The Impossible Picosecond

On Tuesday, the chief executive of Corvil (which specialises in “latency management for financial markets” – they’re a high-frequency-trading technology firm, in short) stood up at a conference and said something a bit ridiculous. eFinancialNews reported it like this:

Speaking at a London conference on Tuesday, Donal Byrne, chief executive of Corvil, a high-speed trading technology company, caused a ripple of audible incredulity throughout the room when he suggested that trading speeds could be reduced to picoseconds in the not too distant future.

This promptly did the rounds of the news websites and started off a whole lot of uninformed blathering – mostly from Tyler Durden at ZeroHedge, who wailed that the evil front-runners would front-run everybody even faster with their magical picosecond front-running machines, and then went off and hid in his nuclear-safe bunker clutching a 400oz gold bar and muttering about Ben Bernanke’s invisible helicopters.

But what nobody seems to have mentioned is that picosecond-speed trading is impossible.

There are two reasons for this.

(Just to be clear: a picosecond is a trillionth of a second; a nanosecond a billionth; there are 1000 picoseconds in a nanosecond.)

Firstly, think about the computers that power an exchange’s matching engine.┬áThe processors in these computers run at clock speeds of a few gigahertz – a few billion clock ticks per second. Flip that around, and you get something in the low single digits of clock ticks per nanosecond – or one clock tick per few hundred picoseconds.

Now, it doesn’t matter what hardware you’re using – it takes more than one clock tick to match a trade and send it to the trade log and start sending the confirmations. You’re not getting your trades done in hundreds of picoseconds – it’s going to be hundreds of nanoseconds. Algo M2, which claims to be the fastest matching engine in the world, has an internal latency of 500 nanoseconds – and if you want to trim that down to picoseconds, you’ll have to speed up your CPUs by a factor of a thousand.

That could happen in fifteen years, give or take, if Moore’s Law holds.

But there’s another, even better reason why picosecond trading will never happen, and it is this: you’d have to break the speed of light to do it.

All the little electrical impulses inside a computer travel at the speed of light, including the one that says “Buy 100 shares of SPY @ 132.40 for account number 31337”. So the fastest your order can possibly travel is the distance that light can travel in that time.

At this point, it might be worth introducing a visual aid. Here, from the Museum of American History, is a nanosecond:
Grace Hopper's nanosecond

The wire at the bottom left is 30 centimetres long, and is roughly the distance that light can travel in a nanosecond. This nanosecond belonged to Rear Admiral Grace Hopper, a pioneer of computer science and someone who really doesn’t get enough credit for everything she did for computing. She used to hand these nanoseconds out at talks, and she’d add that if you wanted to see a picosecond, you could find a packet of picoseconds in the supermarket: “they’re labelled ‘pepper'”.

And that’s the problem.

The little electrical impulses with your trade in them can only travel as fast as the speed of light – and even at the speed of light, with nothing in the way, it takes 70 picoseconds for them to cross a 21mm-wide Itanium chip. It’s physically impossible to do anything with that data in less than 70 picoseconds.

So ZH’s pseudonymous Tyler Durden can come out out of his bunker; the picosecond robots are not going to front-run him to death with their evil front-running HFT algorithms of algorithmic HFT front-running. And I’m sure Donal Byrne is a great bloke, but I do wonder whether he realised he was proposing to break the speed of light when he made that quote.

(And if anyone ever proposes femtosecond-speed trading, you know they’re making it all up.)

This entry was posted in Money, Rant. Bookmark the permalink.

3 Responses to The Impossible Picosecond

  1. Franck says:

    Very clear explanations, superb post thank you !
    The guy from Corvil which is supposed to be a latency expert wanted to show off but it clearly back-fired and exposed him as incompetent, ouch !

  2. David says:

    One little oversight… Processor designs are already approaching maximum transistor density (you can only make a transistor so small), which puts a limit on the maximum data processing rate. No Moore’s law is not dead… it’s just more likely to be fulfilled by scaling out processor cores. Once dual-processor desktop PCs were unheard of, but pundits are now desktops and laptops pretty much all have dual-core processors, PC servers are packing 6 to 8 core per processors and even phones/portable devices are going the multi-core route. This makes processors faster for almost every workload… but the impossible picosecond requires greater single threaded performance than is theoretically possible with silicon technology. Maybe quantum computers with quantum entanglement data-busses will be capable of the necessary throughput, but it’s certainly not going to happen any time in the next couple of decades.

  3. sgen says:

    What if computer memory can be based on electron spin?

Leave a Reply

Your email address will not be published. Required fields are marked *