13:33 <razzy> Should I choose arm or x86 for picolisp computing-heavy server?

13:34 <razzy> can picolisp utilize nvidia graphics card?

13:34 <beneroth> use amd64 architecture (64bit)

13:35 <abu[7]> arm or x86: I think it depends on the hardware

13:35 <abu[7]> nvidia: How? Of course you can, with proper libraries

13:35 <beneroth> picolisp itself doesn't run on the graphics card, but you can use picolisp to run code in graphics card, using the usual libraries for that

13:36 <abu[7]> Haha, same answers overlapped :)

13:36 <beneroth> pil64 used to run well on arm, but now with llvm it's probably not such a big difference anymore?

13:36 <beneroth> I'd guess "standard" amd64 architecture (doesn't matter if amd or intel or whatever) is probably best now

13:37 <abu[7]> Yeah

13:37 <beneroth> ARM if you want to optimize the electricity bill

13:39 <razzy> I buy virtual machine from amazon.

13:39 <razzy> electrical bill does not concern me.

13:39 <abu[7]> Perhaps run some tests on both?

13:40 <beneroth> running number crunching on nvidia cards (e.g. CUDA): I don't think someone has published any picolisp wrappers for CUDA or thelike. You could make wrappers using the C++ interface of those libraries, or simply use python from picolisp (with a bit overhead, but only during I/O I would think)

13:41 <beneroth> but running code on GPU is always limited to some specific kinds of calculations, not general programming.

13:42 <beneroth> I also recommend you to run tests.

13:42 <razzy> beneroth: amazon has really powerfull CPUs, it will be some time, before i need GPUs. I was thinking maybe llvm uses GPUs natively

13:42 <razzy> now

13:42 <beneroth> and Amazon: some companies run tests before they stick with a server. Even when renting the same type of server from AWS there are differences on what kind of server is actually used.

13:43 <beneroth> razzy, GPU cannot run all kinds of programs. GPU is optimized for some specific kinds of highly parallel math computing, e.g. vector/matrix calculations. They are used in 3D rendering (video games, CAD) and in machine learning ("KI")

13:43 <abu[7]> razzy, you mean LLVM has direct Gpu instructions?

13:43 <abu[7]> This would not help

13:43 <abu[7]> You would need to program in llvm-ir

13:44 <abu[7]> Better use some libraries

13:44 <beneroth> yes, better stick to standard libraries

13:44 <razzy> abu[7]: tests are good idea, I will think of them

13:44 <beneroth> the big one is CUDA from nvidia.

13:45 <beneroth> AMD is currently pushing a freshly released competitor, but I don't know the name right now.

13:45 <abu[7]> razzy: Which kind of calculations do you have in mind?

13:45 <beneroth> setting up CUDA is also not very easy. I've done it before, it can be tricky.

13:45 <razzy> beneroth: thx,

13:45 <beneroth> LLVM actually has a AMDGPU backend: https://www.llvm.org/docs/AMDGPUUsage.html

13:45 <abu[7]> interesting!

13:45 <beneroth> but yeah, better use the libraries. Using LLVM means re-inventing those libraries. not worth it unless you have something very specific in mind.

13:46 <abu[7]> T

13:46 <abu[7]> llvm is on the lowest level

13:47 <beneroth> and for start probably easiest is to use python. Python has excellent wrappers of those libraries (python is big because of two reasons: 1) used in teaching 2) they have wrappers for many complex C++ libraries)

13:48 <beneroth> using picolisp to coordinate smaller task-specific python programms and/or using picolisp database for managing some data is probably very effective, probably more effective than using only python.

13:48 <beneroth> razzy, but if you don't use it now, don't bother to much with it.

13:48 <beneroth> and the thing with Amazon servers is that you can change and re-deploy it easily, so I would only pay for it when you use it, not before.

13:49 <razzy> beneroth: yes :)

13:51 <razzy> Thank you all.