Domination has a lead time

June 21, 2023

Suppose that someone makes a prediction like 'PCIE 5.0 will dominate the datacenter in five years' (as a hypothetical example). If we're looking at how likely this is and what would have to happen when in order to get there, one important thing to remember is that domination has a lead time, and this lead time will shorten all of the timelines from what you might otherwise think. The ultimate cause of this lead time is the inertia of the installed base.

Let's take 'dominate the datacenter' as meaning that at least half of the systems in datacenters will use PCIE 5.0. Systems in a datacenter (and things in general) have a lifetime; they're bought, operated for a while, and then replaced with new systems. Thus, at any given time the systems in datacenters are a mix of ages. Some were just bought as replacements or expansions, some are halfway through their expected lifetime, and others are almost at the end of their lifetime and replacements are being planned. This implies that if you want to achieve 50% of currently installed systems having PCIE 5.0 at the five year mark, such systems need to start coming in the datacenter doors in quantity much earlier than five years from now. This is the lead time in action.

This lead time pushes back all of the other timelines involved (even apart from any lead times they may have themselves). If you have to start having PCIE 5.0 systems arriving in datacenters well before five years, obviously those systems have to be available, proven, and attractive by whenever they have to start coming in the door. Before that can happen, you probably need early PCIE 5.0 systems so that people can get experience and shake out bugs (and maybe develop PCIE 5.0 peripherals, get them sold and integrated into systems, and so on). This goes all the way back up the chain of dependencies that leads to people buying PCIE 5.0 systems in quantity in time.

How long a lead time this prediction needs depends in significant part on how fast people turn over their existing systems. As an extreme case, if the typical datacenter system lasted ten years, it's probably already too late to see PCIE 5.0 dominating datacenters in five; even if 100% of the systems sold from now onward were PCIE 5.0, that might only barely reach half of the installed systems in five years (there's some wiggle room for datacenter expansion).

The corollary of this is that you can start to rule out predictions of domination if the lead times necessary given the current state of the world are getting implausibly short. If there are multiple remaining steps necessary between here and there, plausible lead times will depend on how fast you think these steps can be done (or are likely to be done). If you think it will take three years to get PCIE 5.0 fully out into the world and widely available in attractive systems, that leaves only two years to dominate datacenter system sales in order to get the datacenter population up.

PS: This also means there's a big difference between predicting that something will 'dominate the datacenter' and that something will 'dominate sales of datacenter systems'. Dominating sales has basically no required lead time, depending on what sales period you look at.

(However, you may find it implausible that sales will abruptly go from very low level to 'dominating', and therefor want to put in a lead time for sales to ramp up to that level. But abrupt transitions in sales are at least plausible, unlike abrupt transitions in installed bases.)

Written on 21 June 2023.
« Having non-x86 machines makes your life harder, especially in a mixed setup
I don't expect to see competitive RISC-V servers any time soon »

Page tools: View Source, Add Comment.
Login: Password:
Atom Syndication: Recent Comments.

Last modified: Wed Jun 21 21:21:20 2023
This dinky wiki is brought to you by the Insane Hackers Guild, Python sub-branch.