WOA Issue 34
This is the last regular edition of the Works on…
As a longtime observer and participant in the “Arm in the Datacenter” community, I can confidently predict that over the next year we’ll move solidly from “some day” to compelling. While predictions are notably foolish to make, here is what is driving my enthusiasm:
It’s always fun to start with the good stuff: chips! 2021 is definitely shaping up to be the year in which Arm in the datacenter moves solidly from “some day” to compelling.
At the core of this are multi-year investments in designs led by Arm’s “Neoverse”. Amazon’s Graviton2 is based on Neoverse N1, and delivers robust performance and substantial cost advantages. Ampere’s next generation Altra is in early access, with expectations of peppy, consistent performance combined with lower power consumption.
In the HPC market, Fujitsu is leading the way with a top supercomputer based on their A64FX design. Marvell will deliver its next generation ThunderX3 — a monster with up to 384 threads. NVIDIA’s acquisition of Mellanox brings together market-leading SmartNICs with market-leading GPU technology. Add to that new, well-funded (and talent-rich) upstarts like Pensando and Nuvia’s, and we expect to see both trend-setting performance in widespread production, as well as an industry roadmap to build upon and invest in.
Portable software is better software, especially in an world defined by distributed applications and experiences. Code that works on more than one machine type and more than one architecture is more likely to be resilient in the face of change and growth, and today’s environment offers plenty of each. Building portable code takes extra work, but that work pays off in the form of flexibility and added reach.
One of the unexpected benefits of porting code to a new machine is that you get a fresh chance to optimize. As an example, when Amazon looked closely at PHP, they found that by ripping out some of the complex language features in PHP4 that were no longer supported in PHP7, they could pick up double digit performance gains. Similarly, Neocortix’s port of [email protected] to the arm64 environment gave them the opportunity to pull forward years of work on GROMACS, a set of scientific codes that were ripe for optimization.
The Arm ecosystem has focused on tools and tooling to ensure that there are resources available for all developers for continuous integration. Leading CI/CD vendors like Travis CI and Drone Cloud offer hosted arm64 build environments. They are joined by the likes of Github, Gitlab, Gitea and others to also provide a commercially viable option for building your commercial code on private Arm servers.
On the cloud native frontier, the promise of portable, architecture-agnostic software is alive and well. Most all existing CNCF projects perform well on Arm, and many projects see Arm support as a business advantage in a 5G and Edge world. In the core datacenter, Kubernetes is king, and companies like Rancher, Weave, and Kinvolk are at the forefront of ensuring that code built for the Kubernetes environment is robust and portable across architectures. Expect to see more of this sort of collaboration industry-wide.
For Arm to truly make 2021 a breakout year, cheaper is not the only answer — systems need to be as fast, and arguably better than the alternatives. A few unique trends are conspiring to make this possible.
One unique feature of the Arm ecosystem is the relative ease of entry into the market. Rather than build your own fabrication facilities, you can leverage the supply chain of partners who can license an architecture and core designs (Arm) or even build you a complete SoC to your design specifications (Marvell). This extended reach of ecosystem resources was what enabled Amazon to build its own Graviton2 chip, leveraging the experience of an Arm licensee (Annapurna Labs) which Amazon had acquired, and also taking advantage of fabrication facilities from TSMC. Similarly, Microsoft’s SQ1 was built specifically for the needs of its new Surface Pro X in a collaboration with Qualcomm.
Hyperscalers have their own workloads like databases and storage offerings that are consumed as services by their customers, who don’t need to know or care what chip those software-as-a-service workloads are running on. This is an ideal set of workloads for cutting-edge , especially those that have hardware acceleration that can be tuned to the specific task. New high core count Arm designs are more efficient than their competitors and predecessors, especially under heavy loads.
Apple, which controls as much of its supply chain as it can manage, sees the benefits of Arm architecture in unifying their chip story to take advantage of Apple Silicon across iPad, iPhone, and MacOS designs. This will likely have the spillover effect of broadening the world of software and developers that touch Arm-based silicon for datacenter-related development.
And its not just for hyperscalers anymore. Companies like Pensando, Mellanox and others are bringing 2nd or 3rd generation accelerators to market. This allows Enterprises to take advantage of unique technology as they invest in market leadership across mobility, entertainment, financial services, and more.
There are a zillion mobile phones in the market, and nearly all run on Arm-based cores. This means that there’s a huge manufacturing and silicon process technology investment i that provides a competitive advantage as Arm licensees grapple with datacenter use cases..
The Arm ecosystem has been thriving, with wins from supercomputers to laptops. What helps this all along is a process advantage, which means the world’s most advanced manufacturing technology is used not only to build Arm-based chips for mobile phones, but also Arm-based CPUs and specialty chips that are transforming the massive datacenter ecosystem. .