[einpoklum]

> with the C++ API

The funny thing is that the "C++ API" is almost entirely C-like, foregoing almost everything beneficial and convenient about C++, while at the same time not being properly limited to C.

(which is why I wrote this: https://github.com/eyalroz/cuda-api-wrappers/ )

> an awful GPU mailbox design is still the cutting-edge tech

Can you elaborate on what you mean by a "mailbox design"?

[pjmlp]

Depends on which CUDA API one is looking to,

https://docs.nvidia.com/cuda/cuda-c-std/index.html

[einpoklum]

I meant the fundamental ones, mostly:

* CUDA Driver API: https://docs.nvidia.com/cuda/cuda-driver-api/index.html * NVRTC: https://docs.nvidia.com/cuda/nvrtc/index.html * (CUDA Runtime API, very popular but not entirely fundamental as it rests on the driver API)

the CUDA C++ library is a behemoth that sits on top of other things.

[Joel_Mckay]

In general, a modern GPU must copy its workload into/out-of its own working area in vram regardless of the compute capability number, and thus is constrained by the same clock-domain-crossing performance bottleneck many times per transfer.

At least the C++ part of the systems were functional enough to build the current house of cards. Best of luck =3