Interlaken: The ideal high-speed chip-to-chip interface? | Hacker News Reader

Interlaken: The ideal high-speed chip-to-chip interface?

design-reuse.com

RiscV & open chips seem to all be sticking to TileLink, an early intra & inter-chip interface. I keep wondering how long it will hold, or whether well see some other interfaces come play.

I forgot about Interlaken. Might be useful, maybe?

Tilelink seems to be a memory interface, including cache coherency; Interlaken is a packet interface - each could be built on the other, but with less efficiency.

idk if that is strictly true. coherency protocols can using an underlying framing/flit/error handling/flow control scheme just fine.

treating coherency memory like a message protocol is possible, but would not just have constant performance problems, but scaling issues due to remote dictionary lookups. you could turn all that off with flags I guess.

Cache protocols usually have at least some optimisation for getting the accessed bytes first, which isn't normally present in a protocol designed to exchange packets. What I mean by that is that if your CPU reads byte 27 of a 32byte cache line, the cache is going to read 32 bytes from memory but it will start at byte 27 (or more likely bytes 24..27) so that it can unblock the CPU as fast as possible. I'm not actually sure if that makes its way down to the multicore coherency protocols - if the buses are sufficiently wide maybe they don't bother - but it's seems plausible.

How does this compare to something like pcie? Why would someone use on over the other?

We don't see this replacing the memory-processor bus. That gives some indication of how ideal it is. I should add in my opinion memory interfaces are not ideal. Everything is a trade off.

I prefer good old AXI4

AXI4 is only really used within a chip. This is an inter-chip interconnect. It's also an old one, and I'm not sure it's actually a good solution.

Reply on news.ycombinator.com