Amazon's cloud unit readies more powerful data center chip – sources | Hacker News Reader

Amazon's cloud unit readies more powerful data center chip – sources

reuters.com

64 pointsekoutanov18 comments

This is exciting!

I use AWS's existing a1 generation for map-reduce jobs on the Common Crawl. They're slower in absolute terms than the Intel-based c5 generation, but are cheaper on a $/GB-processed basis. This is despite my runtime of choice (the JVM) being much more heavily optimized for x86 than for ARM. I can imagine there's a lot of headroom for improvements both at the hardware level and at the software optimization level as adoption grows.

Basically the playbook for the cloud providers is to pretend that Intel and AMD don’t need to worry about it for as long as possible, then suddenly they will be close to price parity and then it’s bad news for Intel and AMDs server business.. hard to say on timeline though.

Makes sense. In 2012, I was at a Fortune 10 oil and gas company that suddenly got squeezed by Oracle as our licenses were expiring. Given that most of the in house software was running on Oracle and leveraging one off functionality and APIs not standard in competitors or in the open source community and that most of the original devs had moved on (most were IT contractors anyway), we had little options. And the rates Oracle gave us were absurd, but they knew we were locked in, and they understood the industry was doin well and we had money (this is before the markets downward dive).

If you’re a cloud company, you absolutely, 100% do not want to depend on an Intel or AMD. If anything, as CEO, this is probably one of the what if scenarios that would keep you up at night.

It goes in both directions a bit. Because the biggest cloud providers are so huge, they do drive a massive chunk of revenue for Intel especially. I remember watching a Q&A with Scott Guthrie where he mentioned how much friendlier the Intel folks are now that Azure is such an enormous customer.

Yea good point. Intel will be like “we get it, you want to replace us, let us build the chip you want to build. We will give you a sweetheart deal.” And Amazon will be like cool we will but we are still working on our own thing. Which will pressure intel to keep their prices low.

While in theory there could be a lot more competition for arm based compute platforms, is there any risk of the cloud providers adding custom extensions to their silicon to increase performing, creating yet another platform lock-in? I'm thinking if the Intel compiler is somewhat analogous.

Napkins Maths.

Intel make roughly $20B from Datacenter this year, and the estimate was nearly 50% goes to the HyperScaler, which is AWS, Azure, Google, Alibaba etc.

And AWS owns more than 50% of the HyperScaler market, so that is at least 25% of Intel's DC revenue going to Amazon, roughly $5B per year.

IDK what hyperscalar means exactly but if you just mean who are the big customers of intel Microsoft and Google are probably as big as Amazon, people forget that there is more going on in data centers than public cloud.

It really was a matter of time, but I honestly thought we’d see low cost Apple laptops with ARM CPUs before we saw much inroads in the data center.

As a developer, it’s never been easier for me to be cloud, OS, and CPU-architecture agnostic.

I’m sure AWS sees the writing on the wall, and I don’t see any other thing happening now other than a race to the bottom.

Anyone have thoughts on how this might compare to something like Ampere's ARM server chips? Seems like the custom chips are indeed ARM based as well? I saw some benchmarks recently on Ampere chips that had me fairly impressed.

https://kinvolk.io/blog/2019/11/comparative-benchmark-of-amp...

I actually benchmarked Ampere against AWS Graviton for CPU-intensive workloads. The summary is AWS Graviton is 25% faster: https://github.com/sanxiyn/blog/blob/master/posts/2019-11-12...

This new chip would improve over AWS Graviton.

Binary compatibility is the Moat that Intel and AMD have.

Try using ARM based server and you quickly realize the pain.

Besides, other Hyperscalers won't use Amazon's Chips. So no economy of scale.

Do you remember Amazon Phone ?

Personal anecdata: I tried using ARM based server and didn't find any pain.

Ok. I guess it depends on the apps and their dependencies. Typically we run in to library issues.

Can you be more specific? You run closed source libraries that you don't have the source for?

Compatibility issues?

Performance issues?

Something else?

Many libraries, even open source, have hand-rolled x64 assembly for the hot bits, or have transitive dependencies that have hand-rolled assembly. You want to rewrite that AES-NI code for ARM equivalency yourself?

Many open source projects don’t even have build tooling or testing on anything but x64.

That’s a lot of pain, even discounting buggy and poorly tested drivers.

In my experience, most libraries have C fallback for x64 assembly, if not another handrolled NEON assembly. I would like to know which libraries have x64 assembly without C fallback.

I’m sure most have C fallback, but that code is 20x slower and isn’t actually tested or used by anybody else in practice. The devs probably don’t even have unit tests or build environments for ARM server platforms.

As for which libraries, I’m too lazy to do a detailed search. All I can report anecdotally is that we very briefly tried ARM at dayjob on our codebase about a year ago and it seemed everything was immediately broken. Especially interop between high level languages and C libraries. The team tasked with it said almost imm fiat would “way too hard, let’s quit this experiment”.

Reply on news.ycombinator.com