AWS re:Invent came and went, and what a week it was. While the total number of announcements out of the conference was huge, the company itself made 13 significant announcements. These ranged from 5G to edge to core infrastructure to Quantum computing. Moor Insights & Strategy Principal Analyst Patrick Moorhead provided some great coverage that can be found here and here. Likewise, fellow analysts Paul Smith-Goodson and Will Townsend provided coverage on the Quantum Computing and 5G fronts here and here. Additionally, my Moor Insights colleague Steve McDowell and I discussed re:Invent on our latest podcast, found here.
Two topics that spurred lots of buzz were the general availability of Outposts (and Local Zones) and the new EC2 instances based on the newly released Graviton2 processor designed and developed by the folks at AWS (formerly Annapurna Labs). Why are these announcements so significant, and what are the potential longer-term impacts? This article will attempt to unpack both announcements.
6th Generation EC2 instances based on Graviton2 – this is huge
AWS announced six total EC2 instances based on Graviton2:
- General Purpose instances M6g and M6gd (d denotes local NVMe storage)
- Compute Optimized instances C6g and C6gd
- Memory Optimized instances R6g and R6gd
While the general-purpose instance (M6g) is available for preview today, the compute-optimized and memory optimized instances are will be available “soon.” Note that AWS is greatly expanding the workloads that are supported by the Graviton family, tackling the workloads with intense compute and memory needs. This signals that the company views these Arm-based instances as performance peers to the x86 equivalents.
If this sounds like déjà vu all over again, it’s not. The reason is Graviton2 – an AWS designed and developed chip based on Arm’s Neoverse N1 architecture (released in February of this year – see my early coverage here). Graviton2 is built on a 7nm design process, with the features and specs of what one would expect to see in its x86 competitors – 64 cores, 8 memory channels, 64 PCIe lanes, a high speed interconnect to minimize latency and the microarchitectural additions that enable high performance computing (HPC) and analytics to run in an accelerated fashion. The numbers Patrick Moorhead listed in his Graviton2 coverage support these claims. This is how the M6g coverage compares to M5 performance:
- >40% better integer performance on SPECint2017 Rate (estimate)
- >20% better floating point performance on SPECfp2017 Rate (estimate)
- >40% better Java performance on SPECjvm2008 (estimate)
- >20% better web serving performance on NGINX
- >40% better performance on Memcached with lower latency
- >20% better media encoding performance for uncompressed 1080p to H.264 video
- 25% better BERT ML inference
- >50% better EDA performance on Cadence Xcellium EDA tool
These are impressive numbers. Perhaps more impressive is the performance profile these numbers demonstrate for Graviton2 (and EC2 instances based on it). The range demonstrated in the above reflects the range of workload types that populate the enterprise datacenter – from general purpose compute to high performance and all workloads in between. More plainly put, these numbers demonstrate that Graviton2 can support all the applications that power the modern business.
When looking at this level of performance at a reported 20% lower price point, it becomes clear that AWS sees a winner in Graviton2 and is strongly encouraging customers to increase utilization.
A last note on Graviton2 that must be addressed. Arm announced the Neoverse brand and family of products at TechCon in October 2018. It released the N1 and E1 architectures in February 2019. Ten months later, commercial silicon is in production. Even if one assumes the team at Annapurna got early access to the N1 architecture and received support from Arm, this is an incredible timeline to production.
Expect strong adoption of AWS’ 6th Generation EC2 instances. Furthermore, it would make sense that customers who have good experience with these instances will expand utilization when the cost-performance benefit is fully realized.
The question is this – does the launch of Graviton2 open the datacenter floodgates for Arm? In the near-to-mid term, the answer is “no.” The enterprise IT market doesn’t move that fast. However, 10 years from now we could be talking about this launch as the seminal moment in Arm’s datacenter efforts. If this sounds crazy or farfetched, just remember that the same views were held of Intel when it first grabbed a foothold in the datacenter about 30 years ago. According the experts back then, we should still be talking about big iron running Solaris, AIX and HP-UX running our datacenters.
Outposts availability: is AWS an infrastructure vendor?
AWS also announced the general availability of Outposts at re:Invent 2019. Outposts is essentially on-prem AWS. You like the idea of the cloud, but want it within striking distance? Talk to AWS about Outposts. AWS-branded infrastructure is delivered to the datacenter/server room, cabled and configured, provisioned and ready for IT. Its customers run native AWS or VMware environments, with seamless connectivity back to the AWS public cloud. Organizations gain full access to all tools (APIs, etc.) – no purchase of hardware required. IT simply pays for the AWS services consumed by the organization.
Is the above too good to be true? The idea of hands-off infrastructure consumed as a service is attractive. Cost shifts from capital to operational. Headaches associated with infrastructure management are reduced. The ability to expand cloud economic models to workloads and applications that have latency sensitivity issues is real. The ability to better manage the shadow IT cost and complexity issues associated with cloud becomes attainable. This is all good for Enterprise IT, or any IT organization for that matter.
One of the big questions that will take some time to answer is what impact Outposts will have on the traditional IT services companies. The market has already seen companies like HPE announce a strategic direction where every product is consumed as a service. The company also just announced GreenLake Central, a single pane of glass to manage it all – from edge to cloud to core datacenter. This move to become a neutral and unbiased control plane appears to be HPE’s attempt at differentiating. Strategically, this is a smart move. The real test will be how this service plays in the real world (scheduled release is second half of 2020).
Putting things in perspective
It is easy to get caught up in the hype surrounding some of the announcements that came out of AWS re:Invent 2019. Across the board, AWS has set a rate and pace of innovation that very few competitors can match – innovation rooted in relevance for today and impact for tomorrow.
Of the 13 major announcements, the launch of EC2 instances on Graviton2 is most interesting to me. Putting a finer point on this, Graviton2 is the most interesting to me. The last few years saw a series of fits and starts for Arm in the server market. Calxeda, Applied Micro, AMD, Samsung and Qualcomm are just a few of the very recognizable names that made big splashes in developing Arm for servers – only to pull stakes. In some cases, before the product even hit the market.
AWS is different. It is a silicon company and a big consumer of CPUs. Demonstrating a compelling value prop for Graviton2 through price and performance promotes Arm as a credible alternative and player for enterprise IT. I believe AWS embracing Arm and customers having success could be the ultimate “killer app” that spurs adoption longer term.
The availability of Outposts is also very interesting. Will AWS find success with Outposts? I believe the answer is “yes.” Will the deployment of Outposts increase an enterprise IT organization’s dependence on AWS? Again, I believe the answer is “yes” as enterprise IT begins to realize how much operational headaches it eliminates – from budgetary to management. The biggest question is this: longer term, what does an expanding Outposts environment mean for those companies that sell hardware and software into IT organizations? For that, I have no crystal ball – it’s going to fun to watch this all unfold.