> GPUs. Biggest compute engines we have. How do we program them? 1. Carefully. They only give you - roughly and unless you're super-lucky - around one order of

> GPUs. Biggest compute engines we have. How do we program them? 1. Carefully. They only give you - roughly and unless you're super-lucky - around one order of magnitude improvement in raw flops and raw GB/sec memory bandwidth over a perfectly-exploited CPU. (Not that it's easy to properly exploit a CPU of course; it's actually quite difficult.) That means that if you cut even a few corners - you're just going to lose that edge, and a (again, perfectly-programmed) CPU beats you. 2. With a programming language which has zero or very-low cost abstractions, and models well the computations a GPU "thread" can perform. Preferably with some JIT'ing capability to be dynamic enough. For the first two, that trusty warthog, C++, does fairly well - especially with some help from nice libraries. (Shameless self-plug time... this kind of help: https://github.com/eyalroz/cuda-kat https://github.com/eyalroz/cuda-kat ; I'm thinking of submitting that to "Show HN" soon.) JITing for GPUs is maaaaasively under-explored and under-developed IMHO, and if someone is interesting in collaborating on that, feel free to write me. Would other languages do? If there's not enough ability to abstract, you get a mountain of specialized code; if there's too much abstraction, you start paying through your nose. Could Rust work? I don't know. Perhaps one of those "better C"'s like Zig? The fact that CPU-side libraries can't really run on the GPU kind of levels the playing field against languages with a large base of already-written software. > Ownership... C++ now tries to do "move semantics", with modest success, but has trouble checking at compile time for ownership errors Actually, C++ has, over the past decade or so, introduced several measures to address the issue of ownership and resource leakage. If you combine move semantics, library facilities (mostly smart pointers), the problem is half-solved. Static analysis is improving too, especially when you "decorate" parameters, e.g. with `owner<T>` or `non_null<T>` and such. The "C++ Core Guidelines" (https://github.com/isocpp/CppCoreGuidelines https://github.com/isocpp/CppCoreGuidelines) aim to be machine-checkable whenever possible.