Just one question
For Intel, Is it the beginning of the end or the end of the beginning ?
Amazon Web Services has found another use for its home-brewed Graviton2 Arm processors: powering an instance type designed for burstable performance. Most instance types in Amazon’s EC2 service define a server with particular specifications. The T3 instance type instead offers servers with a baseline level of CPU performance, …
ARM servers in the datacenter are not going to be a Thing until there is a sufficiently common systems architecture that important software can be portable between different vendors' implementations without a massive conversion and optimization effort. Compare for example AWS ARM instances and the Fujitsu ARM compute nodes. The market today is a lot like the desktop computer world before the IBM PC (which, incidentally, is the direct ancestor of the Lintel server of today). Contrast the rapid adoption of GPUs for compute, which is facilitated enormously by Nvidia driving a standardized API across a broad range of cards.
The worst case scenario for ARM servers is that the market remains an archipelago of incompatible implementations. The second worst is that AWS ARM instances become the de facto standard by sheer weight of market presence, and everybody else is left trying to "do what AWS does" by inspection, much like S3.
I have to agree, almost none of our cloud-centric developers are writing code that is compiled anymore and those few that are, are usually doing it in Java.
With the rise of "serverless", nodejs and python it's clear that not much is needed to be compiled. Yes, Rust still needs to target a platform but with LLVM intermediates you aren't talking about a major risk when writing.
Arm recommends that Arm server vendors all stick to the Arm SBSA == Server Base System Architecture which should completely eliminate the "massive conversion and optimization effort" that you talk about.
From the Wikipedia page:
"The SBSA seeks to strengthen the ARM ecosystem by specifying a minimal set of standardized features so that an OS built for this standard platform should function correctly without modification on all hardware products compliant with the specification."
"Initial public version of the SBSA was announced on January 29, 2014."
This bodes well for the ARM ecosystem.
For each enterprise system that tunes their apps for more cores - and moves to more cost effective (burstable) systems - Amazon will continue to polish the stack and fill in missing holes in their offering.
(speculation) In 20 months - a Graviton 3 on TSMC 3nm - perhaps with some Nvidia ML IP blocks could provide Intel with more than just a headache.
Biting the hand that feeds IT © 1998–2020