From 82ec8006cc8dad073c88d3f9b9149ac64b7a8647 Mon Sep 17 00:00:00 2001 From: Albert Reuther Date: Wed, 7 Feb 2024 09:50:54 -0500 Subject: [PATCH] 2023 additions --- README.md | 8 +- peak_2022.md | 4 +- peak_2023.md | 156 +++++++++++++++++++++++++++ peak_accelerators_ieee_hpec_2023.csv | 136 +++++++++++++++++++++++ 4 files changed, 300 insertions(+), 4 deletions(-) create mode 100644 peak_2023.md create mode 100644 peak_accelerators_ieee_hpec_2023.csv diff --git a/README.md b/README.md index efab9fd..8bab5ac 100755 --- a/README.md +++ b/README.md @@ -7,7 +7,11 @@ A few colleagues and I at MIT Lincoln Laboratory Supercomputing Center (LLSC) ha ## Papers and Datasets -So far we have published three papers at the IEEE-HPEC Conference and a fourth paper has been accepted at IEEE-HPEC 2022. Each of the papers are available in IEEE Xplore and arXiv.org. The datasets that were compiled for these papers are available here on subpages, and more fields are available as CSV files. +So far we have published five papers at the IEEE-HPEC Conference. Each of the papers are available in IEEE Xplore and arXiv.org. The datasets that were compiled for these papers are available here on subpages, and more fields are available as CSV files. + +### 2023: +A. Reuther, P. Michaleas, M. Jones, V. Gadepally, S. Samsi and J. Kepner, "Lincoln AI Computing Survey (LAICS) Update," *2023 IEEE High Performance Extreme Computing Conference (HPEC)*, 2023, pp. 1-7, \[[IEEE Xplore doi: 10.1109/HPEC58863.2023.10363568](https://doi.org/10.1109/HPEC58863.2023.10363568)\] \[[ArXiv.org/abs/2310.09145]( +https://doi.org/10.48550/arXiv.2310.09145)\] \[[data](peak_2023.md)\]. ### 2022: A. Reuther, P. Michaleas, M. Jones, V. Gadepally, S. Samsi and J. Kepner, "AI and ML Accelerator Survey and Trends," *2022 IEEE High Performance Extreme Computing Conference (HPEC)*, 2022, pp. 1-10, \[[IEEE Xplore doi: 10.1109/HPEC55821.2022.9926331](https://doi.org/10.1109/HPEC55821.2022.9926331)\] \[[ArXiv.org/abs/2210.04055](https://arxiv.org/abs/2210.04055)\] \[[data](peak_2022.md)\]. @@ -25,4 +29,4 @@ Please acknowledge this work with one or more of the papers above. -Copyright 2022 MIT, Albert I. Reuther +Copyright 2023 MIT, Albert I. Reuther diff --git a/peak_2022.md b/peak_2022.md index 472688b..bbc02bc 100644 --- a/peak_2022.md +++ b/peak_2022.md @@ -125,6 +125,8 @@ For the full dataset in CSV format, please download \[[peak-accelerators-ieee-hp | Texas Instruments | TDA4VM | TexInst | 8000 | 20 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/tis-first-automotive-soc-with-an-ai-accelerator-launches/)\] \[[www.ti.com](https://www.ti.com/lit/gpn/tda4vm)\] \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/newsletter_detail.php?num=6130&year=2020&tag=3)\] | | Toshiba | 2015 | Toshiba | 20000 | 10 | int8 | System | \[[www.eetimes.com](https://www.eetimes.com/samsung-toshiba-detail-ai-chips/)\] | | Tsinghua | Tianjic | Tianjic | 1210 | 0.95 | int8 | Chip | \[[www.nature.com](http://www.nature.com/articles/s41586-019-1424-8)\] | +| Untether | TsunAImi | TsunAImi | 2e+06 | 400 | int8 | Card | \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/hewsletter_detail.php?num=6230)\] | +| XMOS | xcore.ai | xcore.ai | 51.2 | 1 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/xmos-adapts-xcore-into-aiot-crossover-processor/#)\] | --- @@ -132,5 +134,3 @@ For the full dataset in CSV format, please download \[[peak-accelerators-ieee-hp Copyright 2022 MIT, Albert I. Reuther -| Untether | TsunAImi | TsunAImi | 2e+06 | 400 | int8 | Card | \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/hewsletter_detail.php?num=6230)\] | -| XMOS | xcore.ai | xcore.ai | 51.2 | 1 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/xmos-adapts-xcore-into-aiot-crossover-processor/#)\] | diff --git a/peak_2023.md b/peak_2023.md new file mode 100644 index 0000000..139093c --- /dev/null +++ b/peak_2023.md @@ -0,0 +1,156 @@ +# 2023 IEEE-HPEC Paper Data + + +This page has a list of some of the data columns/fields that are included in the 2023 IEEE-HPEC paper: + +A. Reuther, P. Michaleas, M. Jones, V. Gadepally, S. Samsi and J. Kepner, "Lincoln AI Computing Survey (LAICS) Update," *2023 IEEE High Performance Extreme Computing Conference (HPEC)*, 2023, pp. 1-7, \[[IEEE Xplore doi: 10.1109/HPEC58863.2023.10363568](https://doi.org/10.1109/HPEC58863.2023.10363568)\] \[[ArXiv.org/abs/2310.09145]( +https://doi.org/10.48550/arXiv.2310.09145)\]. + + +For the full dataset in CSV format, please download \[[peak-accelerators-ieee-hpec-2023.csv](peak-accelerators-ieee-hpec-2023.csv)\]. + +--- + +| Company | Product | Label | Peak Perf. (GOPs/ GFLOPs) | Peak Power (W) | Precision | Form Factor | References | +| ------- | ------- | ----- | :-----------------------: | :------------: | :-------: | :---------: | ---------- | +| Achronix | VectorPath S7t-VG6 | Achronix | 86000 | 300 | int8 | Card | \[[www.eetimes.com](https://www.eetimes.com/fpga-acceleration-card-delivers-on-bandwidth-speed-and-flexibility/)\] | +| Adapteva | Epiphany-V | Adaptiva | 2050 | 29.26 | fp32 | Chip | \[[arxiv.org](https://arxiv.org/abs/1610.01832)\] \[[doi.org](https://doi.org/10.1109/ACSSC.2014.7094761)\] | +| Aimotive | aiWare3 | Aimotive | 100000 | 25 | int8 | Chip | \[[aimotive.com](https://aimotive.com/news/content/1223)\] | +| AIStorm | AIStorm | AIStorm | 2500 | 0.225 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/startup-accelerates-ai-at-the-sensor/)\] | +| Alibaba | HanGuang 800 | Alibaba | 314252 | 157.126 | int8 | Card | \[[medium.com](https://medium.com/syncedreview/alibabas-new-ai-chip-can-process-nearly-80k-images-per-second-63412dec22a3)\] | +| AlphaIC | RAP-C | AlphaIC | 60000 | 40 | int8 | System | \[[www.alphaics.ai](https://www.alphaics.ai/alphaics-introduces-worlds-most-powerful-ai-platform-alphaedgetm-for-l2-driverless-cars-and-autonomous-systems/)\] | +| AlphaIC | RAP-E | AlphaIC | 30000 | 3 | int8 | Chip | \[[www.eenewsanalog.com](https://www.eenewsanalog.com/news/indo-us-startup-preps-agent-based-ai-processor/page/0/1)\] | +| Amazon | Inferentia | AWSi1 | 128000 | 200 | int8 | Card | \[[perspectives.mvdirona.com](https://perspectives.mvdirona.com/2018/11/aws-inferentia-machine-learning-processor/)\] \[[www.cloudmanagementinsider.com](https://www.cloudmanagementinsider.com/amazon-inferentia-for-machine-learning-and-artificial-intelligence/)\] | +| AMD | Radeon Instinct MI6 | AMD-MI8 | 8190 | 150 | fp16 | Card | \[[blog.exxactcorp.com](https://blog.exxactcorp.com/taking-deeper-look-amd-radeon-instinct-gpus-deep-learning/)\] | +| AMD | Radeon Instinct MI60 | AMD-MI60 | 29500 | 300 | fp16 | Card | \[[www.anandtech.com](https://www.anandtech.com/show/13562/amd-announces-radeon-instinct-mi60-mi50-accelerators-powered-by-7nm-vega)\] | +| AMD | MI100 | AMD-MI100 | 92300 | 300 | fp16 | Card | \[[www.anandtech.com](https://www.anandtech.com/show/17054/amd-announces-instinct-mi200-accelerator-family-cdna2-exacale-servers)\] | +| AMD | MI210 | AMD-MI210 | 181000 | 300 | fp16 | Card | \[[www.anandtech.com](https://www.anandtech.com/show/17054/amd-announces-instinct-mi200-accelerator-family-cdna2-exacale-servers)\] | +| AMD | MI250 | AMD-MI250 | 362000 | 550 | fp16 | Card | \[[www.anandtech.com](https://www.anandtech.com/show/17054/amd-announces-instinct-mi200-accelerator-family-cdna2-exacale-servers)\] | +| AMD | MI300 | AMD-MI300 | 766000 | 700 | fp16 | Card | \[[www.nextplatform.com](https://www.nextplatform.com/2023/06/14/the-third-time-charm-of-amds-instinct-gpu/)\] | +| ARM | Ethos N77 | Ethos | 4100 | 0.8 | int8 | Chip | \[[fuse.wikichip.org](https://fuse.wikichip.org/news/3282/arm-ethos-is-for-ubiquitous-ai-at-the-edge/)\] | +| Axelera | Axelera Test Core | Axelera | 39100 | 2.79 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/axelera-demos-ai-test-chip-after-taping-out-in-four-months/)\] | +| Baidu | Baidu Kunlun 200 | Baidu-K1 | 64000 | 160 | fp16 | Chip | \[[doi.org](https://doi.org/10.1109/ISSCC42613.2021.9366056)\] \[[www.eetimes.com](https://www.eetimes.com/baidu-accelerator-rises-in-ai/)\] \[[www.zdnet.com](https://www.zdnet.com/article/baidu-creates-kunlun-silicon-for-ai/)\] | +| Baidu | Baidu Kunlun II | Baidu-K2 | 192000 | 120 | fp16 | Chip | \[[www.tomshardware.com](https://www.tomshardware.com/news/baidu-unveils-kunlun-ii-processor-for-ai)\] | +| Biren Technology | br100 | br100 | 1.02e+06 | 550 | fp16 | Card | \[[www.hpcwire.com](https://www.hpcwire.com/2022/08/22/chinese-startup-biren-details-br100-gpu/)\] \[[www.tomshardware.com](https://www.tomshardware.com/news/biren-mulls-ipo)\] | +| Biren Technology | br104 | br104 | 512000 | 300 | fp16 | Card | \[[www.hpcwire.com](https://www.hpcwire.com/2022/08/22/chinese-startup-biren-details-br100-gpu/)\] \[[www.tomshardware.com](https://www.tomshardware.com/news/biren-mulls-ipo)\] | +| Bitmain | BM1880 | Bitmain | 1000 | 2.5 | int8 | Chip | \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/newsletter_detail.php?num=5975&year=2019&tag=3)\] | +| Blaize | El Cano | Blaize | 1600 | 7 | int8 | Card | \[[www.blaize.com](https://www.blaize.com/wp-content/uploads/2020/09/Blaize-Ignites-Edge-AI-Performance.pdf)\] | +| Cambricon | MLU100 | Cambricon | 64000 | 80 | fp16 | Card | \[[www.chinamoneynetwork.com](https://www.chinamoneynetwork.com/2018/05/04/chinese-ai-chip-maker-cambricon-unveils-new-cloud-based-smart-chip)\] \[[www.anandtech.com](https://www.anandtech.com/show/12815/cambricon-makers-of-huaweis-kirin-npu-ip-build-a-big-ai-chip-and-pcie-card)\] | +| Cambricon | MLU100 | Cambricon | 128000 | 80 | int8 | Card | \[[www.chinamoneynetwork.com](https://www.chinamoneynetwork.com/2018/05/04/chinese-ai-chip-maker-cambricon-unveils-new-cloud-based-smart-chip)\] \[[www.anandtech.com](https://www.anandtech.com/show/12815/cambricon-makers-of-huaweis-kirin-npu-ip-build-a-big-ai-chip-and-pcie-card)\] | +| Canaan | Kendrite K210 | Kendryte | 230 | 0.3 | int8 | Chip | \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/newsletter_detail.php?num=5992)\] | +| Cerebras | CS-1 | CS-1 | 1.64e+06 | 15000 | fp16 | System | \[[www.cerebras.net](https://www.cerebras.net/introducing-the-cerebras-cs-1-the-industrys-fastest-artificial-intelligence-computer/)\] | +| Cerebras | CS-2 | CS-2 | 3.4e+06 | 23000 | fp16 | System | \[[www.hpcwire.com](https://www.hpcwire.com/2021/04/20/cerebras-doubles-ai-performance-with-second-gen-7nm-wafer-scale-engine/)\] | +| Cornami | Cornami | Cornami | 418792 | 30 | fp16 | Chip | \[[cornami.com](https://cornami.com/1416-2/)\] | +| Enflame | Cloudblazer T10 | Enflame | 8e-08 | 225 | bf16 | Card | \[[www.eenewsanalog.com](https://www.eenewsanalog.com/news/globalfoundries-aids-launch-chinese-ai-startup)\] | +| Esperanto | ET-SoC-1 | ES-1 | 500000 | 120 | int8 | Card | \[[www.computer.org](https://www.computer.org/csdl/magazine/mi/2022/03/09670721/1zUUV1zrRUA)\] \[[fuse.wikichip.org](https://fuse.wikichip.org/news/4911/a-look-at-the-et-soc-1-esperantos-massively-multi-core-risc-v-approach-to-ai/)\] | +| Flex Logix | InferX X1 | FlexLogix | 7640 | 13.5 | int8 | Chip | \[[flex-logix.com](https://flex-logix.com/wp-content/uploads/2020/04/Accelerator-Evaluation-on-Real-Edge-Inference-Applications-04_03_2020.pdf)\] | +| Google | TPU Edge | TPUedge | 4000 | 2 | int8 | System | \[[aiyprojects.withgoogle.com](https://aiyprojects.withgoogle.com/edge-tpu)\] | +| Google | TPU1 | TPU1 | 92000 | 75 | int8 | Chip | \[[doi.org](https://doi.org/10.1145/3360307)\] \[[www.nextplatform.com](https://www.nextplatform.com/2018/05/10/tearing-apart-googles-tpu-3-0-ai-coprocessor/)\] | +| Google | TPU2 | TPU2 | 45000 | 280 | bf16.32 | Chip | \[[doi.org](https://doi.org/10.1145/3360307)\] \[[www.nextplatform.com](https://www.nextplatform.com/2018/05/10/tearing-apart-googles-tpu-3-0-ai-coprocessor/)\] | +| Google | TPU3 | TPU3 | 123000 | 450 | bf16.32 | Chip | \[[doi.org](https://doi.org/10.1145/3360307)\] \[[www.nextplatform.com](https://www.nextplatform.com/2018/05/10/tearing-apart-googles-tpu-3-0-ai-coprocessor/)\] | +| Google | TPU4i | TPU4i | 138000 | 175 | bf16.32 | Chip | \[[doi.org](https://doi.org/10.1109/ISCA52012.2021.00010)\] | +| Google | TPU4 | TPU4 | 275000 | 192 | bf16.32 | Chip | \[[www.hpcwire.com](https://www.hpcwire.com/2022/05/16/google-clouds-new-tpu-v4-ml-hub-packs-9-exaflops-of-ai/)\] | +| GraphCore | C2 | GraphCore | 125000 | 300 | fp16.32 | Card | \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/newsletter_detail.php?num=5992)\] \[[www.graphcore.ai](https://www.graphcore.ai/posts/preliminary-ipu-benchmarks-providing-previously-unseen-performance-for-a-range-of-machine-learning-applications)\] | +| GraphCore | C2 | GraphCoreNode | 1.6e+06 | 2400 | fp16.32 | System | \[[www.graphcore.ai](https://www.graphcore.ai/hubfs/Lead%20gen%20assets/DSS8440%20IPU%20Server%20White%20Paper_2020.pdf)\] | +| GraphCore | Colossus Mk2 | GraphCore2 | 250000 | 300 | bf16 | Card | \[[www.eetimes.com](https://www.eetimes.com/graphcore-takes-on-nvidia-with-second-gen-ai-accelerator/)\] | +| GraphCore | Bow-2000 | GraphCoreBow | 350000 | 300 | bf16 | Card | \[[www.tomshardware.com](https://www.tomshardware.com/news/graphcore-tsmc-bow-ipu-3d-wafer-on-wafer-processor)\] | +| GreenWaves | GAP8 | GAP8 | 22.7 | 0.1 | int8 | Chip | \[[greenwaves-technologies.com](https://greenwaves-technologies.com/gap8_gap9/)\] \[[www.eejournal.com](https://www.eejournal.com/article/gap9-for-ml-at-the-edge/)\] | +| GreenWaves | GAP9 | GAP9 | 151 | 0.64 | int8 | Chip | \[[greenwaves-technologies.com](https://greenwaves-technologies.com/gap8_gap9/)\] \[[www.eejournal.com](https://www.eejournal.com/article/gap9-for-ml-at-the-edge/)\] | +| Groq | Groq Node | GroqNode | 1.5e+06 | 3300 | fp16.32 | System | \[[www.nextplatform.com](https://www.nextplatform.com/2020/09/29/groq-shares-recipe-for-tsp-nodes-systems/)\] | +| Groq | Groq Node | GroqNode | 6e+06 | 3300 | int8 | System | \[[www.nextplatform.com](https://www.nextplatform.com/2020/09/29/groq-shares-recipe-for-tsp-nodes-systems/)\] | +| Groq | Tensor Streaming Processor | Groq | 205000 | 300 | fp16 | Card | \[[groq.com](http://groq.com/wp-content/uploads/2020/04/Groq-Rocks-NNs-Linley-Group-MPR-2020Jan06.pdf)\] \[[doi.org](https://doi.org/10.1109/ISCA45697.2020.00023)\] | +| Groq | Tensor Streaming Processor | Groq | 820000 | 300 | int8 | Card | \[[groq.com](http://groq.com/wp-content/uploads/2020/04/Groq-Rocks-NNs-Linley-Group-MPR-2020Jan06.pdf)\] \[[doi.org](https://doi.org/10.1109/ISCA45697.2020.00023)\] | +| Gyrfalcon | Gyrfalcon | Gyrfalcon | 2800 | 0.224 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/gyrfalcon-unveils-fourth-ai-accelerator-chip/)\] | +| Gyrfalcon | Gyrfalcon | GyrfalconServer | 2.15e+06 | 900 | int8 | System | \[[www.hpcwire.com](https://www.hpcwire.com/off-the-wire/solidrun-gyrfalcon-develop-edge-optimized-ai-inference-server/)\] | +| Habana | Gaudi | Gaudi | 50000 | 200 | fp16 | Card | \[[habana.ai](https://habana.ai/wp-content/uploads/2019/06/Habana-Offers-Gaudi-for-AI-Training.pdf)\] \[[doi.org](https://doi.org/10.1109/MM.2020.2975185)\] | +| Habana | Goya HL-1000 | Goya | 100000 | 150 | int8 | Card | \[[habana.ai](https://habana.ai/wp-content/uploads/2019/06/Habana-Offers-Gaudi-for-AI-Training.pdf)\] \[[doi.org](https://doi.org/10.1109/MM.2020.2975185)\] \[[www.linleygroup.com](https://www.linleygroup.com/mpr/article.php?id=12103)\] | +| Habana | Gaudi2 | Gaudi2 | NaN | 650 | | Card | \[[www.nextplatform.com](https://www.nextplatform.com/2022/05/10/intel-pits-new-gaudi2-ai-training-engine-against-nvidia-gpus/)\] | +| Habana | Greco | Greco | NaN | 75 | | Card | \[[www.nextplatform.com](https://www.nextplatform.com/2022/05/10/intel-pits-new-gaudi2-ai-training-engine-against-nvidia-gpus/)\] | +| Hailo | Hailo-8 | Hailo-8 | 2688 | 9.3 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/details-of-hailo-ai-edge-accelerator-emerge/)\] | +| Hailo | Hailo-15H | Hailo-15 | 20000 | 2 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/hailo-adds-vision-processor-socs-for-smart-cameras/)\] | +| Horizon Robotics | Journey2 | Journey2 | 4000 | 2 | int8 | Chip | \[[en.horizon.ai](https://en.horizon.ai/product/journey)\] | +| Huawei HiSilicon | Ascend 310 | Ascend-310 | 8000 | 8 | fp16 | Chip | \[[e.huawei.com](https://e.huawei.com/us/products/cloud-computing-dc/atlas/ascend-310)\] | +| Huawei HiSilicon | Ascend 310 | Ascend-310 | 16000 | 8 | int8 | Chip | \[[e.huawei.com](https://e.huawei.com/us/products/cloud-computing-dc/atlas/ascend-310)\] | +| Huawei HiSilicon | Ascend 910 | Ascend-910 | 512000 | 310 | fp16 | Chip | \[[e.huawei.com](https://e.huawei.com/us/products/cloud-computing-dc/atlas/ascend-910)\] | +| Huawei HiSilicon | Ascend 910 | Ascend-910 | 512000 | 310 | int8 | Chip | \[[e.huawei.com](https://e.huawei.com/us/products/cloud-computing-dc/atlas/ascend-910)\] | +| IBM | TrueNorth | TrueNorth | 1890 | 0.5 | int8 | System | \[[www.top500.org](https://www.top500.org/news/ibm-finds-killer-app-for-truenorth-neuromorphic-chip/)\] \[[doi.org](https://doi.org/10.1073/pnas.1604850113)\] \[[doi.org](https://doi.org/10.1109/TCAD.2015.2474396)\] | +| IBM | TrueNorth | TrueNorthSys | 1890 | 44 | int8 | System | \[[www.top500.org](https://www.top500.org/news/ibm-finds-killer-app-for-truenorth-neuromorphic-chip/)\] \[[doi.org](https://doi.org/10.1073/pnas.1604850113)\] \[[doi.org](https://doi.org/10.1109/TCAD.2015.2474396)\] | +| IBM/NYU | NeuFlow | NeuFlow | 320 | 10 | int16 | Chip | \[[doi.org](https://doi.org/10.1109/CVPRW.2011.5981829)\] | +| Institute for Computing Technology | DaDianNao | DaDianNao | 5590 | 15.97 | int16 | Chip | \[[cacm.acm.org](https://cacm.acm.org/magazines/2016/11/209123-diannao-family/)\] \[[doi.org](https://doi.org/10.1109/MICRO.2014.58)\] | +| Institute for Computing Technology | DianNao | DianNao | 452 | 0.485 | int16 | Chip | \[[cacm.acm.org](https://cacm.acm.org/magazines/2016/11/209123-diannao-family/)\] | +| Institute for Computing Technology | PuDianNao | PuDianNao | 1060 | 0.596 | int16 | Chip | \[[cacm.acm.org](https://cacm.acm.org/magazines/2016/11/209123-diannao-family/)\] \[[dl.acm.org](https://dl.acm.org/doi/10.1145/2775054.2694358)\] | +| Institute for Computing Technology | ShiDianNao | ShiDianNao | 194 | 0.32 | int16 | Chip | \[[cacm.acm.org](https://cacm.acm.org/magazines/2016/11/209123-diannao-family/)\] \[[doi.org](https://doi.org/10.1145/2749469.2750389)\] | +| Intel | Arria 10 1150 | Arria | 283000 | 85 | fp16.32 | Chip | \[[arxiv.org](https://arxiv.org/abs/1807.06434)\] \[[www.nextplatform.com](https://www.nextplatform.com/2018/07/31/intel-fpga-architecture-focuses-on-deep-learning-inference/)\] | +| Intel | Mobileye EyeQ5 | EyeQ5 | 12000 | 5 | int8 | Chip | \[[www.blaize.com](https://www.blaize.com/wp-content/uploads/2020/09/Blaize-Ignites-Edge-AI-Performance.pdf)\] | +| Intel | Movidius Myriad X | MovidiusX | 1000 | 2 | int16 | Chip | \[[www.extremetech.com](https://www.extremetech.com/computing/254772-new-movidius-myriad-x-vpu-packs-custom-neural-compute-engine)\] | +| Intel | Xe-HPC | Intel-Xe-HPC | 360000 | 600 | fp16 | Card | \[[www.tomshardware.com](https://www.tomshardware.com/news/intel-xe-hpc-ponte-vecchio-examined)\] \[[www.hpcwire.com](https://www.hpcwire.com/off-the-wire/intel-unveils-details-about-sapphire-rapids-cpu-ponte-vecchio-gpu-ipu/)\] \[[www.tomshardware.com](https://www.tomshardware.com/news/intels-ponte-vecchio-smiles-for-the-camera)\] | +| Intel | Flex140 | Flex140 | 105000 | 75 | int8 | Card | \[[www.nextplatform.com](https://www.nextplatform.com/2022/10/04/different-gpu-horses-for-different-datacenter-courses/)\] | +| Intel | Flex170 | Flex170 | 256000 | 150 | int8 | Card | \[[www.nextplatform.com](https://www.nextplatform.com/2022/10/04/different-gpu-horses-for-different-datacenter-courses/)\] | +| Intel | Nervana Lake Crest | Nervana1 | 38000 | 210 | fp32 | Card | \[[newsroom.intel.com](https://newsroom.intel.com/editorials/artificial-intelligence-requires-holistic-approach/)\] | +| Intel | Nervana Spring Crest | Nervana2 | 120000 | 210 | fp16.32 | Card | \[[newsroom.intel.com](https://newsroom.intel.com/editorials/artificial-intelligence-requires-holistic-approach/)\] | +| Intel | Xeon Platinum 8180 | 2xXeon8180 | 4480 | 205 | fp32 | Chip | \[[www.anandtech.com](https://www.anandtech.com/show/14466/intel-xeon-cascade-lake-vs-nvidia-turing)\] \[[www.cpu-world.com](http://www.cpu-world.com/CPUs/Xeon/Intel-Xeon%208180.html)\] | +| Intel | Xeon Platinum 8180 | 2xXeon8280 | 38700 | 205 | int8 | Chip | \[[www.anandtech.com](https://www.anandtech.com/show/14466/intel-xeon-cascade-lake-vs-nvidia-turing)\] \[[www.cpu-world.com](http://www.cpu-world.com/CPUs/Xeon/Intel-Xeon%208280.html)\] | +| Kalray | Coolidge | Kalray | 3000 | 20 | fp16.32 | Chip | \[[www.european-processor-initiative.eu](https://www.european-processor-initiative.eu/dissemination-material/1259/)\] \[[www.eenewsanalog.com](https://www.eenewsanalog.com/news/nxp-kalray-demo-coolidge-parallel-processor-bluebox)\] | +| Kalray | Coolidge | Kalray | 1000 | 20 | fp32 | Chip | \[[www.european-processor-initiative.eu](https://www.european-processor-initiative.eu/dissemination-material/1259/)\] \[[www.eenewsanalog.com](https://www.eenewsanalog.com/news/nxp-kalray-demo-coolidge-parallel-processor-bluebox)\] | +| Kalray | Coolidge | Kalray | 24000 | 20 | int8.32 | Chip | \[[www.european-processor-initiative.eu](https://www.european-processor-initiative.eu/dissemination-material/1259/)\] \[[www.eenewsanalog.com](https://www.eenewsanalog.com/news/nxp-kalray-demo-coolidge-parallel-processor-bluebox)\] | +| Kneron | KL520 Neural Processing Unit | KL520 | 300 | 0.5 | int8 | Chip | \[[www.eetasia.com](https://www.eetasia.com/knerons-next-gen-edge-ai-chip-gets-40m-boost/)\] | +| Kneron | KL720 | KL720 | 1400 | 1.556 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/kneron-attracts-strategic-investors/)\] | +| Maxim | Max 78000 | Maxim | 56 | 0.028 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/maxim-debuts-homegrown-ai-accelerator-in-latest-ulp-soc/)\] \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/newsletter_detail.php?num=6274&year=2021&tag=3)\] \[[doi.org](https://doi.org/10.1117/12.2622390)\] | +| MemryX | MX3 | MX3 | 5000 | 1 | fp16 | Chip | \[[www.eetimes.com](https://www.eetimes.com/adding-low-power-ai-ml-interference-to-edge-devices/)\] | +| Microsoft | Brainwave | Brainwave | 2000 | 150 | int8 | Chip | \[[www.nextplatform.com](https://www.nextplatform.com/2017/08/24/drilling-microsofts-brainwave-soft-deep-leaning-chip/)\] | +| MIT | Eyeriss | Eyeriss | 67.2 | 0.278 | int16 | Chip | \[[doi.org](https://doi.org/10.1109/MM.2017.265085944)\] \[[doi.org](https://doi.org/10.1109/JSSC.2016.2616357)\] \[[doi.org](https://doi.org/10.1109/JPROC.2017.2761740)\] | +| MIT | Netcast | Netcast | 10000 | 0.001 | int8 | Chip | \[[arxiv.org](https://arxiv.org/ftp/arxiv/papers/2203/2203.05466.pdf)\] | +| Mythic | M1076 | Mythic76 | 25000 | 3 | analog | Chip | \[[www.eetimes.com](https://www.eetimes.com/mythic-resizes-its-analog-ai-chip/)\] \[[www.nextplatform.com](https://www.nextplatform.com/2018/08/23/a-mythic-approach-to-deep-learning-inference/)\] \[[medium.com](https://medium.com/mythic-ai/mythic-hot-chips-2018-637dfb9e38b7)\] | +| Mythic | M1108 | Mythic108 | 35000 | 4 | analog | Chip | \[[www.eetimes.com](https://www.eetimes.com/mythic-resizes-its-analog-ai-chip/)\] \[[www.nextplatform.com](https://www.nextplatform.com/2018/08/23/a-mythic-approach-to-deep-learning-inference/)\] \[[medium.com](https://medium.com/mythic-ai/mythic-hot-chips-2018-637dfb9e38b7)\] | +| NovuMind | NovuTensor | NovuMind | 15000 | 15 | int8 | Chip | \[[moorinsightsstrategy.com](https://moorinsightsstrategy.com/wp-content/uploads/2019/05/NovuMind-An-Early-Entrant-in-AI-Silicon-By-Moor-Insights-And-Strategy.pdf)\] \[[www.eetimes.com](https://www.eetimes.com/novuminds-ai-chip-sparks-controversy/)\] | +| NVIDIA | Ampere A10 | A10 | 125000 | 150 | fp16.32 | Card | \[[www.nextplatform.com](https://www.nextplatform.com/2021/04/15/nvidia-rounds-out-ampere-lineup-with-two-new-accelerators/)\] | +| NVIDIA | Ampere A100 | A100 | 312000 | 400 | fp16.32 | Card | \[[devblogs.nvidia.com](https://devblogs.nvidia.com/nvidia-ampere-architecture-in-depth/)\] | +| NVIDIA | Ampere A800 | A800 | 218000 | 300 | fp16.32 | Card | \[[www.tomshardware.com](https://www.tomshardware.com/news/nvidia-a800-performance-revealed)\] | +| NVIDIA | Ampere A30 | A30 | 165000 | 165 | fp16.32 | Card | \[[www.nextplatform.com](https://www.nextplatform.com/2021/04/15/nvidia-rounds-out-ampere-lineup-with-two-new-accelerators/)\] | +| NVIDIA | Ampere A40 | A40 | 150000 | 300 | fp16.32 | Card | \[[www.nextplatform.com](https://www.nextplatform.com/2021/04/15/nvidia-rounds-out-ampere-lineup-with-two-new-accelerators/)\] | +| NVIDIA | DGX Station | DGX-Station | 480000 | 1500 | fp16.32 | System | \[[www.tomshardware.com](https://www.tomshardware.com/news/nvidia-volta-v100-dgx-1-hgx-1,34380.html)\] | +| NVIDIA | DGX-1 | DGX-1 | 900000 | 3500 | fp16.32 | System | \[[www.tomshardware.com](https://www.tomshardware.com/news/nvidia-volta-v100-dgx-1-hgx-1,34380.html)\] \[[www.anandtech.com](https://www.anandtech.com/show/12587/nvidias-dgx2-sixteen-v100-gpus-30-tb-of-nvme-only-400k)\] | +| NVIDIA | DGX-2 | DGX-2 | 1.92e+06 | 10000 | fp16.32 | System | \[[www.anandtech.com](https://www.anandtech.com/show/12587/nvidias-dgx2-sixteen-v100-gpus-30-tb-of-nvme-only-400k)\] | +| NVIDIA | DGX-A100 | DGX-A100 | 5e+06 | 6500 | fp16.32 | System | \[[devblogs.nvidia.com](https://devblogs.nvidia.com/defining-ai-innovation-with-dgx-a100/)\] | +| NVIDIA | DGX-H100 | DGX-H100 | 1.6e+07 | 10200 | fp16.32 | System | \[[wccftech.com](https://wccftech.com/nvidia-unveils-hopper-gh100-powered-dgx-h100-dgx-pod-h100-h100-pcie-accelerators/)\] | +| NVIDIA | H100 | H100 | 1e+06 | 700 | fp16.32 | Card | \[[www.anandtech.com](https://www.anandtech.com/show/17327/nvidia-hopper-gpu-architecture-and-h100-accelerator-announced)\] | +| NVIDIA | Jetson AGX Xavier | XavierAGX | 11000 | 30 | fp16 | System | \[[www.anandtech.com](https://www.anandtech.com/show/15070/nvidia-gives-jetson-xavier-a-trim-announces-nanosized-jetson-xavier-nx)\] | +| NVIDIA | Jetson AGX Xavier | XavierAGX | 32000 | 30 | int8 | System | \[[www.anandtech.com](https://www.anandtech.com/show/15070/nvidia-gives-jetson-xavier-a-trim-announces-nanosized-jetson-xavier-nx)\] | +| NVIDIA | Jetson NX Orin | OrinNX | 50000 | 25 | int8 | System | \[[www.hothardware.com](https://www.hothardware.com/news/nvidia-jetson-agx-orin)\] \[[www.nvidia.com](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/)\] | +| NVIDIA | Jetson AGX Orin | OrinAGX | 138000 | 60 | int8 | System | \[[www.hothardware.com](https://www.hothardware.com/news/nvidia-jetson-agx-orin)\] \[[www.nvidia.com](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/)\] | +| NVIDIA | Jetson TX1 | Jetson1 | 408 | 11.7 | fp16.32 | System | \[[devblogs.nvidia.com](https://devblogs.nvidia.com/jetson-tx2-delivers-twice-intelligence-edge/)\] | +| NVIDIA | Jetson TX2 | Jetson2 | 580 | 12.8 | fp16.32 | System | \[[devblogs.nvidia.com](https://devblogs.nvidia.com/jetson-tx2-delivers-twice-intelligence-edge/)\] | +| NVIDIA | Jetson Xavier NX | XavierNX | 6000 | 15 | fp16 | System | \[[www.anandtech.com](https://www.anandtech.com/show/15070/nvidia-gives-jetson-xavier-a-trim-announces-nanosized-jetson-xavier-nx)\] | +| NVIDIA | Jetson Xavier NX | XavierNX | 21000 | 15 | int8 | System | \[[www.anandtech.com](https://www.anandtech.com/show/15070/nvidia-gives-jetson-xavier-a-trim-announces-nanosized-jetson-xavier-nx)\] | +| NVIDIA | DRIVE AGX L2 | AGX-L2 | 200000 | 45 | int8 | System | \[[hothardware.com](https://hothardware.com/news/nvidia-drive-agx-pegasus-orin-ampere-next-gen-autonomous-cars)\] | +| NVIDIA | DRIVE AGX L5 | AGX-L5 | 2e+06 | 800 | int8 | System | \[[hothardware.com](https://hothardware.com/news/nvidia-drive-agx-pegasus-orin-ampere-next-gen-autonomous-cars)\] | +| NVIDIA | L40 | L40 | 90500 | 300 | fp16 | Card | \[[www.techpowerup.com](https://www.techpowerup.com/gpu-specs/l40.c3959)\] | +| NVIDIA | Pascal P100 | P100 | 21200 | 300 | fp16.32 | Card | \[[www.nvidia.com](https://www.nvidia.com/en-us/data-center/tesla-p100/)\] \[[www.anandtech.com](https://www.anandtech.com/show/12809/16gb-nvidia-tesla-v100-gets-reprieve-remains-in-production)\] | +| NVIDIA | T4 | T4 | 131000 | 150 | int8 | Card | \[[devblogs.nvidia.com](https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/)\] | +| NVIDIA | Volta V100 | V100 | 125000 | 300 | fp16.32 | Card | \[[www.nvidia.com](https://www.nvidia.com/en-us/data-center/tesla-v100/)\] \[[www.anandtech.com](https://www.anandtech.com/show/12809/16gb-nvidia-tesla-v100-gets-reprieve-remains-in-production)\] | +| Perceive | Ergo | Perceive | 4000 | 0.0727 | int8 | Chip | \[[www.forbes.com](https://www.forbes.com/sites/tiriasresearch/2020/04/06/perceive-exits-stealth-with-super-efficient-machine-learning-chip-for-smarter-devices/#1b25ab646d9c)\] | +| PEZY Computing | PEZY-SC2 | PEZY-SC2 | 8190 | 546.13 | fp32 | System | \[[fuse.wikichip.org](https://fuse.wikichip.org/news/191/the-2048-core-pezy-sc2-sets-a-green500-record/)\] | +| Preferred Networks | MN-3 | PN-3 | 524000 | 500 | fp16 | Card | \[[projects.preferred.jp](https://projects.preferred.jp/mn-core/en/)\] \[[www.anandtech.com](https://www.anandtech.com/show/15177/preferred-networks-a-500-w-custom-pcie-card-using-3000-mm2-silicon)\] | +| Quadric | q1-64 | Quadric | 6240 | 12 | int8 | Chip | \[[quadric.io](https://quadric.io/supercomputing.pdf)\] | +| Qualcomm | Cloud AI 100 | Qcomm | 400000 | 75 | int8 | Card | \[[www.eetimes.com](https://www.eetimes.com/qualcomm-cloud-ai-100-promises-impressive-performance-per-watt-for-near-edge-ai/)\] \[[www.eetimes.com](https://www.eetimes.com/qualcomm-targets-ai-inferencing-in-the-cloud/#)\] | +| Qualcomm | QRB5165 | RB5 | 15000 | 30 | int8 | System | \[[www.therobotreport.com](https://www.therobotreport.com/qualcomm-robotics-rb5-platform-puts-5g-ai-in-developers-hands/)\] | +| Qualcomm | QRB5165N | RB6 | 200000 | 30 | int8 | System | \[[www.qualcomm.com](https://www.qualcomm.com/products/internet-of-things/industrial/industrial-automation/robotics-rb6-platform)\] | +| Rockchip | RK3399Pro | RK3399Pro | 2400 | 3 | int8 | Chip | \[[www.rock-chips.com](https://www.rock-chips.com/a/en/News/Press_Releases/2018/0108/869.html)\] | +| SiMa.ai | SiMa.ai | SiMa.ai | 9120 | 4 | int8 | Chip | \[[www.linleygroup.com](https://www.linleygroup.com/uploads/sima-machine-learning-moves-to-the-edge-wp.pdf)\] | +| Stanford | EIE | EIE | 102 | 0.6 | int16 | Chip | \[[ieeexplore.ieee.org](http://ieeexplore.ieee.org/document/7551397/)\] | +| Stanford | TETRIS | Tetris | 128 | 8.42 | int16 | Chip | \[[dl.acm.org](https://dl.acm.org/doi/10.1145/3093337.3037702)\] | +| Syntiant | NDP101 | Syntiant | 200 | 0.01 | int4.8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/tech-heavyweights-back-ai-chip-startup/)\] \[[www.eetimes.com](https://www.eetimes.com/document.asp?doc_id=1334301)\] | +| Tachyum | Prodigy | Tachyum | 1.2e+07 | 950 | fp16.32 | Chip | \[[www.tomshardware.com](https://www.tomshardware.com/news/tachyum-teases-128-core-cpu-57-ghz-950w-16-ddr5-channels)\] | +| Tenstorrent | Tenstorrent | Tenstorrent | 368000 | 75 | int8 | Card | \[[www.tenstorrent.com](https://www.tenstorrent.com/wp-content/uploads/2020/04/Tenstorrent-Scales-AI-Performance.pdf)\] | +| Tesla | Tesla Full Self-Driving Computer | Tesla | 72000 | 72 | int8 | System | \[[doi.org](https://doi.org/10.1109/MM.2020.2975764)\] \[[en.wikichip.org](https://en.wikichip.org/wiki/tesla_(car_company)/fsd_chip)\] | +| Texas Instruments | TDA4VM | TexInst | 8000 | 20 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/tis-first-automotive-soc-with-an-ai-accelerator-launches/)\] \[[www.ti.com](https://www.ti.com/lit/gpn/tda4vm)\] \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/newsletter_detail.php?num=6130&year=2020&tag=3)\] | +| Toshiba | 2015 | Toshiba | 20000 | 10 | int8 | System | \[[www.eetimes.com](https://www.eetimes.com/samsung-toshiba-detail-ai-chips/)\] | +| Tsinghua | Tianjic | Tianjic | 1210 | 0.95 | int8 | Chip | \[[www.nature.com](http://www.nature.com/articles/s41586-019-1424-8)\] | +| Untether | TsunAImi | TsunAImi | 2e+06 | 400 | int8 | Card | \[[www.linleygroup.com](https://www.linleygroup.com/newsletters/hewsletter_detail.php?num=6230)\] | +| XMOS | xcore.ai | xcore.ai | 51.2 | 1 | int8 | Chip | \[[www.eetimes.com](https://www.eetimes.com/xmos-adapts-xcore-into-aiot-crossover-processor/#)\] | + + +--- + + +Copyright 2023 MIT, Albert I. Reuther diff --git a/peak_accelerators_ieee_hpec_2023.csv b/peak_accelerators_ieee_hpec_2023.csv new file mode 100644 index 0000000..835f0bc --- /dev/null +++ b/peak_accelerators_ieee_hpec_2023.csv @@ -0,0 +1,136 @@ +Company,Product,Label,PeakPerformance,Precision,Power,Technology,IorT,FormFactor,PlotLocation,Hide,PlotLocationZ,HideZ,Updated,Notes,ReferenceURL,Reference +Achronix,VectorPath S7t-VG6,Achronix,8.60E+13,int8,300,FPGA,inference,Card,SE,,SW,,6/12/20,,https://www.eetimes.com/fpga-acceleration-card-delivers-on-bandwidth-speed-and-flexibility/,roos2019fpga +Adapteva,Epiphany-V,Adaptiva,2.05E+12,fp32,29.26,multicore,inference,Chip,SW,Hide,SW,Hide,6/9/20,,https://arxiv.org/abs/1610.01832;https://doi.org/10.1109/ACSSC.2014.7094761,olofsson2016epiphany;olofsson2015kickstarting +Aimotive,aiWare3,Aimotive,1.00E+14,int8,25,dataflow,inference,Chip,SW,,SW,,6/2/20,,https://aimotive.com/news/content/1223,aimotive2018aiware3 +"AIStorm ",AIStorm,AIStorm,2.50E+12,int8,0.225,dataflow,inference,Chip,SW,,SW,,6/9/20,Claim to do some of the math up at the sensor in analog domain. Originally came to the embedded space scene with biometric sensors and processing.,https://www.eetimes.com/startup-accelerates-ai-at-the-sensor/,merritt2019aistorm +Alibaba,HanGuang 800,Alibaba,3.14252E+14,int8,157.126,tensor,inference,Card,NE,,NE,,6/14/20,no peak,https://medium.com/syncedreview/alibabas-new-ai-chip-can-process-nearly-80k-images-per-second-63412dec22a3,peng2019alibaba +AlphaIC,RAP-C,AlphaIC,6.00E+13,int8,40,dataflow,inference,System,SW,Hide,SW,Hide,6/2/20,,https://www.alphaics.ai/alphaics-introduces-worlds-most-powerful-ai-platform-alphaedgetm-for-l2-driverless-cars-and-autonomous-systems/,trivedi2019alphaics +AlphaIC,RAP-E,AlphaIC,3.00E+13,int8,3,dataflow,inference,Chip,NW,,NW,,6/13/20,,https://www.eenewsanalog.com/news/indo-us-startup-preps-agent-based-ai-processor/page/0/1,clarke2018indo +Amazon,Inferentia,AWSi1,1.28E+14,int8,200,CPU,inference,Card,SE,,SW,,6/14/20,,https://perspectives.mvdirona.com/2018/11/aws-inferentia-machine-learning-processor/;https://www.cloudmanagementinsider.com/amazon-inferentia-for-machine-learning-and-artificial-intelligence/,hamilton2018aws;cloud2020deep +AMD,Radeon Instinct MI6,AMD-MI8,8.19E+12,fp16,150,GPU,inference,Card,SE,Hide,SE,Hide,6/14/20,,https://blog.exxactcorp.com/taking-deeper-look-amd-radeon-instinct-gpus-deep-learning/,exxactcorp2017taking +AMD,Radeon Instinct MI60,AMD-MI60,2.95E+13,fp16,300,GPU,training,Card,SE,Hide,SE,Hide,6/9/20,,https://www.anandtech.com/show/13562/amd-announces-radeon-instinct-mi60-mi50-accelerators-powered-by-7nm-vega,smith2018amd +AMD,MI100,AMD-MI100,9.23E+13,fp16,300,GPU,training,Card,SE,,NW,,7/31/23,,https://www.anandtech.com/show/17054/amd-announces-instinct-mi200-accelerator-family-cdna2-exacale-servers,smith2021amd +AMD,MI210,AMD-MI210,1.81E+14,fp16,300,GPU,training,Card,SE,,SW,,7/31/23,,https://www.anandtech.com/show/17054/amd-announces-instinct-mi200-accelerator-family-cdna2-exacale-servers,smith2021amd +AMD,MI250,AMD-MI250,3.62E+14,fp16,550,GPU,training,Card,SE,,SE,,7/31/23,,https://www.anandtech.com/show/17054/amd-announces-instinct-mi200-accelerator-family-cdna2-exacale-servers,smith2021amd +AMD,MI300,AMD-MI300,7.66E+14,fp16,700,GPU,training,Card,SE,,SE,,7/31/23,,https://www.nextplatform.com/2023/06/14/the-third-time-charm-of-amds-instinct-gpu/,morgan2023third +ARM,Ethos N77,Ethos,4.10E+12,int8,0.8,dataflow,inference,Chip,SW,,SW,,6/13/20,,https://fuse.wikichip.org/news/3282/arm-ethos-is-for-ubiquitous-ai-at-the-edge/,schor2020arm +Axelera,Axelera Test Core,Axelera,3.91E+13,int8,2.79,dataflow,inference,Chip,NW,,NW,,7/9/22,Test chip,https://www.eetimes.com/axelera-demos-ai-test-chip-after-taping-out-in-four-months/,ward2022axelera +Baidu,Baidu Kunlun 200,Baidu-K1,6.40E+13,fp16,160,CGRA,training,Chip,S,,S,,7/1/21,Baidu claims the FLOPs and power numbers in the zdnet article. The chip has been deployed in Baidu data centers as FPGAs and are being converted to ASICs. No release data has been mentioned in the press. ,https://doi.org/10.1109/ISSCC42613.2021.9366056;https://www.eetimes.com/baidu-accelerator-rises-in-ai/;https://www.zdnet.com/article/baidu-creates-kunlun-silicon-for-ai/,ouyang2021kunlun;merritt2018baidu;duckett2018baidu +Baidu,Baidu Kunlun II,Baidu-K2,1.92E+14,fp16,120,CGRA,training,Chip,S,,SW,,7/31/23,,https://www.tomshardware.com/news/baidu-unveils-kunlun-ii-processor-for-ai,shilov2021baidu +Biren Technology,br100,br100,1.02E+15,fp16,550,GPU,training,Card,SW,,SW,,7/31/23,,https://www.hpcwire.com/2022/08/22/chinese-startup-biren-details-br100-gpu/;https://www.tomshardware.com/news/biren-mulls-ipo,peckham2022chinese;shilov2023chinese +Biren Technology,br104,br104,5.12E+14,fp16,300,GPU,training,Card,SW,,SW,,7/31/23,,https://www.hpcwire.com/2022/08/22/chinese-startup-biren-details-br100-gpu/;https://www.tomshardware.com/news/biren-mulls-ipo,peckham2022chinese;shilov2023chinese +Bitmain,BM1880,Bitmain,1.00E+12,int8,2.5,dataflow,inference,Chip,NE,,NE,,6/13/20,,https://www.linleygroup.com/newsletters/newsletter_detail.php?num=5975&year=2019&tag=3,wheeler2019bitmain +Blaize,El Cano,Blaize,1.60E+12,int8,7,dataflow,inference,Card,NE,,NE,,6/13/21,,https://www.blaize.com/wp-content/uploads/2020/09/Blaize-Ignites-Edge-AI-Performance.pdf,demler2020blaize +Cambricon,MLU100,Cambricon,6.40E+13,fp16,80,dataflow,training,Card,SW,Hide,SW,Hide,6/2/20,,https://www.chinamoneynetwork.com/2018/05/04/chinese-ai-chip-maker-cambricon-unveils-new-cloud-based-smart-chip;https://www.anandtech.com/show/12815/cambricon-makers-of-huaweis-kirin-npu-ip-build-a-big-ai-chip-and-pcie-card,wu2018chinese;cutress2018cambricon +Cambricon,MLU100,Cambricon,1.28E+14,int8,80,dataflow,inference,Card,NW,Hide,NW,Hide,6/2/20,,https://www.chinamoneynetwork.com/2018/05/04/chinese-ai-chip-maker-cambricon-unveils-new-cloud-based-smart-chip;https://www.anandtech.com/show/12815/cambricon-makers-of-huaweis-kirin-npu-ip-build-a-big-ai-chip-and-pcie-card,wu2018chinese;cutress2018cambricon +Canaan,Kendrite K210,Kendryte,2.30E+11,int8,0.3,CPU,inference,Chip,NW,,NW,,6/9/20,,https://www.linleygroup.com/newsletters/newsletter_detail.php?num=5992,gwennap2019kendryte +Cerebras,CS-1,CS-1,1.64E+15,fp16,15000,CGRA,training,System,SE,,SE,,6/11/20,Estimated peak,https://www.cerebras.net/introducing-the-cerebras-cs-1-the-industrys-fastest-artificial-intelligence-computer/,hock2019introducing +Cerebras,CS-2,CS-2,3.40E+15,fp16,23000,CGRA,training,System,SW,,SW,,5/24/21,Estimated peak,https://www.hpcwire.com/2021/04/20/cerebras-doubles-ai-performance-with-second-gen-7nm-wafer-scale-engine/,trader2021cerebras +Cornami,Cornami,Cornami,4.18792E+14,fp16,30,dataflow,training,Chip,NW,,NW,,6/14/20,no peak,https://cornami.com/1416-2/,cornami2020cornami +Enflame,Cloudblazer T10,Enflame,8.00E+01,bf16,225,CPU,training,Card,NW,,NW,,6/13/21,,https://www.eenewsanalog.com/news/globalfoundries-aids-launch-chinese-ai-startup,clarke2019globalfoundries +Esperanto,ET-SoC-1,ES-1,5.00E+14,int8,120,CPU,inference,Card,NW,,NW,,7/31/23,,https://www.computer.org/csdl/magazine/mi/2022/03/09670721/1zUUV1zrRUA;https://fuse.wikichip.org/news/4911/a-look-at-the-et-soc-1-esperantos-massively-multi-core-risc-v-approach-to-ai/,ditzel2022accelerating;schor2021look +Flex Logix,InferX X1,FlexLogix,7.64E+12,int8,13.5,dataflow,inference,Chip,N,Hide,N,Hide,6/12/20,,https://flex-logix.com/wp-content/uploads/2020/04/Accelerator-Evaluation-on-Real-Edge-Inference-Applications-04_03_2020.pdf,mehta2020performance +Google,TPU Edge,TPUedge,4.00E+12,int8,2,tensor,inference,System,SE,,SE,,6/9/20,,https://aiyprojects.withgoogle.com/edge-tpu,tpu2019edge +Google,TPU1,TPU1,9.20E+13,int8,75,tensor,inference,Chip,NW,,SE,,6/1/21,,https://doi.org/10.1145/3360307;https://www.nextplatform.com/2018/05/10/tearing-apart-googles-tpu-3-0-ai-coprocessor/,jouppi2020domain;teich2018tearing +Google,TPU2,TPU2,4.50E+13,bf16.32,280,tensor,training,Chip,SE,,SE,,6/1/21,,https://doi.org/10.1145/3360307;https://www.nextplatform.com/2018/05/10/tearing-apart-googles-tpu-3-0-ai-coprocessor/,jouppi2020domain;teich2018tearing +Google,TPU3,TPU3,1.23E+14,bf16.32,450,tensor,training,Chip,SE,,SE,,6/9/21,,https://doi.org/10.1145/3360307;https://www.nextplatform.com/2018/05/10/tearing-apart-googles-tpu-3-0-ai-coprocessor/,jouppi2021ten;jouppi2020domain;teich2018tearing +Google,TPU4i,TPU4i,1.38E+14,bf16.32,175,tensor,training,Chip,NE,,NE,,6/9/21,,https://doi.org/10.1109/ISCA52012.2021.00010,jouppi2021ten +Google,TPU4,TPU4,2.75E+14,bf16.32,192,tensor,training,Chip,NE,,NE,,7/10/22,,https://www.hpcwire.com/2022/05/16/google-clouds-new-tpu-v4-ml-hub-packs-9-exaflops-of-ai/,peckham2022google +GraphCore,C2,GraphCore,1.25E+14,fp16.32,300,CGRA,training,Card,NE,,NE,,6/14/20,,https://www.linleygroup.com/newsletters/newsletter_detail.php?num=5992;https://www.graphcore.ai/posts/preliminary-ipu-benchmarks-providing-previously-unseen-performance-for-a-range-of-machine-learning-applications,gwennap2020groq;lacey2017preliminary +GraphCore,C2,GraphCoreNode,1.60E+15,fp16.32,2400,CGRA,training,System,SW,,SW,,6/14/20,,https://www.graphcore.ai/hubfs/Lead%20gen%20assets/DSS8440%20IPU%20Server%20White%20Paper_2020.pdf,graphcore2020dell +GraphCore,Colossus Mk2,GraphCore2,2.50E+14,bf16,300,CGRA,training,Card,SW,,NW,,7/10/22,,https://www.eetimes.com/graphcore-takes-on-nvidia-with-second-gen-ai-accelerator/,ward2020graphcore +GraphCore,Bow-2000,GraphCoreBow,3.50E+14,bf16,300,CGRA,training,Card,NE,,NE,,7/10/22,,https://www.tomshardware.com/news/graphcore-tsmc-bow-ipu-3d-wafer-on-wafer-processor,tyson2022graphcore +GreenWaves,GAP8,GAP8,2.27E+10,int8,0.10,dataflow,inference,Chip,NE,,NE,,7/12/21,,https://greenwaves-technologies.com/gap8_gap9/;https://www.eejournal.com/article/gap9-for-ml-at-the-edge/,greenwaves2020gap;turley2020gap9 +GreenWaves,GAP9,GAP9,1.51E+11,int8,0.64,dataflow,inference,Chip,NE,,NE,,7/12/21,,https://greenwaves-technologies.com/gap8_gap9/;https://www.eejournal.com/article/gap9-for-ml-at-the-edge/,greenwaves2020gap;turley2020gap9 +Groq,Groq Node,GroqNode,1.50E+15,fp16.32,3300,tensor,training,System,SE,,SE,,7/6/21,,https://www.nextplatform.com/2020/09/29/groq-shares-recipe-for-tsp-nodes-systems/,hemsoth2020groq +Groq,Groq Node,GroqNode,6.00E+15,int8,3300,tensor,inference,System,SW,,SW,,7/6/21,,https://www.nextplatform.com/2020/09/29/groq-shares-recipe-for-tsp-nodes-systems/,hemsoth2020groq +Groq,Tensor Streaming Processor,Groq,2.05E+14,fp16,300,tensor,training,Card,SE,,NW,,5/19/20,,http://groq.com/wp-content/uploads/2020/04/Groq-Rocks-NNs-Linley-Group-MPR-2020Jan06.pdf;https://doi.org/10.1109/ISCA45697.2020.00023,gwennap2020groq;abts2020think +Groq,Tensor Streaming Processor,Groq,8.20E+14,int8,300,tensor,inference,Card,SE,,SE,,5/19/20,,http://groq.com/wp-content/uploads/2020/04/Groq-Rocks-NNs-Linley-Group-MPR-2020Jan06.pdf;https://doi.org/10.1109/ISCA45697.2020.00023,gwennap2020groq;abts2020think +Gyrfalcon,Gyrfalcon,Gyrfalcon,2.80E+12,int8,0.224,PIM,inference,Chip,NW,,NW,,6/14/20,,https://www.eetimes.com/gyrfalcon-unveils-fourth-ai-accelerator-chip/,ward2019gyrfalcon +Gyrfalcon,Gyrfalcon,GyrfalconServer,2.15E+15,int8,900,PIM,inference,System,NW,,NE,,6/14/20,,https://www.hpcwire.com/off-the-wire/solidrun-gyrfalcon-develop-edge-optimized-ai-inference-server/,hpcwire2020solidrun +Habana,Gaudi,Gaudi,5.00E+13,fp16,200,tensor,training,Card,SW,,SW,,6/14/20,8 VLIW cores with MMA unit,https://habana.ai/wp-content/uploads/2019/06/Habana-Offers-Gaudi-for-AI-Training.pdf;https://doi.org/10.1109/MM.2020.2975185,gwennap2019habanagaudi;medina2020habana +Habana,Goya HL-1000,Goya,1.00E+14,int8,150,tensor,inference,Card,SE,,SE,,6/14/20,8 VLIW cores with MMA unit,https://habana.ai/wp-content/uploads/2019/06/Habana-Offers-Gaudi-for-AI-Training.pdf;https://doi.org/10.1109/MM.2020.2975185;https://www.linleygroup.com/mpr/article.php?id=12103,gwennap2019habanagoya;medina2020habana;gwennap2019habana +Habana,Gaudi2,Gaudi2,,,650,tensor,inference,Card,SE,Skip,SE,Skip,7/9/22,,https://www.nextplatform.com/2022/05/10/intel-pits-new-gaudi2-ai-training-engine-against-nvidia-gpus/,peckham2022intel;morgan2022intel +Habana,Greco,Greco,,,75,tensor,inference,Card,SE,Skip,SE,Skip,7/9/22,,https://www.nextplatform.com/2022/05/10/intel-pits-new-gaudi2-ai-training-engine-against-nvidia-gpus/,peckham2022intel;morgan2022intel +Hailo,Hailo-8,Hailo-8,2.688E+12,int8,9.3,dataflow,inference,Chip,SE,,SW,,6/14/20,no peak,https://www.eetimes.com/details-of-hailo-ai-edge-accelerator-emerge/,ward2019details +Hailo,Hailo-15H,Hailo-15,2.00E+13,int8,2,dataflow,inference,Chip,SW,,SE,,7/31/23,,https://www.eetimes.com/hailo-adds-vision-processor-socs-for-smart-cameras/,ward2023hailo +Horizon Robotics,Journey2,Journey2,4.00E+12,int8,2,dataflow,inference,Chip,N,,N,,6/13/20,,https://en.horizon.ai/product/journey,horizon2020journey +Huawei HiSilicon,Ascend 310,Ascend-310,8.00E+12,fp16,8,tensor,training,Chip,SW,Hide,SW,Hide,5/19/20,,https://e.huawei.com/us/products/cloud-computing-dc/atlas/ascend-310,huawei2020ascend310 +Huawei HiSilicon,Ascend 310,Ascend-310,1.60E+13,int8,8,tensor,inference,Chip,SW,,SW,,5/19/20,,https://e.huawei.com/us/products/cloud-computing-dc/atlas/ascend-310,huawei2020ascend310 +Huawei HiSilicon,Ascend 910,Ascend-910,5.12E+14,fp16,310,tensor,training,Chip,NW,,NE,,5/19/20,,https://e.huawei.com/us/products/cloud-computing-dc/atlas/ascend-910,huawei2020ascend910 +Huawei HiSilicon,Ascend 910,Ascend-910,5.12E+14,int8,310,tensor,inference,Chip,NW,Hide,NW,Hide,5/19/20,,https://e.huawei.com/us/products/cloud-computing-dc/atlas/ascend-910,huawei2020ascend910 +IBM,TrueNorth,TrueNorth,1.89E+12,int8,0.5,neuromorphic,inference,System,SW,Hide,SW,Hide,6/9/20,,https://www.top500.org/news/ibm-finds-killer-app-for-truenorth-neuromorphic-chip/;https://doi.org/10.1073/pnas.1604850113;https://doi.org/10.1109/TCAD.2015.2474396,feldman2016ibm;esser2016convolutional;akopyan2015truenorth +IBM,TrueNorth,TrueNorthSys,1.89E+12,int8,44,neuromorphic,inference,System,SE,Hide,SE,Hide,6/9/20,,https://www.top500.org/news/ibm-finds-killer-app-for-truenorth-neuromorphic-chip/;https://doi.org/10.1073/pnas.1604850113;https://doi.org/10.1109/TCAD.2015.2474396,feldman2016ibm;esser2016convolutional;akopyan2015truenorth +IBM/NYU,NeuFlow,NeuFlow,3.20E+11,int16,10,dataflow,inference,Chip,SE,Hide,SE,Hide,6/15/20,,https://doi.org/10.1109/CVPRW.2011.5981829,farabet2011neuflow +Institute for Computing Technology,DaDianNao,DaDianNao,5.59E+12,int16,15.97,dataflow,inference,Chip,SE,Hide,SE,Hide,6/9/20,,https://cacm.acm.org/magazines/2016/11/209123-diannao-family/;https://doi.org/10.1109/MICRO.2014.58,chen2016diannao;chen2014dadiannao +Institute for Computing Technology,DianNao,DianNao,4.52E+11,int16,0.485,dataflow,inference,Chip,W,Hide,W,Hide,6/9/20,,https://cacm.acm.org/magazines/2016/11/209123-diannao-family/,chen2016diannao +Institute for Computing Technology,PuDianNao,PuDianNao,1.06E+12,int16,0.596,dataflow,inference,Chip,SW,Hide,SW,Hide,6/9/20,,https://cacm.acm.org/magazines/2016/11/209123-diannao-family/;https://dl.acm.org/doi/10.1145/2775054.2694358,chen2016diannao;liu2015pudiannao +Institute for Computing Technology,ShiDianNao,ShiDianNao,1.94E+11,int16,0.32,dataflow,inference,Chip,SW,Hide,SW,Hide,6/9/20,,https://cacm.acm.org/magazines/2016/11/209123-diannao-family/;https://doi.org/10.1145/2749469.2750389,chen2016diannao;du2015shidiannao +Intel,Arria 10 1150,Arria,2.83E+14,fp16.32,85,FPGA,training,Chip,SW,,SW,,6/9/20,,https://arxiv.org/abs/1807.06434;https://www.nextplatform.com/2018/07/31/intel-fpga-architecture-focuses-on-deep-learning-inference/,abdelfattah2018dla;hemsoth2018intel +Intel,Mobileye EyeQ5,EyeQ5,1.20E+13,int8,5,dataflow,inference,Chip,SE,,NE,,7/13/21,Expected release in March 2021,https://www.blaize.com/wp-content/uploads/2020/09/Blaize-Ignites-Edge-AI-Performance.pdf,demler2020blaize +Intel,Movidius Myriad X,MovidiusX,1.00E+12,int16,2,manycore,inference,Chip,SE,Hide,SE,Hide,6/9/20,,https://www.extremetech.com/computing/254772-new-movidius-myriad-x-vpu-packs-custom-neural-compute-engine,hruska2017new +Intel,Xe-HPC,Intel-Xe-HPC,3.60E+14,fp16,600,GPU,training,Card,SE,,NE,,6/1/23,aka Ponte Vecchio,https://www.tomshardware.com/news/intel-xe-hpc-ponte-vecchio-examined;https://www.hpcwire.com/off-the-wire/intel-unveils-details-about-sapphire-rapids-cpu-ponte-vecchio-gpu-ipu/;https://www.tomshardware.com/news/intels-ponte-vecchio-smiles-for-the-camera,shilov2021intels;intel2021intel;shilov2022intels +Intel,Flex140,Flex140,1.05E+14,int8,75,GPU,inference,Card,SE,,NE,,7/31/23,,https://www.nextplatform.com/2022/10/04/different-gpu-horses-for-different-datacenter-courses/,morgan2022different +Intel,Flex170,Flex170,2.56E+14,int8,150,GPU,inference,Card,SE,,SE,,7/31/23,,https://www.nextplatform.com/2022/10/04/different-gpu-horses-for-different-datacenter-courses/,morgan2022different +Intel,Nervana Lake Crest,Nervana1,3.80E+13,fp32,210,manycore,training,Card,SE,Hide,SE,Hide,6/9/20,Hide,https://newsroom.intel.com/editorials/artificial-intelligence-requires-holistic-approach/,rao2018beyond +Intel,Nervana Spring Crest,Nervana2,1.20E+14,fp16.32,210,manycore,inference,Card,SE,Hide,SE,Hide,6/9/20,Hide,https://newsroom.intel.com/editorials/artificial-intelligence-requires-holistic-approach/,rao2018beyond +Intel,Xeon Platinum 8180,2xXeon8180,4.48E+12,fp32,205,multicore,training,Chip,SW,Hide,SW,Hide,6/14/20,,https://www.anandtech.com/show/14466/intel-xeon-cascade-lake-vs-nvidia-turing;http://www.cpu-world.com/CPUs/Xeon/Intel-Xeon%208180.html,degelas2019intel;cpuworld2020xeon8180 +Intel,Xeon Platinum 8180,2xXeon8280,3.87E+13,int8,205,multicore,inference,Chip,SW,Hide,SW,Hide,6/14/20,,https://www.anandtech.com/show/14466/intel-xeon-cascade-lake-vs-nvidia-turing;http://www.cpu-world.com/CPUs/Xeon/Intel-Xeon%208280.html,degelas2019intel;cpuworld2020xeon8280 +Kalray,Coolidge,Kalray,3.00E+12,fp16.32,20,manycore,inference,Chip,E,Hide,E,Hide,5/30/20,,https://www.european-processor-initiative.eu/dissemination-material/1259/;https://www.eenewsanalog.com/news/nxp-kalray-demo-coolidge-parallel-processor-bluebox,"dupont2019kalray, clarke2020nxp" +Kalray,Coolidge,Kalray,1.00E+12,fp32,20,manycore,inference,Chip,SE,Hide,SE,Hide,5/30/20,,https://www.european-processor-initiative.eu/dissemination-material/1259/;https://www.eenewsanalog.com/news/nxp-kalray-demo-coolidge-parallel-processor-bluebox,"dupont2019kalray, clarke2020nxp" +Kalray,Coolidge,Kalray,2.40E+13,int8.32,20,manycore,inference,Chip,NW,,NW,,5/30/20,,https://www.european-processor-initiative.eu/dissemination-material/1259/;https://www.eenewsanalog.com/news/nxp-kalray-demo-coolidge-parallel-processor-bluebox,"dupont2019kalray, clarke2020nxp" +Kneron,KL520 Neural Processing Unit,KL520,3.00E+11,int8,0.5,dataflow,inference,Chip,NW,Hide,NW,Hide,6/2/20,,https://www.eetasia.com/knerons-next-gen-edge-ai-chip-gets-40m-boost/,ward2020kneron +Kneron,KL720,KL720,1.40E+12,int8,1.556,dataflow,inference,Chip,SW,,SW,,6/2/21,,https://www.eetimes.com/kneron-attracts-strategic-investors/,ward2021kneron +Maxim,Max 78000,Maxim,5.60E+10,int8,0.028,dataflow,inference,Chip,SW,,SW,,7/10/22,,https://www.eetimes.com/maxim-debuts-homegrown-ai-accelerator-in-latest-ulp-soc/;https://www.linleygroup.com/newsletters/newsletter_detail.php?num=6274&year=2021&tag=3;https://doi.org/10.1117/12.2622390,ward2020maxim;jani2021maxim;clay2022benchmarking +MemryX,MX3,MX3,5.00E+12,fp16,1.000,dataflow,inference,Chip,NW,,NE,,7/4/23,,https://www.eetimes.com/adding-low-power-ai-ml-interference-to-edge-devices/,leibson2023adding +Microsoft,Brainwave,Brainwave,2.00E+12,int8,150,dataflow,inference,Chip,NE,Hide,NE,Hide,6/13/20,,https://www.nextplatform.com/2017/08/24/drilling-microsofts-brainwave-soft-deep-leaning-chip/,morgan2017drilling +MIT,Eyeriss,Eyeriss,6.72E+10,int16,0.278,dataflow,inference,Chip,S,Hide,S,Hide,6/9/20,,https://doi.org/10.1109/MM.2017.265085944;https://doi.org/10.1109/JSSC.2016.2616357;https://doi.org/10.1109/JPROC.2017.2761740,chen2018eyeriss;chen2017eyeriss;sze2017efficient +MIT,Netcast,Netcast,1.00E+13,int8,0.001,dataflow,inference,Chip,NE,Hide,NE,Hide,5/4/22,Not truly embedded,https://arxiv.org/ftp/arxiv/papers/2203/2203.05466.pdf,sludds2022delocalized +Mythic,M1076,Mythic76,2.50E+13,analog,3,PIM,inference,Chip,NE,,SW,,6/30/21,,https://www.eetimes.com/mythic-resizes-its-analog-ai-chip/;https://www.nextplatform.com/2018/08/23/a-mythic-approach-to-deep-learning-inference/;https://medium.com/mythic-ai/mythic-hot-chips-2018-637dfb9e38b7,ward2021mythic;hemsoth2018mythic;fick2018mythic +Mythic,M1108,Mythic108,3.50E+13,analog,4,PIM,inference,Chip,NE,,NE,,6/30/21,,https://www.eetimes.com/mythic-resizes-its-analog-ai-chip/;https://www.nextplatform.com/2018/08/23/a-mythic-approach-to-deep-learning-inference/;https://medium.com/mythic-ai/mythic-hot-chips-2018-637dfb9e38b7,ward2021mythic;hemsoth2018mythic;fick2018mythic +NovuMind,NovuTensor,NovuMind,1.50E+13,int8,15,dataflow,inference,Chip,SE,,SE,,6/13/20,,https://moorinsightsstrategy.com/wp-content/uploads/2019/05/NovuMind-An-Early-Entrant-in-AI-Silicon-By-Moor-Insights-And-Strategy.pdf;https://www.eetimes.com/novuminds-ai-chip-sparks-controversy/,freund2019novumind;yoshida2018novumind +NVIDIA,Ampere A10,A10,1.25E+14,fp16.32,150,GPU,inference,Card,SW,,SW,,6/29/21,,https://www.nextplatform.com/2021/04/15/nvidia-rounds-out-ampere-lineup-with-two-new-accelerators/,morgan2021nvidia +NVIDIA,Ampere A100,A100,3.12E+14,fp16.32,400,GPU,training,Card,SE,,SE,,6/9/20,,https://devblogs.nvidia.com/nvidia-ampere-architecture-in-depth/,krashinsky2020nvidia +NVIDIA,Ampere A800,A800,2.18E+14,fp16.32,300,GPU,training,Card,SE,,NE,,8/1/23,,https://www.tomshardware.com/news/nvidia-a800-performance-revealed,shilov2023nvidias +NVIDIA,Ampere A30,A30,1.65E+14,fp16.32,165,GPU,inference,Card,NW,,NW,,6/29/21,,https://www.nextplatform.com/2021/04/15/nvidia-rounds-out-ampere-lineup-with-two-new-accelerators/,morgan2021nvidia +NVIDIA,Ampere A40,A40,1.50E+14,fp16.32,300,GPU,inference,Card,NE,,NE,,6/29/21,,https://www.nextplatform.com/2021/04/15/nvidia-rounds-out-ampere-lineup-with-two-new-accelerators/,morgan2021nvidia +NVIDIA,DGX Station,DGX-Station,4.80E+14,fp16.32,1500,GPU,training,System,SE,,SE,,6/9/20,,"https://www.tomshardware.com/news/nvidia-volta-v100-dgx-1-hgx-1,34380.html",alcorn2017nvidia +NVIDIA,DGX-1,DGX-1,9.00E+14,fp16.32,3500,GPU,training,System,SE,,SE,,6/9/20,,"https://www.tomshardware.com/news/nvidia-volta-v100-dgx-1-hgx-1,34380.html;https://www.anandtech.com/show/12587/nvidias-dgx2-sixteen-v100-gpus-30-tb-of-nvme-only-400k",alcorn2017nvidia;cutress2018nvidias +NVIDIA,DGX-2,DGX-2,1.92E+15,fp16.32,10000,GPU,training,System,NW,,NW,,6/9/20,,https://www.anandtech.com/show/12587/nvidias-dgx2-sixteen-v100-gpus-30-tb-of-nvme-only-400k,cutress2018nvidias +NVIDIA,DGX-A100,DGX-A100,5.00E+15,fp16.32,6500,GPU,training,System,NE,,NE,,6/14/20,,https://devblogs.nvidia.com/defining-ai-innovation-with-dgx-a100/,campa2020defining +NVIDIA,DGX-H100,DGX-H100,1.60E+16,fp16.32,10200,GPU,training,System,NW,,SE,,8/4/23,,https://wccftech.com/nvidia-unveils-hopper-gh100-powered-dgx-h100-dgx-pod-h100-h100-pcie-accelerators/,mujtaba2022nvidia +NVIDIA,H100,H100,1.00E+15,fp16.32,700,GPU,training,Card,SE,,SE,,7/5/22,Need to update when Hopper releases,https://www.anandtech.com/show/17327/nvidia-hopper-gpu-architecture-and-h100-accelerator-announced,smith2022nvidia +NVIDIA,Jetson AGX Xavier,XavierAGX,1.10E+13,fp16,30,GPU,inference,System,NE,,SE,,6/29/21,,https://www.anandtech.com/show/15070/nvidia-gives-jetson-xavier-a-trim-announces-nanosized-jetson-xavier-nx,smith2019nvidia +NVIDIA,Jetson AGX Xavier,XavierAGX,3.20E+13,int8,30,GPU,inference,System,SE,,SE,,6/29/21,,https://www.anandtech.com/show/15070/nvidia-gives-jetson-xavier-a-trim-announces-nanosized-jetson-xavier-nx,smith2019nvidia +NVIDIA,Jetson NX Orin,OrinNX,5.00E+13,int8,25,GPU,inference,System,SW,,SW,,7/10/22,,https://www.hothardware.com/news/nvidia-jetson-agx-orin;https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/,"funk2022nvidia,nvidia2022embedded" +NVIDIA,Jetson AGX Orin,OrinAGX,1.38E+14,int8,60,GPU,inference,System,SW,,SW,,7/10/22,,https://www.hothardware.com/news/nvidia-jetson-agx-orin;https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/,"funk2022nvidia,nvidia2022embedded" +NVIDIA,Jetson TX1,Jetson1,4.08E+11,fp16.32,11.7,GPU,inference,System,SW,,SE,,6/9/20,,https://devblogs.nvidia.com/jetson-tx2-delivers-twice-intelligence-edge/,franklin2017nvidia +NVIDIA,Jetson TX2,Jetson2,5.80E+11,fp16.32,12.8,GPU,inference,System,SE,,SE,,6/9/20,,https://devblogs.nvidia.com/jetson-tx2-delivers-twice-intelligence-edge/,franklin2017nvidia +NVIDIA,Jetson Xavier NX,XavierNX,6.00E+12,fp16,15,GPU,inference,System,SE,,SE,,6/29/21,,https://www.anandtech.com/show/15070/nvidia-gives-jetson-xavier-a-trim-announces-nanosized-jetson-xavier-nx,smith2019nvidia +NVIDIA,Jetson Xavier NX,XavierNX,2.10E+13,int8,15,GPU,inference,System,SE,,SE,,6/29/21,,https://www.anandtech.com/show/15070/nvidia-gives-jetson-xavier-a-trim-announces-nanosized-jetson-xavier-nx,smith2019nvidia +NVIDIA,DRIVE AGX L2,AGX-L2,2.00E+14,int8,45,GPU,inference,System,SE,,SE,,7/9/22,,https://hothardware.com/news/nvidia-drive-agx-pegasus-orin-ampere-next-gen-autonomous-cars,hill2020nvidia +NVIDIA,DRIVE AGX L5,AGX-L5,2.00E+15,int8,800,GPU,inference,System,SE,,SW,,7/9/22,,https://hothardware.com/news/nvidia-drive-agx-pegasus-orin-ampere-next-gen-autonomous-cars,hill2020nvidia +NVIDIA,L40,L40,9.05E+13,fp16,300,GPU,training,Card,SE,,SE,,7/31/23,,https://www.techpowerup.com/gpu-specs/l40.c3959,techpowerup2023nvidia +NVIDIA,Pascal P100,P100,2.12E+13,fp16.32,300,GPU,training,Card,SE,,SE,,6/29/21,,https://www.nvidia.com/en-us/data-center/tesla-p100/;https://www.anandtech.com/show/12809/16gb-nvidia-tesla-v100-gets-reprieve-remains-in-production,pascal2018nvidia;smith201816gb +NVIDIA,T4,T4,1.31E+14,int8,150,GPU,inference,Card,NW,,NW,,6/9/20,,https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/,kilgariff2018nvidia +NVIDIA,Volta V100,V100,1.25E+14,fp16.32,300,GPU,training,Card,SE,,SE,,6/9/20,,https://www.nvidia.com/en-us/data-center/tesla-v100/;https://www.anandtech.com/show/12809/16gb-nvidia-tesla-v100-gets-reprieve-remains-in-production,volta2019nvidia;smith201816gb +Perceive,Ergo,Perceive,4.00E+12,int8,7.27E-02,dataflow,inference,Chip,NW,,NW,,6/2/20,concurrent low-power video and audio processing,https://www.forbes.com/sites/tiriasresearch/2020/04/06/perceive-exits-stealth-with-super-efficient-machine-learning-chip-for-smarter-devices/#1b25ab646d9c,mcgregor2020perceive +PEZY Computing,PEZY-SC2,PEZY-SC2,8.19E+12,fp32,546.13,manycore,training,System,SE,Hide,SE,Hide,6/2/20,,https://fuse.wikichip.org/news/191/the-2048-core-pezy-sc2-sets-a-green500-record/,schor2017pezy +Preferred Networks,MN-3,PN-3,5.24E+14,fp16,500,multicore,training,Card,SE,,SE,,6/2/20,,https://projects.preferred.jp/mn-core/en/;https://www.anandtech.com/show/15177/preferred-networks-a-500-w-custom-pcie-card-using-3000-mm2-silicon,"preferred2020mncore, cutress2019preferred" +Quadric,q1-64,Quadric,6.24E+12,int8,12,dataflow,inference,Chip,SW,,SW,,6/13/20,no peak,https://quadric.io/supercomputing.pdf,firu2019quadric +Qualcomm,Cloud AI 100,Qcomm,4.00E+14,int8,75,GPU,inference,Card,NE,,NE,,7/1/21,,https://www.eetimes.com/qualcomm-cloud-ai-100-promises-impressive-performance-per-watt-for-near-edge-ai/;https://www.eetimes.com/qualcomm-targets-ai-inferencing-in-the-cloud/#,ward2020qualcomm;mcgrath2019qualcomm +Qualcomm,QRB5165,RB5,1.50E+13,int8,30,GPU,inference,System,NE,,NE,,8/2/23,30W is 12V @ 2.5A,https://www.therobotreport.com/qualcomm-robotics-rb5-platform-puts-5g-ai-in-developers-hands/,crowe2020qualcomm +Qualcomm,QRB5165N,RB6,2.00E+14,int8,30,GPU,inference,System,NE,,NE,,8/2/23,,https://www.qualcomm.com/products/internet-of-things/industrial/industrial-automation/robotics-rb6-platform,qualcomm2023robotics +Rockchip,RK3399Pro,RK3399Pro,2.40E+12,int8,3,GPU,inference,Chip,SW,,SW,,6/9/20,,https://www.rock-chips.com/a/en/News/Press_Releases/2018/0108/869.html,rockchip2018rockchip +SiMa.ai,SiMa.ai,SiMa.ai,9.12E+12,int8,4,dataflow,inference,Chip,SW,,SW,,6/14/20,no peak,https://www.linleygroup.com/uploads/sima-machine-learning-moves-to-the-edge-wp.pdf,gwennap2020machine +Stanford,EIE,EIE,1.02E+11,int16,0.6,dataflow,inference,Chip,S,Hide,S,Hide,6/15/20,,http://ieeexplore.ieee.org/document/7551397/,han2016eie +Stanford,TETRIS,Tetris,1.28E+11,int16,8.42,dataflow,inference,Chip,N,Hide,N,Hide,6/15/20,,https://dl.acm.org/doi/10.1145/3093337.3037702,gao2017tetris +Syntiant,NDP101,Syntiant,2.00E+11,int4.8,1.00E-02,PIM,inference,Chip,SE,,SE,,6/2/20,"4-bit weights, 8-bit activations",https://www.eetimes.com/tech-heavyweights-back-ai-chip-startup/;https://www.eetimes.com/document.asp?doc_id=1334301,mcgrath2018tech;merritt2018syntiant +Tachyum,Prodigy,Tachyum,1.20E+16,fp16.32,9.50E+02,CPU/GPU,training,Chip,SW,,SE,,7/6/22,,https://www.tomshardware.com/news/tachyum-teases-128-core-cpu-57-ghz-950w-16-ddr5-channels,shilov2022tachyum +Tenstorrent,Tenstorrent,Tenstorrent,3.68E+14,int8,75,multicore,inference,Card,SW,,SE,,6/13/20,,https://www.tenstorrent.com/wp-content/uploads/2020/04/Tenstorrent-Scales-AI-Performance.pdf,gwennap2020tenstorrent +Tesla,Tesla Full Self-Driving Computer,Tesla,7.20E+13,int8,72,tensor,inference,System,NW,,SE,,6/9/20,,https://doi.org/10.1109/MM.2020.2975764;https://en.wikichip.org/wiki/tesla_(car_company)/fsd_chip,talpes2020compute;wikichip2020fsd +Texas Instruments,TDA4VM,TexInst,8.00E+12,int8,20,dataflow,inference,Chip,SE,,SE,,6/29/21,Includes onboard C7x DSP,https://www.eetimes.com/tis-first-automotive-soc-with-an-ai-accelerator-launches/;https://www.ti.com/lit/gpn/tda4vm;https://www.linleygroup.com/newsletters/newsletter_detail.php?num=6130&year=2020&tag=3,ward2020ti;ti2021tda4vm;demler2020ti +Toshiba,2015,Toshiba,2.00E+13,int8,10,multicore,inference,System,NE,,NE,,6/13/20,,https://www.eetimes.com/samsung-toshiba-detail-ai-chips/,merritt2019samsung +Tsinghua,Tianjic,Tianjic,1.21E+12,int8,0.95,neuromorphic,inference,Chip,S,Hide,S,Hide,6/13/20,ANN/SNN research chip,http://www.nature.com/articles/s41586-019-1424-8,pei2019towards +Untether,TsunAImi,TsunAImi,2.00E+15,int8,400,PIM,inference,Card,SW,,SW,,9/2/21,,https://www.linleygroup.com/newsletters/hewsletter_detail.php?num=6230,gwennap2020untether +XMOS,xcore.ai,xcore.ai,5.12E+10,int8,1,dataflow,inference,Chip,SE,Hide,SE,Hide,6/5/20,,https://www.eetimes.com/xmos-adapts-xcore-into-aiot-crossover-processor/#,ward2020xmos \ No newline at end of file