High-end Macs have moved to PCIe 5.0 speeds in their internal drives. Thunderbolt 5 is not fast enough to get the same performance from external ones.
Thunderbolt is also too slow for higher-end networks. A single port is already insufficient for 100-gigabit speeds.
When people talk about 100gigabit networks for Macs, im really curious what kind of network you run at home and how much money you spent on it. Even at work I’m generally seeing 10gigabit network ports with 100gigabit+ only in data centers where macs don’t have a presence
Local AI is probably the most common application these days.
Apple recently added support for InfiniBand over Thunderbolt. And now almost all decent Mac Studio configurations have sold out. Those two may be connected.
> Apple recently added support for InfiniBand over Thunderbolt.
TIL:
* https://developer.apple.com/documentation/technotes/tn3205-l...
Or maybe I forgot:
* https://news.ycombinator.com/item?id=46248644
100 Gb/s Ethernet is likely to be expensive, but dual-port 25 Gb/s Ethernet NICs are not much more expensive than dual-port 10 Gb/s NICs, so whenever you are not using the Ethernet ports already included by a motherboard it may be worthwhile to go to a higher speed than 10 Gb/s.
If you use dual-port NICs, you do not need a high-speed switch, which may be expensive, but you can connect directly the computers into a network, and configure them as either Ethernet bridges or IP routers.
I work in media production and I have the same thought constantly. Hell I curse in church as far as my industry is concerned because I find 2.5 to be fine for most of us. 10 absolutely.
100gbps is going to be for mesh networks supporting clusters (4 Mac Studios let's just say) - not for LAN type networks (unless it's in an actual datacenter).
I suppose the throughput is not the key, latency is. When you split ann operation that normally ran within one machine between two machines, anything that crosses the boundary becomes orders of magnitude slower. Even with careful structuring, there are limits of how little and how rarely you can send data between nodes.
I suppose that splitting an LLM workload is pretty sensitive to that.