Repository navigation
Help for LDPC Accelerated #1146
Replies: 5 comments 2 replies
|
*I use JETSON AGX ORIN 64G |
|
Interesting observation, I also expected GPU decoding to outperform CPU in this case. Could this be related to overheads like kernel launch or data transfer, especially for smaller workloads? I’ve been exploring Sionna recently and trying to understand how performance scales across different setups this is a really useful case to follow. |
|
Hi @Miguelclunardi, the CPU vs. GPU performance depends on several system parameters. Note that for simplicity the CUDA LDPC decoding tutorial uses the OAI default interface, which results in a single codeword being decoded per kernel launch. This workload is too small to fully realize the advantages of GPU acceleration; but offloading still frees CPU resources for other tasks. |
|
A good starting point is the OAI LDPC overview and the OAI CUDA LDPC implementation. The latter belongs to a separate OAI branch; I cannot comment on its current status. |
|
Yes, that is likely your current bottleneck. This functionality is not currently implemented in SionnaRK. You would need to verify with the OpenAirInterface community which operating modes are supported by their CUDA-based LDPC decoder. |
Uh oh!
There was an error while loading. Please reload this page.
Hi, I need help with GPU-accelerated LDPC decoding. When I deploy the tutorial, I get unexpected results.
When using the CPU, the performance is:
LDPC decoder: 201.32 us (201.32 us / seg).
However, with the GPU (CUDA), the latency increases significantly:
1872.380781 [NR_PHY] I CUDA LDPC decoder: 530.06 us (530.06 us / seg)
1873.613738 [NR_PHY] I CUDA LDPC decoder: 526.12 us (526.12 us / seg)
Is this correct? I expected the GPU to be faster than the CPU, but it is currently more than twice as slow.
All reactions