MLPerf_Compatibility_Table.adoc

Table of Contents

• Training
• HPC
• Inference
• Mobile
• Tiny

Training

Model	0.5	0.6	0.7	1.0	1.1	2.0
ResNet-50 v1.5	X	X
SSD-ResNet34	X	X				N/A
RetinaNet-ResNeXt50	N/A					X
MaskRCNN	X	X
NCF	X	N/A
NMT	X	X		N/A
Transformer	X	X		N/A
MiniGo	X	X	X
DLRM	N/A		X
BERT	N/A			X
RNN-T	N/A			X	X
3D U-Net	N/A			X

Metric: Time-to-train (measured in minutes)

Note: v0.6 ResNet-50 v1.5, SSD-ResNet34, NMT increased accuracy targets, all v0.6 benchmarks changed initializition timing, and v0.7 MiniGo moved to 19x19 board

HPC

Model	0.7	1.0
CosmoFlow	X	X
DeepCAM	X
Open Catalyst	N/A	X

Metrics: Time-to-train (measured in minutes) and throughput (weak scaling - measured in models/minute)

Inference

Model	0.5	0.7	1.0	1.1	2.0
MobileNet-v1	X	N/A
ResNet-50 v1.5	X
SSD-MobileNets	X
SSD-ResNet34	X
NMT	X	N/A
DLRM	N/A	X
BERT	N/A	X
RNN-T	N/A	X
3D U-Net	N/A	X			X

Metrics: Queries/second (server), Samples/second (offline), Latency (measured in milliseconds) (single stream), Streams (multi-stream v0.5-v1.1), Latency (measured in milliseconds) (multi-stream 2.0+)

Additional power metrics: System power (measured in watts) (server and offline), system energy per stream (measured in joules) (single stream and multi-stream)

Note: Performance metrics for inference and power submissions are not comparable

Note: Multistream v0.5-v1.1 is not compatible with v2.0 and newer

Mobile

Model	0.7	1.0	1.1	2.0
MobileNetEdge	X
SSD-MobileNetsV2	X	N/A
MobileDET	N/A	X
MOSAIC	N/A			X
MobileBERT	X

Primary metrics: Latency (measured in milliseconds) (single stream), Samples/second (offline)

Note: Submission requires all benchmarks in single stream and MobileNetEdge in single stream and offline

Tiny

Model	0.5	0.7
MobileNetV1	X
ResNet-V1	X*
DSCNN	X
FC Autoencoder	X

Primary metric: Latency (measured in milliseconds)

Secondary metric: Energy per inference (measured in microjoules)

*Latency Compatible, not accuracy: v0.5 and v0.7 use the same model, but changed the evaluation set to improve balance.