4 ms·

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

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.

chvid 6y ago
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.
- chrisseaton 6y ago
  > 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.
  - amalter 6y ago
    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)
    - klelatti 6y ago
      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?
    - chrisseaton 6y ago
      > 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 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!
      - amalter 6y ago
        Mostly reporting what I’m seeing with our workload. Is it possible OpenJDK/Hotspot has more ARM coverage than Graal?
        chrisseaton 6y ago
        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.

chvid 6y ago

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.

chrisseaton 6y ago
> 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.
- amalter 6y ago
  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)
  - klelatti 6y ago
    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?
  - chrisseaton 6y ago
    > 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 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!
    - amalter 6y ago
      Mostly reporting what I’m seeing with our workload. Is it possible OpenJDK/Hotspot has more ARM coverage than Graal?
      - chrisseaton 6y ago
        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.