hewumars
5/29/2019 - 6:57 AM

AI芯片对比

readme.md
content_copyfile_download
  • Rendered
  • Source
算力内存带宽（GB/s）解码编码功耗备注
昇腾31016TOPS INT8LPDDR4x 8/4GB3Gb/s每核心H.264/H.265 16路1080P@30FPSJPEG 1080P@256FPS1路1080P@30FPS9.5WCaffe/TensorFlow
昇腾910256TOPS FP16,512TOPS350W
Atlas3004x4TFLOPS,64TOPS@INT8LPDDR4x 4x 8GBH.264/H.265 4x 16路1080P@30FPSJPEG 4x 1080P@256FPS4路1080P@30FPS67W
HI3559A4TOPS@INT8-NNIE+2TOPS@INT8-DSPLPDDR4x 4GB视频：8路1080P@30FPSJPEG：1080P@240FPS5W
算丰BM1880终端芯片1TOPS@INT8（Winograd算计可达2TOPS@INT8）LPDDR4 最大4GB2路1080P@30FPSFHD:1080P@75FPS2.5W多框架支持Caffe\ONNX
算丰BM1682云端芯片3TFLOPS8路1080P@30FPS可堆叠多颗芯片
算丰BM1684y云端芯片17.6 TOPS INT8
2.2 TFLOPS FP32		32路硬解码20W
算丰SC1/SC1+2TFLOPs/4TFLOPs16/32GB501颗BM1680/2颗
Nvidia jetson TX21TFLOPS4TOPS@INT8LPDDR4x 8GBD2D 35D2H/H2D 208路1080P@30FPS7.5W
地平线旭日二代4TOPS
依图questcore15 TOPS64路硬解码20W
AI.md
content_copyfile_download
  • Rendered
  • Source
  • Alibaba: Ali-NPU
  • Amazon: AWS Inferentia
  • Apple: Neural Engine
  • ARM: ML Processor
  • Baidu: Kunlun
  • Bitmain: Sophon
  • Cambricon: MLU
  • Flex Logix: InferX
  • GreenWaves: GAP8
  • Google: TPU
  • Graphcore: IPU
  • Groq:
  • Hailo, Hailo-8
  • Huawei: Ascend
  • Intel: NNP, Myriad, EyeQ, GNA
  • Kendryte: K210
  • NationalChip: Neural Processing Unit (NPU)
  • Nvidia: NVDLA, Xavier
  • Samsung: Neural Processing Unit (NPU)
  • Wave Computing: DPU