Russ White considers the challenges of using GPU clusters in high performance computing. Aside from possibly lacking software to take advantage of it, the other challenge lies in the connection network. Ethernet is the default standard for this, but requires additional overhead. Russ sees using PCI Express as a much more efficient solution. He considers a PCIe switch from Liqid to dynamically compose infrastructure.
In response to a reader question on his look at Liqid’s composable infrastructure, Russ White frames an interesting question: is it easier to extend PCIe to support switching, and longer runs, or is it easier to design an entire protocol to (effectively) run PCIe over Ethernet? Liqid developed their solution based on former, but other composable infrastructure projects prefer an Ethernet based approach. It’s an interesting look into the benefits and drawbacks of both.
The idea behind composable infrastructure is so cool, it seems like it has to be made up. The basic concept it to be able to dynamically use pooled resources to make servers that fit your current need, rather than make applications and use cases conform to fixed hardware. If I had to personify composable infrastructure, it would be a transformer that’s made up of grey goo nanobots.
Liqid’s composable infrastructure bridges the gap with this fantastic idea with PCIe Fabric and bare metal goodness. Sadly no nanobots.