External GPU on thunderbolt - and CUDA/MAGMA?

[Boxed Cylon]

I've mostly wandered away from scientific computing (on GPUs and generally), but I've just learned of external GPUs connected to laptops on USB-C Thunderbolt. I've searched this forum and find no mention of anyone attempting scientific computing on such a device (or even if linux supports external GPUs on USB-C thunderbolt). This would be a mighty convenient way to do scientific computing as needed, without having to rebuild one's system to use a high-end graphics card. Can anyone comment on using CUDA/MAGMA for scientific computing on external GPUs?

[Stan Tomov]

I am not sure about the exact mechanism of connecting them and what GPUs can be used as eGPUs, but if one makes the setup the issues of using the GPUs will be similar to systems where you have powerful GPUs connected to a "slow" host CPU (due to slow CPU or slow CPU-GPU data transfers). MAGMA can be used in this setup but best algorithms will be the ones that are "GPU-only", which do not use the CPUs. In MAGMA these would be the BLAS routines, LU, QR, and Cholesky, and all batched routines.

[Boxed Cylon]

It seems using an eGPU - a Gaming Box apparently - is straightforward, using Ubuntu at least:

https://egpu.io/forums/thunderbolt-linu ... lawlessly/

It seems these "Gaming Boxes" are just a USB Thunderbolt to PCIe adapter, with a big power supply:

https://www.amazon.com/Razer-Core-Thund ... 009&sr=8-8

Interesting... How convenient to be able to plug a big GPU in one's laptop, or share it among friends and colleagues, when it is needed.

Presumably the Thunderbolt bandwidth is comparable to PCIe.