The P collection material change is designed to optimize knowledge switch between varied elements of an AI server.
“The largest use case is NIC-to-GPU knowledge ingest,” Danesh mentioned. “How do you get from the scale-out community into the GPU so you possibly can maximize GPU utilization, preserve the GPU fed with knowledge, in addition to getting it again out?”
Danesh defined that the basic job of the P collection Scorpio is having the ability to present excessive efficiency peer-to-peer visitors from the GPU to all of its key sources, whether or not it’s the CPU, the NICs, the SSDs.
The X collection targets scale-up clusters, facilitating excessive pace GPU-to-GPU communication. Not like general-purpose switches, the X collection affords platform-specific customizations.
“We are able to truly do platform-specific customization, relying on the totally different GPU, to actually eke out efficiency, eke out one of the best bandwidth effectivity in addition to the very best reliability,” Danesh mentioned.
How the Scorpio change material is deployed
Whereas Astera Labs refers to Scorpio as a change material, the precise know-how the corporate at present offers is the silicon and software program.
