About GPU memory required

Hello, I checked runtime environment page (Instance Types) and realized that I can use only 1 GPU with 16GB memory.

Then, are the participants can only use only one 16GB memory of GPU on their submission??

Hello,

Yes, this is correct: 1 T4 GPU (16gb) available. Please consider the following points to make your algorithm match hardware specs:

• Mixed Precision Inference switching from 32-bit floating-point (FP32) to 16-bit floating-point (FP16) or even 8-bit integer (INT8)
• Model Pruning and Quantization
• Layer-by-layer processing where intermediate results are not stored if they are not needed later
• Allocating memory only when needed and reusing memory buffers when possible
• Model Compression Techniques (e.g. distillation)
• Use Smaller Input Resolutions
• Using more optimised inference engines such as TensorFlow Lite, ONNX Runtime, or NVIDIA TensorRT which have optimisations turned on out of the box
• Offload Computation to CPU
• Custom Memory Allocators to better manage the GPU memory

Best, Ezequiel