ONT basecaller SUP (super high accuracy) CUDA error on GPU

Question

[GGG_Alex]

Hi all,

I experience an error when using SUP (super high accuracy) mode. HAC and FAST are running fine on my data (~2GB fast5 files). The error looks like a memory error:

[guppy/error] basecall_core::CUDACaller::thread_main: Exception in CUDACaller worker thread: Could not allocate CUDA shared buffer, size 6404734208 terminate called after throwing an instance of 'std::runtime_error'\ what(): Could not allocate CUDA shared buffer, size 6404734208

I am working with:

• Guppy 6.0.1+652ffd179
• Ubuntu 18.04.
• Nvidia GTX1070
• Driver Version: 510.39.01
• CUDA Version: 11.6

Did anyone experience a similar issue. I cannot find anything in the www.

Best, Alex

[andres.firrincieli]

Maybe a couple fo solutions can be found here

Answer 1

[GGG_Alex]

Just to close this question:

I found that doing demultiplexing and basecalling at the same time using guppy_basecaller will take much more GPU memory. If you experience the same problem, you can do on after an other.