> I don't fully understand. So 1 SSD need 1 lane.

M.2 SSDs supports up to PCIe 4x, though there are also PCIe card formats which support 8x and 16x: http://blog.seagate.com/intelligent/seagate-demonstrates-fas...

> How many lanes need something like NVisia 1080Ti?

16 per (technically they need 4 per — that's what TB3 gives — or possibly less, but they'll use 16 if they can)

> If I want to build something like 1 CPU/1 GPU/1 SSD configuration without additional peripherals, 62 lanes will be effectively wasted?

You're at 24 lanes for the GPU (16), SSD (4) and southbridge (4). And unused lanes are no more wasted than unused USB slots, seriously.

But remember that Threadripper is an enthusiast's platform, people who might have 2~4 SSDs (8~16 lanes) or SLI GPUs (2x16).

This is for high-end workstations after all. Many people are suffering with video production. New cameras this year output 8K video raw (or slightly compressed). Even the guys in RED couldn't produce hardware to properly feed the stream into a computer realtime or at least without having to wait.

These comments baffle me. Doesn't the majority of HN work in the industry? Clearly the driving factor regarding IO at this point is machine learning and networking for cloud applications, not gaming.

Desktop/workstation (e.g. gaming) performance hasn't been driving the industry for easily 5 years now. Intel and AMD both simply use their high performance desktop offerings as proofs for their server solutions, which typically come a few months later.

The newer M.2 SSDs suck up 4 lanes. That's 16 lanes for four drives, burning through all available lanes on a Kaby Lake CPU.

> burning through all available lanes on a Kaby Lake CPU.

Don't you also get something like 4x via the chipset through DMI?

This is an interesting possibility:

https://www.reddit.com/r/Amd/comments/6ee5qw/thanks_to_threa...

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)

Maybe for games, but for cuda this is certainly not the case. Double the lanes is double the host to device bandwidth.

"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.

CUDA 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.

And 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.

My 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.

How are you measuring this?