Don MacAskill+ Your Authors @DonMacAskill Chief Geek (and CEO) at @SmugMug & @Flickr. Thrilling our customers. Accidental inventor of consumer SaaS. He/him. Anti-hate. #OpenSafely #BlackLivesMatter Jun. 19, 2020 4 min read + Your Authors

I'm excited to hear about @ARM disrupting laptops and desktops from @Apple next week! But I'd also like to talk about how ARM is disrupting servers in datacenters, especially at @AWSCloud. It's eerily similar, just as profound, and similarly rooted in major @Intel missteps. 1/

This was triggered by @benthompson's excellent @stratechery posts on the subject Tuesday ( ) & Wednesday (if you don't subscribe, you should! ). Go read those first. I'll wait. Good? Good. Let's roll. 2/

By way of background, I'm @AWSCloud's first customer, and I helped invent Internet co-location (aka Internet datacenters) in ~1993, hosting @ebay and more. I also write code regularly. I've been doing this for awhile, and predicting these trends is a mjaor part of my job. 3/

Most importantly, we're using @ARM at large-scale, with production workloads, and have been doing so for years. Graviton2 from @AWSCloud most recently (), but this isn't new for us. We predicted this trend more than 7 years ago. 4/

There are two big bucket topics @stratechery mentions with @ARM on servers. Let's call them "deployment" (cloud services & datacenters) and "development" (coding). Like everything in business, they both boil down to customers, revenue, & profit, but let's tease them apart. 5/

Re: "deployment", this is where things are really interesting. Everyone is aware of how advanced @Apple's ARM chips are, how much value they get from this "tight control of its entire stack", and how they got here (@Intel wouldn't deliver the performance-per-watt needed). 6/

The exact same thing is happening on servers. @Intel has failed to deliver performance-per-dollar (aka per-watt) in a similar way. Out of necessity, companies like @AWSCloud have needed to innovate around this problem. ARM is the solution for them, just like it was for @Apple. 7/

Now, @AWSCloud has been building a tight stack for awhile. It began with custom networking, then extended to custom server hardware with Nitro, and now has moved to the next layer: CPUs. Much like @Apple each layer added to their bespoke stack delivers value & differentiation. 8/

This likely began with a simple ask to @Intel: "We need better perf-per-dollar". And like @Apple a decade earlier, Intel's products in the market the last few years suggest their answer was "No." Which opens up opportunity for @AWSCloud, one they're just beginning to maximize. 9/

What began as driving perf-per-$, now that it has arrived (~40% cheaper!), will again almost certainly mirror @Apple: @AWSCloud will begin adding differentiated value (features you can't get elsewhere) built-in to their own chips given they're now part of their tight stack. 10/

Put another way, the future generation @ARM instances from @AWSCloud will almost certainly not only be cheaper, but empirically better, and different, than their @Intel counterparts in a few generations. Apples-to-apples comparisons across cloud services will be harder. 11/

Not only that, but you won't be *able* to do apples-to-apples comparisons with servers, because these parts won't be available at retail. (Just like @AWSCloud custom switches and Nitro hardware aren't). Cloud services will be definitively *better* than on-premise datacenters. 12/

Note that that's largely a function of scale. The cloud providers (I'm confident @Azure and @GCPcloud will follow suit) drive such significant chip volume (and growing!) that a competitive retail part (and innovation caliber) from others, given lower volumes, is unlikely. 13/

Let's pause for a reminder, dear reader, that this is eerily similar to @Apple and @Intel at the dawn of the iPhone. It's nearly the exact same story, with very similar problems from @Intel, opening up opportunities for those with the scale to tackle them. Ready for more? 14/

Re: "development" Linus Torvalds (creator of Linux) shared his opinion in 2019 about ARM-on-servers "That’s bullshit. If you develop on x86, then you’re going to want to deploy on x86" ( ). I disagree. Nearly every ARM app on iOS/Android is built on x86. 15/

The tooling (IDEs, compilers, emulation, containers, virtualization, etc) have gotten so good for developers this is basically a non-starter. Deploying to an ARM server isn't materially harder than deploying to an App Store. It's trivial even at large-scale production. 16/

Which means it's "just" a value problem. If enough developers see enough value in deploying to ARM, they will. Just like with iOS. With up to 40% cost savings (we're seeing ~this in large-scale production today), that's inevitable. Businesses will demand it. 17/

Re-framing Linus from "you'll happily pay a bit more for x86 cloud hosting" to "you'll happily pay 40% more for x86 cloud hosting" clarifies this. It seems unlikely that would make anyone happy. @ARM is here to stay on servers. And likely to win. /18

Finally, I owe much of the credit for this insight and a debt of gratitude to James Hamilton for this 2012 blog post which got me off the bench and onto "team @ARM" all those years ago. Even if I was a little too early. Better too early than too late.  /end

And it looks like he’s joined Twitter. You should all follow @JrhAtMvDirona. You can thank me later. (Thanks James!)

You can follow @DonMacAskill.


Tip: mention @threader_app on a Twitter thread with the keyword “compile” to get a link to it.

Enjoy Threader? Sign up.

Since you’re here...

... we’re asking visitors like you to make a contribution to support this independent project. In these uncertain times, access to information is vital. Threader gets 1,000,000+ visits a month and our iOS Twitter client was featured as an App of the Day by Apple. Your financial support will help two developers to keep working on this app. Everyone’s contribution, big or small, is so valuable. Support Threader by becoming premium or by donating on PayPal. Thank you.

Follow Threader