Things are really parallelizable
Conversation
Sharding / Layer 2s / DAGs are making good use of parallelism, but as soon as you need any type of global consensus on something, you just can't avoid that slowest part of the system standing in your way.
1
That's true in CeFi too though -- and it mostly seems to get on ok
1
In CeFi exchanges usually the order matching engine is the bottleneck which is already highly optimized - Think how spent quarters swapping their Java implementation out with a Rust one to speed that up.
In DeFi the network delay is already orders of magnitudes higher..
1
1
that doesn't really hurt throughput that much though -- it means *latency* is 400ms but block producers can still compute a bunch of order matching at computer speed and then report them, it just takes 400ms for the world to find out
1
Latency affects throughput. In common BFT systems the validators need to chitchat among themselves to reach consensus, and that could take 400 milliseconds. In technically the network can fork many times within that 400ms so it still needs some time to reach consistency.
1
in the timeline that ends up winning (which is almost always the only timeline), there is 400ms of latency sometimes -- but only in between 400ms blocks. So you might get a factor of ~2 slowdown if half the critical path is chitchat but still spend the other half computing.
1
1
My understanding is that the latency in is a few times better than other BFT systems with the assumption that adversaries can't generate the hashes much faster than the majority of the validators. That puts someone without access to a GPU outside of the cluster.
2
1
The amazing throughput part comes with the fact that all validators are using high end servers with large bandwidth, which are quite expensive and not very decentralized by nature.
2
1
The requirements right now are roughly "some sort of desktop", though yeah if you want to increase critical-path speed those will go up.
But realistically speaking, I would be shocked if e.g. 51% of the stakeweight of ETH2.0 wasn't held by people with access to a fast computer.
1
And compare that to current ETH where close to 100% of the hashrate is controlled by people spending *way* more on their setup than you'd need on Solana.
So yeah it does price out 0.01% of the block production coming from cell phones -- but that wasn't changing consensus anyway.