DirectML Performance Benchmarks - Windows Development

Image Classification (ResNet-50)

Measures inference performance for a common image classification model.

Model: ResNet-50
Dataset: ImageNet (subset)
Batch Size: 32
Precision: FP16

Average Latency: 8.5 ms

                    
// Example Snippet (Conceptual)
auto result = directmlDevice.ExecuteInference(model, inputTensor, outputTensor);
if (result.success) {
    // Process results...
}
                    
                

Object Detection (YOLOv4)

Evaluates throughput for real-time object detection tasks.

Model: YOLOv4
Dataset: COCO (subset)
Batch Size: 16
Precision: FP32

Throughput: 120 FPS

                    
// Example Snippet (Conceptual)
auto frames = captureCameraFeed();
for (const auto& frame : frames) {
    auto detections = directmlDevice.DetectObjects(yoloModel, frame);
    // Render bounding boxes...
}
                    
                

Natural Language Processing (BERT)

Assesses performance for complex NLP tasks like text classification.

Model: BERT-base
Task: Sentiment Analysis
Batch Size: 64
Precision: INT8

Queries per Second: 500 QPS

                    
// Example Snippet (Conceptual)
auto sentences = loadSentences();
for (const auto& sentence : sentences) {
    auto sentiment = directmlDevice.AnalyzeSentiment(bertModel, sentence);
    // Log sentiment...
}
                    
                

Generative Models (StyleGAN2)

Reports generation speed for high-fidelity synthetic data.

Model: StyleGAN2
Task: Image Generation
Latent Vector Size: 512
Precision: FP16

Images per Second: 15 IPS

                    
// Example Snippet (Conceptual)
auto noise = generateLatentVector();
auto generatedImage = directmlDevice.GenerateImage(styleganModel, noise);
saveImage(generatedImage, "output.png");