Jump to content

Talk:Thread block (CUDA programming)

Page contents not supported in other languages.
From Wikipedia, the free encyclopedia

Contains outdated information -- should be updated to current devices

[edit]

I made one or two minor corrections to this. CUDA thread blocks are currently limited to 1024 threads, not 512, as it said in this article, and the max dimensions of the blocks have also been increased. The documents this article cites are out of date, probably by several generations. I've made a very small attempt at bringing parts of it more in line with current hardware, but I certainly didn't check everything in it, and I'm not sure the single reference I added (which is to NVIDIA's documentation) is an acceptable source. I suspect it's considered a "primary source" which is, at least, less than ideal.

The wimpy "citation needed" tag I dropped in is because I'm pretty sure the statement so tagged is incorrect but I'd need to go digging to be sure that I'm not just misunderstanding the terminology used. There are about 80 streaming multiprocessors on a current-generation device and the warps in a block surely don't all have to run on the same one, but the "stream processor" referred to could be a stream scheduler rather than the SMs themselves. (The terminology in this area is nothing if not confusing. That's not Wikipedia's fault, of course.)

Regarding the importance of articles like this one, which is implicitly questioned in the box at the top of this page, those of us working with the hardware find Wikipedia's articles on CUDA devices extremely valuable. They provide a coherent, readable summary of what the terminology means and what various devices can do in a way that no other source I've seen does. (If it weren't for the value of articles like this, I would never have noticed the 512/1024 glitch -- I was looking something up when I stumbled across that.) This particular article, on the other hand, does seem to be a bit of a backwater; it's possible that it could reasonably be merged into one of the other (extremely useful) NVIDIA/CUDA articles. Salaw (talk) 19:16, 19 July 2019 (UTC)[reply]

A little "original research" supported the article's assertion that all threads in a block are restricted to running within a single SM. However, the assertion still needs a citation (and I'm still not sure where that's documented) so the 'citation needed' tag should stay there. Salaw (talk) 23:42, 19 July 2019 (UTC)[reply]