Classification
Revision
Public
PA1
Document ID
Print date
19-2241
2019-05-02
A
A
A
A
A
B
B
A
A
A
A
A
A
A
A
B
B
A
A
A
A
A
A
A
A
B
B
1
1
1
1
1
0
Figure 2: Records in GPU memory where several A-scans are missing, indicated in the data valid buffer.
second buffer. When data processing of the first buffer has completed the second buffer can be passed
to the GPU and the process repeats again.
This method ensures that there is always a buffer available for data transfer, avoiding any wait time.
Transfer rates above 7 GB/s have been measured using the attached example code.
3.3
Backplane Peer-To-Peer Transfer
Data is written from the digitizer directly to the GPU without going through the host CPU or host memory.
This reduces requirement on the host system significantly while still utilizing the full transfer rate capability
of the PCIe back plane. The ADQ7 supports up to eight PCIe generation 3.0 lanes.
ADQ7 GPU P2P
GPU
PCIe
switch
Host
CPU
Root
Complex
Figure 3: Peer-to-peer transfer from the digitizer to the GPU.
Fig.
shows transfer from a digitizer to a GPU. Note that transfer takes place via a PCIe switch.
While peer-to-peer transfer is sometimes supported directly through the root complex without involving
ADQ7 GPU Peer-To-Peer – User Guide
Page 4 of 11
