I'd certainly like to know how to get started w/GPUs. Most of my day to day work doesn't require it. I do wonder if it would make anything faster.

> @U0525KG62 I do wonder if it would make anything faster. Aside from 3D graphics, it seems that any project working with GPU needs to acknowledge its limitations: • GPU has bandwidth and latency communicating with main memory/CPU/SSD - so avoid problems where these will become the bottleneck. • GPU Memory is pretty small on consumer hardware - (8GB to 24GB) . So avoid where you need lots of memory • GPUs are fast at parallel computation of the same thing - compared to CPUs that can do single/multi-threaded calculations of arbitrary different things. • Programming languages for GPU are often less user friendly compared to CPU based programming languages like Clojure. So in the end, it often feels like quite a limited number of niches (e.g. neural network training) where GPU programming is the best solution.

@U06MDAPTP - I agree 4090 has made a decent performance improvement over 3090 (possibly 2x in some usecases). The ethereum mining craze is now over so it might be possible to get these near MSRP.

I fully agree with the first two points. For point 3, CPU also has SIMD and MIMD which can make a huge difference compared to standard instructions. These are also for parallel crunching - just no where near a GPU. For point 4, Dragan's work gives you a lot of friendly Clojure. Neanderthal works on both CPU and GPU with basically no code changes. Though you still have to get your GPU setup to just work (with correct drivers, and CUDA or OpenCL). Generally, due to the first two points, and the pain of just getting a GPU working for your setup, I just go with CPU (which uses the *IMD instructions via MKL). A 100x plus performance improvement over Java is still excellent.