Comment by micw 10 days ago
As far as I can see, sse4.1 has been introduced in CPUs in 2011. That's more than 10 years ago. I wonder why such old servers are still in use. I'd assume that a modern CPU would do the same amount of work with a fraction of energy so that it does not even make economical sense to run such outdated hardware.
Does anyone know the numbers of build servers and the specs?
Those single memory channel shitboxes aren't even fast enough to be usable during big windows updates let alone used in production.
It has been introduced in Intel Penryn, in November 2007.
However the AMD CPUs did not implement it until Bulldozer, in mid 2011.
While they lacked the many additional instructions provided by Bulldozer, also including AVX and FMA, for many applications the older Opteron CPUs were significantly faster than the Bulldozer-based CPUs, so there were few incentives for upgrading them, before the launch of AMD Epyc in mid 2017.
SSE 4.1 is a cut point in supporting old CPUs for many software packages, because older CPUs have a very high overhead for divergent computations (e.g. with if ... else ...) inside loops that are parallelized with SIMD instructions.
On the server side, probably not, but I'd like to point out that old hardware is not uncommon, and it's going to be more and more likely as time passes especially in the desktop space.
I was hit by this scenario in the 2000s with an old desktop pc I had, also in the 10ys range, I was using just for boring stuff and random browsing, which was old, but perfectly adequate for the purpose. With time programs got rebuilt with some version of SSE it didn't support. When even firefox switched to the new instruction set, I had to essentially trash a perfectly working desktop pc as it became useless for the purpose.
Because setting up servers is an annoying piece of grunt-work that people avoid doing more than absolutely necessary, there's an reason the expensive options of AWS,Azure and Google cloud make money because much "just works" when focusing on applications rather than the infra (until you actually need to do something advanced and the obscure commands or clicking bites you in the ass).
Hardware after the first couple of generations of x86_64 muliticore processors are perfectly capable machines to use as servers, even for tasks you want to put off to a build farm.
A few months ago Adobe finally updated Lightroom Classic to require these processor extensions. To squeeze all of the matrix mults it can for AI features also in CPU mode.
It's amazing how long of a run top end hardware from ~2011 has had (just missed the cutoff by a few months). It's taken this long for stuff to really require these features.
> I'd assume that a modern CPU would do the same amount of work with a fraction of energy so that it does not even make economical sense to run such outdated hardware.
There are 8,760 hours in a non-leap year. Electricity in the U.S. averages 12.53 cents per kilowatt hour[1]. A really power-hungry CPU running full-bore at 500 W for a year would thus use about $550 of electricity. Even if power consumption dropped by half, that’s only about 10% of the cost of a new computer, so the payoff date of an upgrade is ten years in the future (ignoring the cost of performing the upgrade, which is non-negligible — as is the risk).
And of course buying a new computer is a capital expense, while paying for electricity is an operating expense.
1: https://www.eia.gov/electricity/monthly/epm_table_grapher.ph...