🖥️ Device Transfer Visualizer
Watch tensor data physically cross the PCIe bus between CPU RAM and GPU VRAM.
📋 How To Use This Lab
Flip the
Device
switch to choose
to('cuda')
or
to('cpu')
.
Flip
Pinned Memory
on to see faster transfers — real PyTorch behavior, not just cosmetic.
Press
Next Step ▶
to commit — data packets fly across the PCIe bus with a real computed transfer time.
Try pressing transfer to the device it's
already
on — PyTorch treats it as a free no-op.
Piper
CPU
4-core host
RAM
pageable
GPU
cuda:0
VRAM
GDDR6
PCIe bus
Transfer Bandwidth Comparison
◀ Previous Step
Next Step ▶
Auto Run ▶▶
🤖 PIPER
Set a target device and memory mode, then press Next Step — I'll carry the data across the bus myself.
🧑🏫 PROF. TORCH
.to('cuda') and .to('cpu') copy a tensor across the PCIe bus — pinned host memory roughly doubles the real bandwidth.
🎛️ Choose a Transfer
Target Device
to('cpu')
to('cuda')
Host Memory Mode
Pageable
Pinned