Hacker News Reader

Comment by mav3rick | Hacker News Reader

Who is replacing Intel servers ? AMD ?

Also ARM chips for server use is rapidly starting to become a serious thing.

It's - of course - very small scale and early days at the moment, but the pricing is very positive and it's something people are starting to think about as a real viable option.

The servers will flip once ARM gets cheap enough to x86.

If I have a database and some Java or Ruby running on some Linux then why would I care whether it is on x86 or anything else?

However are those ARM CPUs going to be made by NVIDIA?

It seems to me that there are other fabless chip producers out there capable of doing this.

> If I have a database and some Java or Ruby running on some Linux then why would I care whether it is on x86 or anything else?

Databases, Java, and Ruby, are well-optimised for running on AMD64. Are they all as well-optimised for running on AArch64, for example? I'm not so sure. I still see basic intrinsics being implemented for Java on ARM. I think people may be in for a bit of a shock when they try ARM with some common software.

We saw pretty decent performance with AWS Gravitron. The pricing is a bit opaque as AWS might be taking a loss, but M6g.2xlarge steps all over M5.2xlarge on your basic Tomcat/Spring Boot app.

Remember there has been a tremendous amount of effort made to make VM's like the JVM and v8 performant on ARM for mobile. Turns out all those intrinsics scale just fine when presented with your typical "Business Application" workload.

I think Intel is especially in trouble with JVM workloads as it's very rare for libraries to call out to native code (unlike Python).

You really can just drop you jar on an ARM (m6g), AMD (m5a) or Intel (m5, m5n) expect the same results. After a bit of hesitation it will become just a standard part of cloud price optimization.

Cloud operators win - Chip vendors get commoditized.

(Modulo specialized workloads like HPC, ML and such - but to the topic, that's not to Intel's strength)

Plus having ARM widely deployed on the desktop via Apple may start to incentivise developers to fill in the remaining gaps where native code libraries are used?

> Remember there has been a tremendous amount of effort made to make VM's like the JVM ... performant on ARM for mobile.

This doesn't reflect what I'm seeing at all - in fact I'm seeing basic ARM intrinsics still being added to the Graal compiler in the last few weeks!

https://github.com/oracle/graal/pull/2588

> You really can just drop you jar on an ARM (m6g), AMD (m5a) or Intel (m5, m5n) expect the same results.

Not without these kind of basic intrinsics you won't!

Mostly reporting what I’m seeing with our workload. Is it possible OpenJDK/Hotspot has more ARM coverage than Graal?

OpenJDK is more complete, but that is also pretty recent work. I think it's an indication of icebergs that Oracle's flagship new compiler is still having random drive-by commits by non-core-team members that add pretty fundamental intrinsics.

It's not that they're going away, it's that they are turning into a more commodity market. You used to have 5000-10000 enterprises buying high-margin Xeons. Every dentist office had 5-6 Xeon chips sold at high margin.

Now, a significant, growing chunk of the market is in AWS, Azure, GCP, etc. Those vendors get favorable terms and make it harder for Intel to segment the market to maximize profits. When Intel was calling on me a couple of years ago, they were pushing storage hard (ruler, etc), but I don't see OEMs embracing Intel branding as storage, and Samsung seems to be winning deals that I'm involved in.

> Every dentist office had 5-6 Xeon chips sold at high margin.

What on earth for? My dentists office has a couple of old Opterons. What were yours using their Xeons for?

I go to a small practice. They have a 3D xray machine, whatever a Microsoft small business server is called, and a practice management system.

When it's time to refresh, it will be 100% cloud.

> Every dentist office had 5-6 Xeon chips sold at high margin.

That's a great observation.

For instance, I recently worked on a project to provide services to insurance agents via the cloud.

To me, it was interesting because it offered a lot of value for an insurance agency. Basically the last thing an insurance agent wants to worry about is some server in a closet that has a Blue Screen of Death.

AMD may not replace Intel, but it is certainly exerting a severe downward price pressure in the datacenter, Intel's most profitable market. The servers my company is deploying in its current refresh are all AMD EPYCs, after a 6–8 month qualification cycle.

As for client computing, ARM is taking over the increasingly irrelevant PC/Mac after having overwhelmingly won mobile, and IoT will also be ARM.

Furthermore, Trump's trade war against China will claim Intel as collateral damage. The Chinese will switch to RISC-V or the AMD-licensed but made in China x64 chips:

https://www.anandtech.com/show/15493/hygon-dhyana-reviewed-c...

Given that the sanctions also blocked Huawei's license of the ARM design. I think the Chinese will create their own processor family.

China has a processor from Sunway powering the #4 supercomputer as of June 2020, it was #1 when it launched in 2016.

https://www.top500.org/system/178764/

https://en.m.wikipedia.org/wiki/Sunway_(processor)

Or fork ARM, ignoring that they're not paying for a license.

something like now in progress

https://asia.nikkei.com/Business/China-tech/How-SoftBank-s-s...

The Power ISA is open source and well supported by a variety of software isn't it? That's what I'd go with if I were in their shoes.

Yes, the new EPYC cpus offer very good performance and value.

Surprisingly, POWER is now drawing even in TCO for small servers, if your shop has the software development chops to do it.

Reply on news.ycombinator.com