We as an industry are finally having an open and honest discussion with ourselves, about why Arm Servers have not been as successful as we had all hoped, and why that is. Not excuses, just an analysis of facts and the consequences they have had.
To set the stage, let’s rewind 4 years. Arm had very high expectations for the Arm Server ecosystem, and made public statements on a regular basis that they intended to capture 20% of the Server market by 2020. We are now 9 months away from that date, and Arm Servers are closer to a 1% novelty still being explored and tested for workload compatibility by users.
In early 2015, the number was 20%, but by January of 2016 the number actually grew to 25%! Here is the opening statement from an article on the Next Platform at the time:
“So 2016 is the year, or at least it is supposed to be. The year when 64-bit ARM chips finally make their way into servers and perhaps start getting wheeled through the loading bays of actual datacenters to start running real workloads alongside the Xeon processors that by and large dominate in the glass house.”
And here are the 2015 claims:
- ARM plans to win 20 per cent of the server market by the year 2020
- Chip designer ARM sees 20 percent market share in servers by 2020
That sentiment and the high hopes continued through the next several years, with Arm Servers always ‘just around the corner’, ‘ready to make a breakthrough’, or ‘poised to take marketshare away from Intel’. AMD entered and left the ARM Server SoC business. Qualcomm entered and left the Arm Server SoC business. Applied Micro was acquired and their Server SoC was abandoned, but fortunately came back via an acquisition of their IP. Either way, time and knowledge were lost.
That brings us to the present day.
Last week, Arm further revealed the E1 and N1 Neoverse processors based on the Ares architecture, which were previously announced last fall at ArmTechCon 2018. Performance numbers and the processor design surely look great, and the product announcement generated plenty of media attention, as has been the norm. The dialogue has shifted, however, as the media has once again seen a new product announcement, but has begun to wonder when these Arm Servers will actually be competetive and ship in any real volumes. This conversation then exploded a few days later, when Linus Torvalds weighed in on the topic, claiming that Arm has already lost, due to the fact that developers and IT teams (for the most part) build and develop on x86 PC’s and workstations, then deploy to the organization’s own servers or the cloud providers. The natural tendency then, is to purchase x86 servers, to match the architecure and eliminate any chance of hardware incompatibility. The x86 PC is, more or less, just a smaller and cheaper version of the exact same server hardware running in the datacenter. Software stacks might vary between local and cloud, but if you can at least minimize differences and just focus on any bugs that may arise due to software variances, that is still better than pushing code up to servers and having to troubleshoot both software AND hardware variances. Thus, the ultimate takeway in Linus’ view, is that until there are cheap, standards-compliant Arm desktop PC’s that developers can use locally, Arm Servers will not stand a chance in the datacenter.
And he is right.
Arm Server enthusiasts have been complaining about this for years, and the product void and differences between the ultra-small and ultra-cheap Arm single board computers (think, Raspberry Pi and similar) and the big, enterprise Arm Servers is too vast. The missing piece in the ecosystem has been a “PC-like” system, that is standards compliant with SBSA and SBBR, and essentially mimics the entire user experience of the x86 PC platforms. This means a normal UEFI boot up and BIOS / GUI that allows user configuration, and a hardware design that allows end users to add whatever hard drives and SSD’s they’d like, choose their brand and size of memory, add in whatever graphics card they prefer, put it all together, and it “just works”. Then, they can install whatever (ARM64) OS they would like, such as RedHat, CentOS, Fedora, Debian, Ubuntu, Suse, etc. Again, it “just works”. The end result is a very “PC-like” experience, and a fully operational Arm desktop computer.
No custom kernels, no board-specific software, no vendor-provided code, no bugs, quirks, hacks, odd shapes and sizes, no fighting with the system, no frustration, and no time spent troublseshooting why the system doesn’t boot or hardware doesn’t work.
Once this is solved, and developers can easily build and test applications and code on an Arm Desktop PC locally, then Arm Servers become a more attractive target to deploy those apps to. Without an Arm Desktop PC, Arm Servers are just a novelty that individuals will continue to test, evaluate, and ultimately discard as not worth the switching costs, as has occured for the past 5 years.
Fortunately, Linus’ timing was excellent, as he spurred this debate just in advance of Linaro Connect, where miniNodes, Packet.net, and Linaro are delivering a joint session titled “Designing a next generation ARM Developer Platform“. In this talk, we will be discussing this problem in detail, hopefully spurring ideas, investment, and starting the process of correcting the problem and getting a small “NUC-like” Arm product engineered and built. The event is streamed live and recorded, so, tune in and watch if you can!
Hopefully, this is the tipping point, and industry change occurs as a result!