4 ms·
A GPU can use 16x 3.0 lanes, theoretically. In practice 8 lanes is enough, people have measured performance between 16x and 8x, it's almost the same. SSDs use
by floatboth 9y ago
A GPU can use 16x 3.0 lanes, theoretically. In practice 8 lanes is enough, people have measured performance between 16x and 8x, it's almost the same.
SSDs use 4x, if you do buy an NVMe drive that is. (In my country, these things are TWICE as expensive as SATA SSDs lol)
- shaklee3 9y agoMaybe for games, but for cuda this is certainly not the case. Double the lanes is double the host to device bandwidth.
- Klathmon 9y agoAnd even for games, the only reason it's not a difference is because game developers target less bandwidth between CPU and GPU so they don't use it all. If we move to a future where every GPU has a lot more bandwidth then games would most likely start taking advantage of it.
- floatboth 9y ago"CUDA" is not a particular application. You could very well be doing something with CUDA that does extremely little data transfer to/from the host, but a lot of computation on the GPU. Like mining :D Also, host to device can be bottlenecked by memory speed on either side.
- shaklee3 9y agoCUDA meaning GPGPU. All I was pointing out is that there is definitely a bandwidth doubling going from 8 to 16 lanes. It's used extremely often by people using the cards for compute. Mining may not have that issue, but deep learning and other applications do.
- TazeTSchnitzel 9y agoMy current “gaming machine” is still using PCIe 1.0 and it doesn't look like it's proving much at all of a bottleneck, because why would it? The game loads most things into memory once and then the rest of the time it's just sending draw calls and maybe geometry updates.
- ionised 9y agoHow are you measuring this?