Emin Gun Sirer expressed:  When you have a system whose performance matters, you need to have a way of measuring and characterizing that performance. There are many aspects to performance, and there are lots of colloquial words (e.g. “speed”) that lack precise definitions.

The main metric that most laypeople are familiar with is “transactions per second” aka TPS. Most people think this measures “speed.” It doesn’t.

TPS measures the ability of a system to complete work per second. It is thus a measure of capacity. Not speed.

The giant Tesla factory in California produces a lot of cars per second. It has high capacity. But the completion time for any individual car has little to do with the capacity of the factory. Starting from steel and rubber, it takes months for a car to be created.

Pipelining, batching, and other techniques can improve capacity. Capacity is important, because high capacity chains like Avalanche burn fees in proportion to their capacity. And users of high capacity systems experience lower fee spikes.

But the metric most people actually care about is, simply put, *speed*. That is, the latency from starting some action to the completion of that action.

High capacity does not mean high speed. Many people think that simply inverting the TPS numbers will give you a measure of latency. This is almost invariably wrong for distributed systems, where multiple things take place concurrently.

As a pandemic cook, I can make 48 pizzas per day. From this, you *can* deduce that it takes me about 30 minutes to make one. Inverting the TPS number yields my latency because there is just plain old me in the kitchen.

The measure of speed for a blockchain is *latency*. The time from submission to completion, also known as time to finality.

Latency, a measure of speed, is measured in seconds, not TPS. High capacity doesn’t necessarily imply low latency.

If someone says they are the *fastest*, and offers numbers measured in TPS, they are either trying to pull a fast one on you or they don’t know what they are talking about.

Low latencies are critical. An AMM that takes 20 seconds to finalize a transaction will always be hopelessly behind real time prices. It will offer lots of opportunities for MEV, which robs users.

An AMM, like Trader Joe, Pango, and YetiSwap, on a fast chain, like Avalanche (yes, it’s both high capacity and low latency), will track real time prices better and offer better trading opportunities to users.

Don’t confuse latencies with other things measured in seconds. Notably, “block time” is meaningless. That is a measure of how fast you start work. Not a measure of how fast you finish it, as every procrastinator knows all too well.

If anyone says they are fast because they have low block times, they are either trying to fool you or are clueless themselves.

For instance, Avalanche doesn’t even have block times, it only issues a block when there’s work to do. What matters is how quickly that block is finalized. Turns out, finalization is around 750ms for Avalanche these days, making it the fastest chain I’ve measured.

Can one do better? Maybe, but that will probably require centralization. If I build a chain that’s implemented on my laptop, I can get even lower numbers, but such systems are worthless. With thousands of validators, Avalanche is also one of the most decentralized chains.

Its high speed, high capacity, and ability to scale to large numbers of validators all stem from its novel consensus protocol.

That concludes the micro lecture, hope it’s useful. Come join me on Fridays at 1pm EST and I talk shop about blockchains and how to evaluate them, among other things.