Linus Torvalds Likes His New AMD Threadripper System (zdnet.com) 44
This week Linus Torvalds and Dirk Hohndel re-created their keynote conversation for a special all-virtual edition of the Open Source Summit and Embedded Linux Conference North America. ZDNet reports:
While COVID-19 has slowed down many technologies, while speeding up other tech developments, it hasn't affected Linux development much at all. "None of my co-developers have been hugely impacted either. I was worried for a while because one of our developers was offline for a month or two.... [But,] it turned out that it was just RSI [repetitive strain injury], and RSI is kind of an occupational hazard to deal with." He added. "One of the things that is so interesting about the Linux community is how much it has always been email-based and remote, how rarely we get together in person...."
Torvalds trusts this new build. Indeed, he ran his end of the videoconference from his new developer machine running the first release candidate of 5.8.
Thinking of his new AMD Threadripper 3970x-based processor-powered developer desktop, Torvalds later added that, although he had been concerned about its fan noise it actually works well for him. Torvalds moved to this new homebrew computer because he needed the speed. "My 'allmodconfig' test builds are now three times faster than they used to be." That's important because Torvalds "strives to do about 20 to 30 [pull requests] a day, which is above my limit, ... [but] in order to do that, [he needs] a lot of computing power.
Torvalds trusts this new build. Indeed, he ran his end of the videoconference from his new developer machine running the first release candidate of 5.8.
Thinking of his new AMD Threadripper 3970x-based processor-powered developer desktop, Torvalds later added that, although he had been concerned about its fan noise it actually works well for him. Torvalds moved to this new homebrew computer because he needed the speed. "My 'allmodconfig' test builds are now three times faster than they used to be." That's important because Torvalds "strives to do about 20 to 30 [pull requests] a day, which is above my limit, ... [but] in order to do that, [he needs] a lot of computing power.
slow news day, ZD Net? (Score:3, Informative)
Programmer likes faster computer, news at 11.
Re: (Score:2)
Besides, we already had an article about Linus' new Threadripper-based processor-powered developer desktop computer machine rig box.
Re: (Score:2)
I understood that he does a serious amount of code review. I would call him an architect.
Re: (Score:3)
One that read all the patchs, builds all the linux kernel tree after merging each one and test the kernel, to make sure it works.. sure, one of those project managers!
Re: (Score:1)
What's not to love? (Score:2)
I'd LOVE a new AMD Threadripper 3970x-based processor-powered desktop.
Re: (Score:2)
I have roughly 32 cores :-)
Yes I was including phones! Only a measly 12 cores in my top PC :-/
Re: (Score:3, Informative)
In fact, if you are buying a new Ryzen or Threadripper ignore completely what the TDP says. Buy the biggest, baddest cooling system you can fit in that.
All the recent steppings have an update to the "Boost" Mechanism. They do not overheat with the stock heatsink, but by god, they go through some crazy crank-up speeds at any chance. I have a raft of the older ones around the house and I upgraded the
Re:What's not to love? (Score:4, Informative)
Re: (Score:2)
Re: (Score:2)
Rookie mistake. Anyone with TR that cares about thermals and low noise levels should be going with an AIO water cooler. The thermal mass of water alone is enough to tamper down the occasional thread spikes.
I'd argue that isn't true these days. Decent custom loop with bigger rads and fans are best but failing that the very large thermal mass heatsinks with heatp[ipes on air tend to do better. Likes of Noctua D15 air in a lot of tests does better in appropriate airflow case (which is needed with water anyway) than a lot of aio water setups; both in temps and noise performance as well as not hitting a threshhold where water in small resevoir has warmed significantly after a continued load like I tend to get with
Re: (Score:2)
Have a look at the spec of that monster and get a new aircon. You will need it.
But if he's switching from a recent Intel desktop, he already has it! Did you notice how the recent "65W" CPUs from Intel have ~230W "short-term" power limits? Sure, the "short-term" shouldn't exceed one minute, but when the default setting on many boards you can buy is actually infinity...you get the picture.
Time != Money? (Score:1)
If build times are of such concern to Linus then it seems like he could easily accelerate them far more than just what a Threadripper offers. Dual CPU Epyc maybe, or just rent some cluster time to do a distributed build in the cloud.
Re: (Score:2)
There is a balance there that includes diminishing returns and I think we can assume that Linus Torvalds doesn't need AniMoJo to make the right decision for him in that regard. Call me adventurous but I'll just assume Linus has done his homework.
Re: Time != Money? (Score:1)
You're adventurous.
You're welcome :-). Lol
Re: (Score:2)
Call me adventurous
Hi, Adventurous, I'm dad.
Re: (Score:2)
Noice!
Re: (Score:2)
Also even though a kernel compile is embarrassingly parallel it probably runs faster on higher clocked CPUs rather than more slower clocked CPUs.
Re: (Score:2)
Threadripper should be better then Epyc for these workloads as Threadripper will also provide reasonably good single threaded performance. The task can only be made so parallel before additional cores will not help. If more cores helped then the obvious CPU choice would be the Threadripper 3990. Dual Epyc would kill some workloads but probably not this one.
Anyone want to report the gains resulting from compiling on a 3990 versus a 3970 (or 3990 with 1/2 the cores disabled)? I imagine many here would
Re: (Score:3)
An other reason for going with this machine specifically is probably because this is still the kinds of system you can make silence/quiet.
A distributed build takes money (which he probably doesn't lack) and might actually not be much faster as you have to go over the network to get the code.
Re: (Score:2)
Frankly, if that's a concern he can put a Beowulf cluster of Threadrippers into his basement and use a Ryzen 3600 as a desktop machine.
Re: (Score:2)
A distributed build takes money (which he probably doesn't lack) and might actually not be much faster as you have to go over the network to get the code.
He could use a distributed filesystem, or just rsync, and have local copies of all of the code.
Re: Time != Money? (Score:1)
Re: Time != Money? (Score:2)
Re: Time != Money? (Score:4, Informative)
Write no, compile yes. He's pulling code from all the subsystem maintainers (lieutenants) to create the master branch. And they again are mostly pulling other people's patches. They're supposed to make sure it's clean and tested before it reaches Linus but that's not always what happens. If it breaks or he finds bad code it goes back down the chain of command again so they can fix their code again and resubmit.
Re: (Score:2)
He also wanted something reasonably silent, rack-mount stuff rarely is. Plus he has to pay for the it.
Re: (Score:3)
Maybe you can graph and point out how long it takes to compile the kernel all the way from 1 thread up to 64 threads so Linus can make an informed decision.
Re: (Score:3)
If build times are of such concern to Linus then it seems like he could easily accelerate them far more than just what a Threadripper offers. Dual CPU Epyc maybe, or just rent some cluster time to do a distributed build in the cloud.
If you're building multiple times a day, then renting that power is going to cost more over two months than simply buying it.
Oh good (Score:3)
Now software can continue to bloat and load just as slow as it did two decades ago.
Re: (Score:3)
One man's bloat is another's critical feature. I remember us mocking WebRTC when it was proposed as bloatware, yet now it could have been a major contributor in preventing COVID-19 from crashing the economy even worse than it did before.
You may not do anything differently, but I sure as heck don't use my computer like I did 2 decades ago.
Re: (Score:2)
Loading up SolidWorks today is just as slow as back when I used AutoCAD on a 386. The hardware is orders of magnitude faster but software still loads as slow from an SSD today as it did from a ESDI platter drive back then.
Re: (Score:2)
Loading up SolidWorks today is just as slow as back when I used AutoCAD on a 386.
I'm sure it has all identical functionality as well and there's no difference in features or the underlying engine what so ever.
*Dear reader this post contains sarcasm.
All those cores (Score:3)
Probably make him sending pseudo-code in email faster :-)
Direct contradiction of earlier article (Score:2)
"""
Linus Torvalds: 'I Do No Coding Any More'
"""
From this article:
""" ... [but] in order to do that, [he needs] a lot of computing power.
"My 'allmodconfig' test builds are now three times faster than they used to be." That's important because Torvalds "strives to do about 20 to 30 [pull requests] a day, which is above my limit,
"""
Re: (Score:2, Insightful)
If all he does is merges, test runs, and commits, then he's not really contradicting himself.
Re: (Score:3)
Could still be quiet (Score:1)
I would have thought the guy would have been remoting into a beefier box to do his builds this whole time before COVID-19 instead of waiting around for a hot laptop to do it and hopefully not glitch up from heating fatigue and/or lack of ECC memory. After all Linux is supposed to be really good at remote control, plus if you are running on batteries, doing large compiles tend
"three times faster" (Score:1)