News for hackers. Stuff that matters.
Re:Unlaunching the 12GB 4080 by dinobones on Friday October 14, 2022 @03:32PM (#33207861)
I'm still mad about CUDA cores. I thought I was going to be able to write some epic 1000x level parallelism running individually on all cores.

It turns out, a CUDA core is not actually a "core."

Re:Unlaunching the 12GB 4080 by Melatonic on Friday October 14, 2022 @03:38PM (#33207923)
Why do they call it that then? I never really looked into it that much and just took as a measure of a certain type of compute capability ( FP16 or FP64 right? )
Re:Unlaunching the 12GB 4080 by wmf on Friday October 14, 2022 @04:08PM (#33208255)
The SIMT architecture makes it look to the programmer like each FPU is a separate core, but all the cores in an SM have to run in lockstep to get good performance.
Re:Unlaunching the 12GB 4080 by monocasa on Friday October 14, 2022 @04:36PM (#33208523)
They use that metric because it makes their marketing specs look nicer.

Ultimately it's a count of the number of SIMD lanes.