MetaRoCE is a transport protocol written from scratch for the conversations between thousands of AI accelerators over plain Ethernet. By the company's own numbers, it keeps about 86 percent of throughput even at one percent packet loss. The spec, the reference code and the compliance test are going public.
- A new protocol for direct memory-to-memory transfer, written specifically for AI over plain Ethernet.
- By Meta's own numbers, it keeps about 86 percent of throughput at 1 percent packet loss.
- Goes public through the Open Compute Project: the spec, the reference implementation libsoftmetaroce, and a compliance test.
A thousand chips train the same model, and the whole time they tell each other what they computed. Lose one packet from that conversation and the standard protocol stops hurrying.
MetaRoCE is Meta's answer. Not a patch on the old RoCE, but a protocol written from a blank page, specifically for the collective operations between thousands of accelerators, the conversations where everyone waits for everyone.
Why put it out in the open
A network protocol written inside one company usually stays there. Meta does the opposite and puts the spec, the code and the test into the Open Compute Project, where any chipmaker can check its own silicon against them.
The math is simple. An open protocol only wins if enough chips speak it. AMD Pensando is the first partner named. If three big chipmakers speak it within a year, Meta has won itself a standard. If not, it just gave away good engineering for free.