Hacker News Reader

Comment by tovej | Hacker News Reader

Well this seems awfully surface level. This blogpost is essentially just dropping diagrams from AMDs and NVIDIAs documentation and reading the labels on them, with multiple mistakes that tell me that this person hasn't probably hasn't done GPGPU programming (what kind of instruction is "reciprocal"? it's the reciprocal sqrt)

Go read NVIDIAs developer docs and AMDs training material for Frontier for some actually useful introductory material.

For a deeper dive, Jia et al.'s "Dissecting the NVIDIA Volta GPU Architecure via Microbenchmarking" is your goto for NVIDIA material, along with NVIDIAs own documentation. For AMD, you should probably read the CDNA 2 ISA docs

View on HN

Hi, Author here. Thanks for the feedback. I'm happy to correct any mistakes.

Sorry you don't think this is useful. However, as the post clearly points out it's not intended to make you a GPGPU programmer. That really isn't possible in a single blog post. Rather, it tries to give a general overview of how the GP part of GPUs works for the general reader, and to try to relate terminology used by AMD and Nvidia. I'm not aware of another source that tries to do this succinctly.

The end of the post gives lots of links for further reading for readers who want more.

And FWIW I have done quite a bit of GPGPU programming.

Thanks for responding, and sorry for the snide remark, I see a lot of blogspam on this topic and my bullshit detectors are sensitive.

My main gripe is that the blog post does not seem to explain any of the terms it's inteoducing. The text keeps promising to explain concepts that aren't brought up again, and the main differences between AMD and NVIDIA are not discussed (shared memory, warpsize). I did appreciate the history in the introduction, that part was well-written.

If you want to explain the GP part of GPGPUs, I would suggest starting with the threadgrid and moving onto threadblock register footprints and occupancy, it's been a successful recipe for me when I've been teaching, and it keeps the mental load low.

Also for none paid subscribers, is this the full article?

(Does it end with:

"After the break we’ll discuss GPU instruction sets and registers, memory and software." (also side question, what break?) )?

Maybe people reacted this way because they assumed this was the full article

Yes there is a paywall at this point. Probably clearer if it said “the rest of this post is for paid subscribers’ but it’s not something I can change.

Reply on news.ycombinator.com