 3 397.59 397.57 396.58 396.44
Neural Magic DeepSparse Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: BERT-Large, NLP Question Answering, Sparse INT8 - Scenario: Asynchronous Multi-Stream c d a b 5 10 15 20 25 SE +/- 0.02, N = 3 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 SE +/- 0.04, N = 3 20.10 20.10 20.15 20.16
Neural Magic DeepSparse Model: ResNet-50, Sparse INT8 - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: ResNet-50, Sparse INT8 - Scenario: Synchronous Single-Stream a c d b 200 400 600 800 1000 SE +/- 2.80, N = 3 SE +/- 3.06, N = 3 SE +/- 5.41, N = 3 SE +/- 7.00, N = 3 1134.77 1133.50 1132.32 1131.64
Neural Magic DeepSparse Model: ResNet-50, Sparse INT8 - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: ResNet-50, Sparse INT8 - Scenario: Synchronous Single-Stream a c d b 0.1982 0.3964 0.5946 0.7928 0.991 SE +/- 0.0021, N = 3 SE +/- 0.0025, N = 3 SE +/- 0.0042, N = 3 SE +/- 0.0054, N = 3 0.8788 0.8796 0.8806 0.8810
Neural Magic DeepSparse Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Asynchronous Multi-Stream a d c b 80 160 240 320 400 SE +/- 0.17, N = 3 SE +/- 1.03, N = 3 SE +/- 0.33, N = 3 SE +/- 0.43, N = 3 363.53 363.59 363.76 364.43
Neural Magic DeepSparse Model: CV Classification, ResNet-50 ImageNet - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: CV Classification, ResNet-50 ImageNet - Scenario: Synchronous Single-Stream d b c a 1.3145 2.629 3.9435 5.258 6.5725 SE +/- 0.0043, N = 3 SE +/- 0.0134, N = 3 SE +/- 0.0039, N = 3 SE +/- 0.0009, N = 3 5.8281 5.8302 5.8312 5.8423
Neural Magic DeepSparse Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Asynchronous Multi-Stream a c d b 40 80 120 160 200 SE +/- 0.12, N = 3 SE +/- 0.08, N = 3 SE +/- 0.10, N = 3 SE +/- 0.18, N = 3 202.02 202.37 202.46 202.51
Neural Magic DeepSparse Model: CV Classification, ResNet-50 ImageNet - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: CV Classification, ResNet-50 ImageNet - Scenario: Synchronous Single-Stream d b c a 40 80 120 160 200 SE +/- 0.12, N = 3 SE +/- 0.40, N = 3 SE +/- 0.12, N = 3 SE +/- 0.03, N = 3 171.36 171.31 171.28 170.96
Neural Magic DeepSparse Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Asynchronous Multi-Stream d a c b 5 10 15 20 25 SE +/- 0.04, N = 3 SE +/- 0.02, N = 3 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 21.95 21.92 21.90 21.90
Neural Magic DeepSparse Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Synchronous Single-Stream a d b c 13 26 39 52 65 SE +/- 0.06, N = 3 SE +/- 0.04, N = 3 SE +/- 0.03, N = 3 SE +/- 0.05, N = 3 58.07 58.14 58.14 58.19
Neural Magic DeepSparse Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: NLP Token Classification, BERT base uncased conll2003 - Scenario: Synchronous Single-Stream a d b c 4 8 12 16 20 SE +/- 0.02, N = 3 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 17.22 17.20 17.20 17.18
Neural Magic DeepSparse Model: NLP Text Classification, DistilBERT mnli - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: NLP Text Classification, DistilBERT mnli - Scenario: Asynchronous Multi-Stream d a c b 40 80 120 160 200 SE +/- 0.00, N = 3 SE +/- 0.07, N = 3 SE +/- 0.05, N = 3 SE +/- 0.02, N = 3 188.44 188.33 188.21 188.15
Neural Magic DeepSparse Model: NLP Text Classification, DistilBERT mnli - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: NLP Text Classification, DistilBERT mnli - Scenario: Asynchronous Multi-Stream d a c b 10 20 30 40 50 SE +/- 0.00, N = 3 SE +/- 0.02, N = 3 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 42.42 42.45 42.48 42.49
Neural Magic DeepSparse Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Synchronous Single-Stream b c d a 7 14 21 28 35 SE +/- 0.01, N = 3 SE +/- 0.00, N = 3 SE +/- 0.02, N = 3 SE +/- 0.02, N = 3 29.82 29.79 29.79 29.77
Neural Magic DeepSparse Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: CV Segmentation, 90% Pruned YOLACT Pruned - Scenario: Synchronous Single-Stream b c d a 8 16 24 32 40 SE +/- 0.01, N = 3 SE +/- 0.00, N = 3 SE +/- 0.03, N = 3 SE +/- 0.02, N = 3 33.52 33.55 33.55 33.57
Neural Magic DeepSparse Model: CV Classification, ResNet-50 ImageNet - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: CV Classification, ResNet-50 ImageNet - Scenario: Asynchronous Multi-Stream d a b c 7 14 21 28 35 SE +/- 0.04, N = 3 SE +/- 0.03, N = 3 SE +/- 0.00, N = 3 SE +/- 0.01, N = 3 29.11 29.11 29.13 29.14
Neural Magic DeepSparse Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Synchronous Single-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Synchronous Single-Stream b a c d 13 26 39 52 65 SE +/- 0.05, N = 3 SE +/- 0.03, N = 3 SE +/- 0.03, N = 3 SE +/- 0.07, N = 3 58.02 58.02 58.02 58.09
Neural Magic DeepSparse Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Synchronous Single-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: NLP Document Classification, oBERT base uncased on IMDB - Scenario: Synchronous Single-Stream b a c d 4 8 12 16 20 SE +/- 0.02, N = 3 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 SE +/- 0.02, N = 3 17.23 17.23 17.23 17.21
Neural Magic DeepSparse Model: CV Classification, ResNet-50 ImageNet - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: CV Classification, ResNet-50 ImageNet - Scenario: Asynchronous Multi-Stream d a b c 60 120 180 240 300 SE +/- 0.36, N = 3 SE +/- 0.26, N = 3 SE +/- 0.04, N = 3 SE +/- 0.12, N = 3 274.60 274.59 274.39 274.28
Neural Magic DeepSparse Model: ResNet-50, Baseline - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org ms/batch, Fewer Is Better Neural Magic DeepSparse 1.7 Model: ResNet-50, Baseline - Scenario: Asynchronous Multi-Stream b c d a 7 14 21 28 35 SE +/- 0.02, N = 3 SE +/- 0.01, N = 3 SE +/- 0.02, N = 3 SE +/- 0.00, N = 3 29.12 29.12 29.12 29.15
Neural Magic DeepSparse Model: ResNet-50, Baseline - Scenario: Asynchronous Multi-Stream OpenBenchmarking.org items/sec, More Is Better Neural Magic DeepSparse 1.7 Model: ResNet-50, Baseline - Scenario: Asynchronous Multi-Stream c b d a 60 120 180 240 300 SE +/- 0.11, N = 3 SE +/- 0.15, N = 3 SE +/- 0.20, N = 3 SE +/- 0.04, N = 3 274.47 274.47 274.46 274.26
Phoronix Test Suite v10.8.5