Hacker News new | ask | show | jobs
by 0-_-0 2098 days ago
So it will have 5120 SPs in 80 clusters vs. the 3080's 8704 CUDA Cores in 68 clusters. Does this imply better performance?
1 comments

Not necessarily either way. Aside from the standard caveats about graphics cards being more complicated than just their shader performance, they are fairly different beasts.

For example: The RX 5700 XT is about or par with or a hair faster than the RTX 2070, for instance. Where the RX 5700 XT has 2,560 SPs the RTX 2070 has 2,304 CUDA cores. The difference between the 3080 and the 2070 appears much larger but it's worth bearing in mind that Ampere shaders are different to Turing shaders.

In Ampere, what were once dedicated 32-bit integer units now can do floating point operations as well, and as such are counted as extra CUDA cores. See: https://www.guru3d.com/articles-pages/geforce-rtx-3080-and-3...

So depending on workload the RTX 3080 may, in the worst case scenario (equal or more integer operations than floating point operations), act like it 'only' has 4,352 Turing-equivalent CUDA cores. Presumably most workloads are optimised such that doing this makes sense.

For comparison, the Navi/RDNA SPs in the RX 5700XT are apparently all already shared int32/fp32 units but there are more differences, still: https://www.techspot.com/article/1874-amd-navi-vs-nvidia-tur...