MLPerf Inference v0.5 - AshokBhat/ml GitHub Wiki

About

  • Inference Benchmark v0.5 published in Nov 2019

Components

Vision - Image Classification

Load Ref Model Params GOPS/Input Data Set Quality Target Metric
Heavy [ResNet-50]] v1.5 ](/AshokBhat/ml/wiki/25.6M- -7.8- -[[ImageNet) (224x224) 99% OF [FP32]] (76.456%) ](/AshokBhat/ml/wiki/[[Top-1-Accuracy)
Light [MobileNet-v1]] 224 ](/AshokBhat/ml/wiki/4.2M- -1.138- -[[ImageNet) (224x224) 98% OF FP32 (71.676%) Top-1 Accuracy

Vision - Object detection

Load Ref Model Params GOPS/Input Data Set Quality Target Metric
Heavy [SSD]]-ResNet34 ](/AshokBhat/ml/wiki/36.3M- -433- -[[COCO) (1200x1200) 99% OF [FP32]] ](/AshokBhat/ml/wiki/0.20-[[mAP)
Light [SSD]]-MobileNet-V1 ](/AshokBhat/ml/wiki/6.91M -2.47- -[[COCO) (300x300) 99% OF [FP32]] ](/AshokBhat/ml/wiki/0.22-[[mAP)

Language - Machine translation

Load Ref Model Params GOPS/Input Data Set Quality Target Metric
? GNMT 210M WMT16 EN-DE 99% OF FP32 23.9 SacreBleu

Scenarios and Metrics

Scenario Query Generation Metric Samples/Query Examples
Single-Stream (SS) Sequential 90th-Percentile Latency 1 Typing Autocomplete, Real-Time AR
Multistream (MS) Arrival Interval With Dropping Number Of Streams, Subject To Latency Bound N Multicamera Driver Assistance, Large-Scale Automation
Server (S) Poisson Distribution Queries Per Second Subject To Latency Bound 1 Translation Website
Offline (O) Batch Throughput At Least 24,576 Photo Categorization

Closed vs Open systems

Closed division - For comparison of different systems

  • Same models, data sets, and quality targets to ensure comparability across wildly different architectures.

Open division - For fostering innovation

  • Innovation in ML systems, algorithms, optimization, and hardware/software co-design.
  • Same ML task but can change the model architecture and the quality targets

Further information

See also